Metaheuristic-optimized interaction-aware deep learning with large language model assistance for data-driven water quality prediction Authors: Ebrahim A Mattar, El-Sayed M El-Kenawy, Sarah M Alhammad, Marwa M Eid Published: 2026 ABSTRACT AbstractAccurate prediction of water-quality indicators remains challenging in small tabular environmental datasets because physicochemical variables can exhibit nonlinear interdependencies and modern deep-learning models are sensitive to hyperparameter configuration. In this study, the supervised regression task is defined as predicting dissolved oxygen (mg/L) from the remaining measured physicochemical variables using a modest public dataset of 200 samples. To address this task, the Automatic Feature Interaction Network (AutoInt) is coupled with the Ninja Optimization Algorithm (NiOA) for wrapper-based hyperparameter optimization under a controlled and reproducible experimental protocol. All evaluations use a fixed train/validation/test split (70%/15%/15%) with leakage-safe preprocessing based only on training-set statistics, and results are aggregated over independent runs with controlled initialization seeds and mean ± standard deviation reporting. Baseline AutoInt without metaheuristic optimization achieves a mean squared error of , whereas the NiOA-optimized AutoInt configuration reaches under the matched optimization budget of 1500 function evaluations (, ). These findings indicate that NiOA-guided hyperparameter tuning can substantially improve AutoInt performance within this fixed benchmark setting. However, because the dataset is small, cross-sectional, and lacks explicit spatial or temporal structure, the results should be interpreted as benchmark-specific evidence rather than broad operational validation. Further evaluation on larger, independent, and spatiotemporally diverse water-quality datasets is required before generalizing the fitted model to wider environmental monitoring or decision-support applications.Keywords: Water quality prediction, Tabular deep learning, Metaheuristic optimization, Hyperparameter tuning, Environmental data modeling, Reproducible benchmarkingSubject terms: Engineering, Mathematics and computing Introduction Water quality assessment constitutes a foundational element of contemporary environmental monitoring and risk-management practice, underpinning ecosystem conservation, sustainable water-resources governance, and public health protection1,2. In operational terms, the concept of “water quality” refers to the extent to which physicochemical conditions support designated uses (e.g., drinking-water supply, irrigation, aquatic habitat integrity, and industrial use) and remain within acceptable risk thresholds. Because aquatic systems integrate signals from catchment processes, atmospheric exchanges, and human activities, water-quality monitoring functions not only as a diagnostic activity but also as an early-warning component of environmental stewardship. Both surface-water and groundwater systems are subject to continuous modification by interacting natural drivers (e.g., lithology-controlled geochemical weathering, seasonal hydrologic regimes, and ecological cycling) and anthropogenic pressures (e.g., agricultural intensification, urban wastewater discharges, industrial effluents, and land-use change). These coupled influences generate complex spatial and temporal variability in water physicochemistry, often producing nonstationary patterns that complicate interpretation and regulatory compliance. For instance, the same site may exhibit distinct water-quality regimes across hydrologic seasons, with dilution, evapoconcentration, and biogeochemical cycling altering concentrations and relationships among variables in ways that are not constant through time. Accordingly, systematic monitoring and evaluation of water-quality indicators remain indispensable for detecting contamination events, tracking ecological condition, supporting enforcement of water-quality standards, and informing prevention-oriented management strategies2. At the same time, the present study focuses on a compact cross-sectional benchmark rather than on a spatially or temporally resolved monitoring network; therefore, the environmental motivation is used to justify the dissolved-oxygen regression task, whereas broad claims about site transfer, seasonal dynamics, or operational deployment require independent validation on datasets that explicitly contain spatial and/or temporal structure. In routine monitoring programs, a set of core physicochemical parameters is commonly employed to characterize water-quality status. These include acidity–alkalinity balance (pH), oxygen regime (dissolved oxygen), thermal conditions (temperature), optical and particulate behavior (turbidity), ionic strength and salinity proxies (electrical conductivity), and nutrient-related indicators (e.g., nitrate concentration where relevant). Collectively, such variables provide an operationally efficient yet information-rich description of water conditions because they encode multiple interacting processes: pH influences chemical speciation, solubility equilibria, and the viability of many aquatic organisms; dissolved oxygen reflects metabolic balance and aeration–consumption dynamics; temperature modulates reaction kinetics, gas solubility, and biological activity; turbidity serves as a proxy for suspended matter that can affect light penetration and habitat quality; and conductivity reflects dissolved ionic content that is frequently associated with geogenic contributions and anthropogenic inputs. Beyond their individual interpretations, these indicators are informative precisely because they co-vary under shared drivers, meaning that multivariate patterns can reveal process-relevant structure even when individual measurements appear within nominal bounds. For example, shifts in conductivity alongside changes in turbidity and dissolved oxygen may indicate altered mixing, runoff contributions, or changes in dominant biogeochemical pathways, while pH can modulate how other constituents partition between dissolved and particulate phases. The selection, measurement, and interpretation of these indicators are supported by established monitoring guidance and standardized analytical procedures, which help ensure methodological comparability across sites and time1,3. Standardization is particularly important in regulatory and cross-site synthesis contexts, where differences in sampling, preservation, and analytical conditions can create apparent trends that do not reflect environmental change. Historically, water-quality assessment has relied on descriptive statistics, threshold-based compliance checks, composite indices, and deterministic or empirical models grounded in physical and chemical reasoning. These approaches remain essential because they provide interpretable benchmarks, regulatory alignment, and process-informed understanding. In practice, threshold comparisons translate scientific knowledge and policy priorities into actionable criteria, while indices offer concise summaries that can support communication and screening across many sites or time points. However, such methods can be limited when relationships among parameters are strongly nonlinear, when interactions are high-order and context dependent, or when heterogeneous conditions introduce regime shifts that are difficult to encode explicitly. In environmental systems, these challenges are common because multiple processes can act simultaneously, and the effective relationship between variables may change as boundary conditions (e.g., flow regime, temperature range, or disturbance intensity) shift. Such limitations have contributed to increased interest in complementary data-driven approaches, particularly as open-access environmental datasets have become more available and as modern learning algorithms have matured. In this setting, machine learning provides a pragmatic framework for learning predictive relationships directly from observations, enabling rapid estimation of target variables, supporting exploratory analysis, and facilitating comparative evaluation across alternative modelling choices4. Importantly, the value of such models in monitoring contexts is often practical as well as scientific: reliable surrogate prediction can reduce measurement burden for difficult-to-monitor variables and can assist in prioritizing follow-up sampling or intervention. Nevertheless, when datasets are small and cross-sectional, predictive gains must be interpreted conservatively because they may reflect strong within-sample correlations, sensitivity to a particular train/validation/test partition, or validation-set over-specialization during wrapper-based hyperparameter optimization rather than generalizable environmental structure. Tabular environmental datasets—typical of many water-monitoring contexts—are especially relevant for assessing modern learning methods because they comprise structured numeric variables with potentially intricate dependency patterns. These datasets often combine heterogeneous units, differing magnitudes, and partially redundant signals, so predictive performance depends on the capacity to model interactions and conditional effects rather than only linear trends. At the same time, it is increasingly recognized that deep learning does not automatically outperform traditional methods on tabular data; success depends on appropriate architectures, inductive biases for interaction modeling, and careful training configuration5. This observation is particularly salient for environmental applications where sample sizes are frequently modest and where the cost of overfitting is high in decision-support settings. Consequently, tabular water-quality datasets provide a meaningful testbed for evaluating interaction-aware learning models and robust optimization strategies under realistic data constraints. Public benchmarks hosted on platforms such as Kaggle further enhance transparency by enabling repeatable experimentation and standardized comparison across studies6. Such benchmarks also encourage clearer reporting of preprocessing and validation choices, which is essential when transferring methods from general-purpose machine learning to domain-sensitive environmental inference. In this work, the public benchmark is used primarily to evaluate methodological transparency, controlled hyperparameter optimization, and repeated-run reporting in a small-sample setting; it is not treated as sufficient evidence for broad deployment across heterogeneous water bodies. Recent progress in artificial intelligence also shows that modern learning systems increasingly rely on structured representations, geometric modeling, and data-efficient optimization when learning from complex or limited observations. Hypergraph learning, for example, provides flexible tools for modeling high-order correlations beyond pairwise relationships7, while recent work on deep loss convexification studies how learning can reshape nonconvex iterative-model objectives to improve local optimization behavior8. Related advances in deep learning for point-cloud instance segmentation, remote-sensing interpretation, virtual–real 3D reconstruction, and data-scarce generative modeling further illustrate the broader movement toward representation learning under structural complexity and limited-data conditions9–12. Although these studies address application domains different from water-quality regression, they motivate the present emphasis on interaction-aware representation learning, explicit optimization design, and reproducible evaluation under constrained data availability. Despite substantial progress in environmental machine learning, predictive modelling of water-quality variables remains challenging for both scientific and practical reasons. First, water-quality parameters are interdependent: they interact through feedback mechanisms shaped by hydrology, biogeochemistry, and external stressors, and these interactions can vary across seasons, sites, and disturbance regimes. This means that predictive structure is often conditional: the effect of one variable on a target may depend on the levels of other variables, and the same multivariate configuration may have different implications under different hydroclimatic contexts. As a result, accurate prediction often requires models capable of capturing nonlinear dependencies and high-order interactions rather than only marginal effects of individual predictors4. From a problem-definition perspective, the task can therefore be framed not merely as fitting a regression surface, but as learning a multivariate mapping that remains stable under limited data, measurement noise, and potential regime heterogeneity that is typical of monitoring programs. Because the present dataset is cross-sectional and does not include explicit temporal or spatial covariates, the modelling task here is restricted to learning relationships within the observed sample distribution, and broader claims about seasonality- or site-driven regimes are treated as motivating context rather than directly tested properties. Second, model performance is frequently sensitive to configuration choices. In particular, advanced tabular deep-learning architectures are designed to learn complex interaction structures, but their predictive reliability depends strongly on hyperparameters governing representational capacity (e.g., embedding size and depth), training dynamics (e.g., learning-rate schedules and optimizer settings), and regularization (e.g., dropout and weight decay). Sensitivity is not only a practical concern but also a methodological one, because unstable tuning can lead to performance estimates that are difficult to reproduce and conclusions that depend on accidental choices rather than principled design. This sensitivity can be amplified in small to medium-sized environmental datasets, where limited sample size increases the risk of overfitting and makes training dynamics less stable under suboptimal configurations5,13. In such settings, hyperparameters effectively determine the balance between underfitting (insufficient capacity to represent interactions) and overfitting (memorization of idiosyncratic patterns), and small changes in regularization or learning rate can materially alter the learned interaction structure. Accordingly, credible performance claims require evaluation protocols that quantify variability across repeated runs, disclose the selected hyperparameters, and explicitly acknowledge the risk of validation-set over-specialization when wrapper-based tuning is applied to small datasets. Third, conventional hyperparameter tuning methods such as grid search or random search can become inefficient as dimensionality increases and as interactions among hyperparameters become more pronounced. Grid search scales poorly because it allocates equal effort to all dimensions regardless of their influence, while random search can still require many trials to reliably discover high-performing regions when the objective is noisy and the effective dimension is high. While random search can be more effective than exhaustive grids in many high-dimensional scenarios, it may still be computationally burdensome when each configuration requires iterative training and validation14. This motivates the use of metaheuristic optimization algorithms, which are widely used for derivative-free optimization in complex, multimodal landscapes and can provide a practical balance between exploration of the search space and exploitation of promising configurations15. Canonical families of such methods include swarm intelligence approaches (e.g., particle swarm optimization) and evolutionary strategies (e.g., differential evolution), which have been extensively studied and adopted across engineering optimization contexts16,17. In the present context, the objective surface is further complicated by stochastic training effects and by the possibility that multiple hyperparameter combinations yield similar validation performance, creating plateaus and local basins that challenge purely local search strategies. In small-sample settings, this challenge is compounded by the risk that wrapper-based tuning over-specializes to a particular validation subset; therefore, the optimization protocol must be paired with transparent reporting of evaluation budgets, selected hyperparameters, run-level variability, and statistical comparisons across optimizers. Recent metaheuristic research increasingly emphasizes adaptive search control, hybridization, population-diversity preservation, learning-guided operators, and stability-aware mechanisms rather than relying only on metaphorical inspiration. For example, elite-pool and Brownian-motion mechanisms have been introduced to strengthen the sine cosine algorithm for global optimization18, while the snow ablation optimizer has been proposed as a population-based method for numerical and engineering design problems19. Improved slime mould variants further illustrate the use of dynamic random search, adaptive mutation, adaptive grouping, and learning operators to balance exploration and exploitation20,21. Other recent developments incorporate reinforcement learning and gradient-search schemes into JAYA22, hybridize grasshopper optimization with complementary search mechanisms23, or introduce stability-oriented social learning strategies for photovoltaic-parameter estimation24. These studies collectively indicate that meaningful progress in metaheuristic optimization is increasingly associated with explicit search-behavior design, robustness analysis, and reproducible benchmarking. At the same time, the broader optimization literature has raised important concerns about metaphor-based metaheuristics, especially when algorithmic novelty is justified primarily by biological, physical, or behavioral analogies rather than by clearly specified operators, transparent benchmarks, and statistically defensible comparisons. The present study therefore does not treat the metaphor underlying any optimizer as evidence of theoretical superiority. Instead, all metaheuristic methods are considered as black-box search procedures acting on the same hyperparameter vector, the same feasible bounds, and the same evaluation budget. This framing is important because the central question is not whether a metaphor is plausible, but whether a published optimizer can provide useful, reproducible hyperparameter search behavior for an interaction-aware tabular model under controlled experimental conditions. Within this methodological landscape, integrating metaheuristic search with interaction-aware tabular models offers a coherent pathway to improving robustness and generalization in water-quality prediction tasks. Specifically, metaheuristics can be used to systematically explore hyperparameter spaces that are nonconvex, noisy, and potentially constrained, without relying on gradient information and without requiring analytically tractable objective functions15. This is particularly relevant when computational budgets are limited and when the objective surface is shaped by stochastic training behavior, as is common in neural-network optimization. Conceptually, the problem becomes one of allocating limited evaluations to identify stable, well-generalizing configurations that capture meaningful feature interactions while maintaining sufficient regularization to avoid spurious associations that can arise in small environmental datasets. To make such searches auditable, the hyperparameter domain and optimization budget must be explicitly disclosed, the selected configuration must be reported, and aggregate metrics must be accompanied by repeated-run dispersion and statistical testing. The overarching goal of this study is to develop a robust and methodologically transparent framework for data-driven water-quality prediction using tabular deep-learning techniques. Robustness here refers to achieving strong predictive performance while maintaining stability across runs and avoiding undue sensitivity to arbitrary training choices, whereas methodological transparency refers to clearly specifying preprocessing, validation, and tuning steps so that results can be audited and reproduced. To support reproducibility and facilitate benchmarking, the study uses a publicly available Kaggle dataset comprising routine water-quality measurements, thereby enabling independent verification and extension by the research community6. The use of a public dataset also helps separate methodological contributions from site-specific idiosyncrasies, allowing the modelling pipeline to be evaluated as a general-purpose approach for small-sample tabular water-quality prediction problems. In this study, the prediction target is fixed as dissolved oxygen, and all reported results correspond to this single-target regression setting. A central objective is to employ the Automatic Feature Interaction Network (AutoInt) as the primary predictive model. AutoInt is specifically designed to learn feature interactions through self-attention mechanisms, providing an explicit inductive bias for capturing high-order dependencies in structured tabular data. Unlike approaches that require manual construction of feature crosses, AutoInt aims to learn interaction structure end-to-end by weighting relationships among embedded feature representations, enabling flexible modeling of conditional effects among physicochemical variables. This property is particularly relevant in water-quality contexts, where predictive performance can depend on interaction effects among physicochemical variables rather than on additive contributions alone13. In practical terms, an interaction-aware model is expected to better reflect coupled behavior among variables (e.g., regimes in which the influence of one predictor strengthens or weakens depending on another), which is a common characteristic of aquatic-system measurements. Given that strong correlations may exist among water-quality parameters in small observational datasets, the study emphasizes leakage-safe preprocessing and transparent validation to reduce the likelihood that near-deterministic associations or preprocessing artifacts inflate performance estimates. To address the sensitivity of AutoInt to hyperparameter selection, the study integrates the Ninja Optimization Algorithm (NiOA) as a hyperparameter optimization mechanism. NiOA is treated as an established, previously published metaheuristic, and its role in this work is confined to guiding hyperparameter search rather than introducing a new optimizer. By leveraging NiOA for derivative-free tuning, the study aims to improve the stability and effectiveness of AutoInt configuration under constrained computational budgets25. NiOA is selected because it is the focal optimizer evaluated in the proposed workflow and because the study examines whether this published population-based optimizer can serve as an effective wrapper-based hyperparameter-optimization engine for AutoInt under a fixed evaluation budget. The study does not claim that NiOA is theoretically superior to all metaheuristics; rather, it compares its empirical behavior against selected population-based optimizers under harmonized search spaces, population sizes, iteration limits, and validation objectives. This objective emphasizes the optimization process as a methodological component: the intent is to systematically identify hyperparameter settings that yield reliable generalization rather than to rely on ad hoc defaults. Finally, the study conducts a systematic comparison between NiOA-assisted AutoInt, baseline AutoInt training without metaheuristic optimization, and alternative optimizer-assisted configurations under consistent experimental conditions, ensuring fairness and methodological rigor. This comparative design is intended to clarify whether performance improvements arise from the choice of interaction-aware architecture, from the choice of optimizer, or from their coupling within a controlled experimental protocol. In addition, the study reports both predictive accuracy and measured computational cost under a matched budget, while treating computational-resource comparisons as implementation-specific observations rather than universal complexity rankings. A further emerging direction for understanding metaheuristic behavior is the use of complex-network analysis to study population interactions, information flow, and collective search dynamics in evolutionary computation and swarm intelligence. Network-based views of evolutionary computation and particle swarm optimization provide a route toward analyzing optimizer behavior beyond metaphorical descriptions by examining how candidate solutions interact, cluster, diversify, and converge over time. This perspective is outside the empirical scope of the present small-sample benchmark, but it motivates future work in which the search dynamics of NiOA and comparator optimizers could be analyzed using network-science descriptors in addition to final predictive metrics. The contributions of this work are methodological and empirical. First, the study establishes a structured preprocessing and feature-handling workflow tailored to tabular water-quality data, emphasizing data consistency, numerical stability, and methodological transparency in alignment with standardized measurement principles3. This includes adopting practices that support fair evaluation, such as applying scaling procedures consistently and maintaining strict separation between training and evaluation subsets to mitigate leakage-driven optimism. Second, it introduces a hybrid modelling workflow that couples AutoInt with NiOA for systematic hyperparameter optimization, thereby combining an interaction-aware tabular architecture with a metaheuristic search strategy suitable for complex, nonconvex tuning landscapes13,15,25. The combination is positioned as a controlled benchmarking workflow rather than as a new optimizer or a new neural architecture. The combination is positioned as a practical response to the observed sensitivity of tabular deep models, aiming to discover well-performing configurations efficiently while remaining compatible with the black-box nature of validation-based objectives. Third, it frames comparative evaluation against established optimizers using a standardized metric suite and consistent computational constraints, focusing on relative behavior and robustness rather than isolated claims. This framing supports interpretation of results as comparative evidence under matched budgets, which is especially important for metaheuristics whose effectiveness can depend strongly on evaluation allocation. Finally, the study aligns with best practices for interpretability by incorporating diagnostic and explainable modelling tools to support transparency in environmental decision-support contexts26. In applied environmental modeling, such interpretability support is essential for linking predictive patterns back to plausible domain relationships and for enabling responsible use of model outputs in monitoring and management workflows. Given the limited size and cross-sectional nature of the benchmark dataset used here, the study treats its findings as evidence of methodological feasibility under controlled conditions, with broader environmental generalization reserved for follow-on validation on larger, diverse monitoring datasets. The remainder of this paper is organized as follows. Section Related work reviews relevant literature on water-quality modelling, tabular deep learning, and metaheuristic hyperparameter optimization. Section Materials and methods describes the dataset, preprocessing procedures, learning models, and optimization methodology. Section Experimental setup outlines the experimental setup and validation strategy. Section Results presents the experimental results. Section Discussion discusses the findings, implications, and limitations. Finally, Section Conclusion concludes the paper and outlines future research directions. Related work Water-quality prediction and assessment occupy a well-established methodological space spanning environmental science, ecosystem protection, and public-health risk management. In practical monitoring settings, agencies and research programs routinely collect multivariate physicochemical measurements—including indicators of acidity/alkalinity, oxygen regime, ionic-content proxies, and optical surrogates of suspended particulate matter—to characterize water-system status and diagnose potential stressors. These measurements are rarely independent: they are linked through coupled geochemical equilibria, transport processes, and biologically mediated transformations, and they often respond jointly to hydrologic variability and human pressures. As a result, water-quality datasets are typically characterized by nonlinear dependencies, context-specific interactions, heteroscedastic noise, and occasional regime shifts driven by seasonal cycles or episodic disturbances. These characteristics have progressively motivated the adoption of data-driven methods that can (i) represent nonlinear functional relationships without requiring fully specified process models, (ii) remain effective when observations are limited, noisy, or unevenly distributed across conditions, and (iii) support interpretability and methodological transparency so that predictions can be scrutinized and responsibly used in management contexts. In alignment with these needs, the prior work most directly motivating the present study can be synthesized into three connected themes: (a) machine learning (ML) for water-quality prediction and index-based assessment, (b) deep learning for tabular data with explicit or implicit feature-interaction modeling, and (c) hyperparameter optimization (HPO), emphasizing metaheuristic strategies as practical alternatives when conventional tuning becomes inefficient. Given the small-sample and cross-sectional characteristics typical of many publicly available environmental benchmarks, including the dataset used in this study, particular emphasis is placed on validation rigor, uncertainty quantification, and transparent reporting of optimization procedures. Machine learning for water-quality prediction and index-based assessment A substantial portion of the ML literature on water quality can be organized around two complementary objectives. The first objective is direct prediction of specific water-quality variables (e.g., dissolved oxygen, nutrient concentrations, turbidity proxies, conductivity-related measures), typically framed as regression over multivariate measurements. The second objective is composite assessment, often operationalized through water quality indices, where multiple variables are aggregated into a single diagnostic quantity used for communication, screening, and management prioritization. Although these objectives differ in their outputs—variable-level estimates versus synthesized indicators—they share common methodological requirements, including robust handling of correlated predictors, principled validation under realistic data constraints, and transparency regarding uncertainty and sensitivity to modelling assumptions. For direct prediction, comprehensive evidence indicates that ML methods are adopted primarily because the governing processes behind water quality are strongly coupled and nonlinear, and therefore difficult to capture with simple linear statistical models across heterogeneous conditions. A broad review by Zhu et al.4 synthesizes ML applications across multiple water-system contexts (surface water, groundwater, drinking water, wastewater, and marine environments), emphasizing both the performance potential and recurring methodological pitfalls (e.g., limited representativeness, inconsistent validation protocols, and underdeveloped interpretability practice). Beyond cataloguing algorithms, this line of work highlights that modelling success in environmental systems is frequently constrained by data-generation realities: sampling designs may be opportunistic rather than statistically balanced; measurement error and censoring can be nontrivial; and the joint distribution of predictors may differ across sites or time periods, thereby challenging generalization. At a complementary level of granularity, Huang et al.27 provides a wide-angle review of ML in natural and engineered water systems, situating water-quality prediction within a broader methodological pipeline that includes data preparation, model selection, evaluation, and deployment considerations. Importantly, such pipeline-oriented accounts emphasize that performance claims are only as credible as the preprocessing and validation design that supports them, particularly when models are intended for operational decision support. In parallel, reviews focused on deep learning (DL) for water-quality management emphasize that modern neural architectures are increasingly used to capture high-dimensional dependencies and temporal dynamics, while also cautioning that success is often contingent on careful training design and fit-for-purpose validation28. This caution is especially pertinent when complex models are applied to modest datasets common in environmental monitoring, where apparent gains can reflect favourable splits or tuning artifacts rather than robust predictive improvement. Recent methodological critiques further stress that repeated resampling, dispersion reporting (mean ± standard deviation), and clear disclosure of target scaling are necessary safeguards when reporting very small error magnitudes on limited environmental datasets. Applied research further illustrates that predictive performance depends not only on the choice of learner, but also on the extent to which training regimes respect the statistical and physical structure of water-quality data. For example, dissolved oxygen modeling is frequently used as a benchmark due to its central ecological role and its sensitivity to interacting drivers; Dehghani et al.29 demonstrates how hybrid ML strategies can be used for dissolved oxygen prediction in running waters, highlighting the value of hybridization when single learners struggle under complex, interacting covariates. The methodological motivation for such hybridization is typically that different model classes capture different aspects of the mapping from predictors to targets (e.g., smooth nonlinear trends versus localized responses or interaction effects), and combining them can reduce variance or mitigate systematic bias under shifting conditions. In the broader literature, these studies collectively reinforce a key implication: water-quality datasets often exhibit interaction effects that are scientifically expected but difficult to encode manually (e.g., temperature modulating oxygen solubility and reaction kinetics; ionic-strength proxies correlating with pollutant loads; optical surrogates reflecting particulate transport). Consequently, modelling strategies that can discover and exploit interactions directly from data are attractive, provided that they are validated carefully and interpreted responsibly. However, when datasets are small and lack explicit temporal or spatial stratification, careful split design and sensitivity analysis are indispensable to ensure that learned interactions reflect generalizable structure rather than idiosyncratic sample patterns. For composite assessment, the water quality index (WQI) literature underscores that index definition, parameter weighting, and aggregation choices can materially affect both uncertainty and interpretability. A detailed review by Uddin et al.30 emphasizes that WQI pipelines introduce uncertainty and subjectivity that can dominate downstream conclusions if not carefully documented and justified. This concern is not merely philosophical: alternative weighting schemes, normalization choices, and aggregation operators can yield meaningfully different rankings or classifications even when based on the same underlying measurements, which complicates comparisons across studies and can affect policy communication. More recent applied studies have increasingly coupled ML-based predictors with explainability tooling to bridge predictive accuracy and interpretability—particularly when index outputs support decision-making and risk communication. For instance,, Choudhary et al.31 integrates stacked ensemble regression with Shapley Additive explanations (SHAP) to interpret WQI prediction, exemplifying a broader trend toward explainable environmental ML. The methodological significance of such work is that it treats interpretability as a first-class requirement rather than an afterthought, thereby supporting more defensible use of ML outputs in contexts where stakeholders must understand which measured factors drive predicted risk or status. From the standpoint of methodological design, these strands converge on a practical challenge that is highly relevant to the present study: water-quality datasets are frequently tabular, moderate-to-small in sample size, heterogeneous in units, and characterized by nonlinear couplings. These properties create demand for tabular learning models that (i) can represent feature interactions without brittle manual feature construction and (ii) can be tuned and validated with particular care to avoid overstated performance, especially when datasets are small enough that resampling variance and split sensitivity become prominent. Accordingly, robust hyperparameter optimization and transparent evaluation protocols are increasingly viewed not as peripheral refinements but as central methodological requirements in environmental ML research. Deep learning for tabular data and interaction modeling Although gradient-boosted decision trees remain strong baselines on many tabular problems, there is sustained interest in DL architectures that learn representations and interactions end-to-end. A unifying perspective is that tabular data are not merely “vectors of numbers”; rather, they are structured collections of heterogeneous fields whose predictive signal can depend on higher-order interactions among features. In many scientific applications, the same numeric feature can have different predictive meaning depending on the configuration of other variables, and the relevant signal may be concentrated in interaction terms rather than in strong marginal effects. A comprehensive survey by Borisov et al.5 organizes tabular DL into architectural and training paradigms, including specialized inductive biases, attention-based mechanisms, and regularization strategies aimed at improving competitiveness and stability. Within this framing, tabular DL is not viewed as a monolithic alternative to trees, but as a family of designs that attempt to inject useful structure (e.g., interaction discovery, feature selection, or hierarchical abstraction) into neural learning pipelines. In addition, an influential viewpoint by y Shwartz-Ziv and Armon32 argues that DL is not universally superior for tabular learning, thereby sharpening the methodological requirement that DL-based choices should be justified, carefully tuned, and compared against strong baselines. In practical terms, this viewpoint reinforces the importance of experimental discipline: when DL is applied to tabular environmental datasets, credible gains require controlled comparisons, transparent hyperparameter reporting, and validation strategies appropriate to the data-generating process. This is particularly critical in small-sample settings, where performance differences may fall within the variance induced by stochastic optimization and data partitioning. Within tabular DL, explicit interaction modeling is particularly relevant for environmental data because physical and chemical variables often co-determine outcomes through coupled mechanisms. AutoInt (Automatic Feature Interaction Learning via Self-Attentive Neural Networks) is a prominent architecture designed explicitly to learn feature interactions through multi-head self-attention operating on embedded feature representations13. The conceptual advantage of this mechanism is that attention can assign varying importance to pairwise and higher-order interactions in a data-adaptive manner, effectively learning which cross-feature relations are most predictive under the observed distribution. While introduced in the context of feature crossing for recommender systems, its central mechanism—attention-based discovery and weighting of interactions among embedded features—generalizes naturally to regression settings where cross-feature dependencies are expected but difficult to specify analytically. This is directly aligned with water-quality prediction tasks, where interactions can reflect underlying chemistry (e.g., joint controls on solubility or equilibrium), hydrologic conditions (e.g., dilution and transport effects), or measurement couplings (e.g., proxies for common latent processes). Nevertheless, the ability of attention-based models to capture such dependencies does not obviate the need for careful validation; interaction learning can amplify overfitting risk when sample sizes are limited and regularization is insufficient. AutoInt also fits into a broader family of attention-centric tabular models. TabTransformer adapts Transformer-style contextual embeddings to tabular inputs, emphasizing representation robustness (especially for mixed feature types) via contextualization across fields33. In such architectures, contextualization can be interpreted as learning feature representations that are conditional on the broader feature set, which can help when the relevance of a given variable depends on the state of other variables. SAINT introduces attention across rows and columns and explores contrastive pretraining to improve label efficiency, which is directly relevant to scientific datasets where labeled samples may be limited34. From a methodological standpoint, this introduces an additional degree of freedom: performance may be improved not only by architectural choices but also by leveraging representation learning objectives that exploit unlabeled structure. Other notable architectures seek different trade-offs between performance and interpretability. TabNet introduces sequential attentive feature selection, yielding step-wise feature usage patterns that can support interpretation35. Such step-wise selection can be particularly valuable in environmental applications where model users may wish to understand which measured parameters were emphasized at different stages of the prediction process. NODE (Neural Oblivious Decision Ensembles) blends tree-ensemble structure with differentiable optimization, aiming to combine benefits of tree-like inductive bias and neural training36. This hybridization is relevant to environmental tabular learning because it attempts to retain the partitioning behaviour and interaction handling typical of trees while maintaining the end-to-end trainability of neural methods. Finally, empirical reassessments such as Gorishniy et al.37 (which introduces the FT-Transformer among other baselines) emphasize that training protocols and architectural details can be decisive for DL competitiveness on tabular data. This observation is particularly important for small datasets, where differences in normalization, early stopping, learning rate schedules, and regularization can dominate outcomes and obscure the contribution of the model class itself. Two methodological lessons from this literature are especially salient for the present work. First, interaction-oriented tabular DL can reduce the need for extensive manual feature engineering, which is appealing in environmental systems where interactions are expected but their functional form is uncertain. This can support more standardized modelling pipelines across monitoring programs, provided that learned interactions are validated and interpreted with care. Second, performance is often highly sensitive to hyperparameters (e.g., embedding size, number of attention heads, depth, dropout, learning rate schedules, and regularization), especially on small datasets. This sensitivity elevates HPO from a “routine tuning step” to a core methodological component that can determine validity, robustness, and reproducibility, particularly when model selection decisions are based on incremental performance differences. Consequently, detailed disclosure of the hyperparameter search space, optimization budget, and selected configuration is necessary to ensure that reported improvements are auditable and reproducible. Hyperparameter optimization and metaheuristic search HPO is central to ML reliability because hyperparameters control model capacity, regularization, and optimization behavior, thereby shaping generalization. In tabular DL settings, hyperparameters govern both representational structure (e.g., embedding dimensionality and attention configuration) and training dynamics (e.g., step sizes, batch effects, and regularization strength), and they can interact in ways that produce non-intuitive performance landscapes. Classical baselines include grid search and random search. Importantly,, Bergstra and Bengio14 establishes that random search is often more efficient than grid search when only a subset of hyperparameters strongly influences performance—an insight that remains practically relevant for deep tabular models whose search spaces are high-dimensional and anisotropic. In such cases, exhaustive grids tend to waste evaluations on unimportant dimensions, whereas random sampling can allocate more trials to exploring influential factors. Bayesian optimization provides a more sample-efficient paradigm by fitting a surrogate model over the hyperparameter space and selecting new trials using an acquisition rule38. Conceptually, this substitutes direct exploration with probabilistic modelling of the objective surface, balancing exploitation of known good regions against exploration where uncertainty remains high. While powerful, Bayesian approaches can be less straightforward when the space is mixed (continuous, integer, categorical), the objective is noisy (e.g., stochastic training, resampling variance), and evaluation budgets are limited—conditions commonly encountered in deep tabular learning for environmental datasets. In these circumstances, surrogate modelling assumptions may be strained, and the overhead of maintaining and optimizing a surrogate can be less attractive relative to direct, population-based exploration. Metaheuristic optimization has therefore been widely explored as a complementary HPO strategy. A review by Abd Elaziz et al.15 synthesizes how advanced metaheuristics are used in deep neural network contexts, emphasizing their derivative-free global search capability and their practical adaptability to rugged, multimodal landscapes. In HPO terms, the appeal is that metaheuristics can operate directly on black-box objectives, tolerate noisy evaluations, and flexibly accommodate constraints and mixed-variable encodings. Canonical population-based algorithms frequently serve as competitive baselines in HPO studies. Particle Swarm Optimization (PSO) is widely used due to its simple velocity–position update rules and effective exploration–exploitation behavior in continuous domains16. Differential Evolution (DE) is valued for its mutation-and-recombination operators that maintain diversity while driving convergence17. More recent metaheuristics, such as the Multi-Verse Optimizer (MVO), propose novel operators motivated by cosmological analogies and are commonly adopted as comparative baselines in engineering optimization39. Stochastic Fractal Search (SFS) uses diffusion-inspired mechanisms and has also been applied broadly in global optimization settings40. In environmental modeling contexts, the appeal of metaheuristics is amplified by constraints typical of real water-quality problems: limited sample sizes, heterogeneous measurement regimes, and noisy objectives induced by resampling-based validation. In such settings, an HPO strategy must not only search effectively but must do so under strict evaluation budgets where each fitness evaluation corresponds to a full training-and-validation cycle. A concrete example is Kushwaha et al.41, which studies metaheuristic approaches in water-quality index prediction and reinforces the empirical importance of robust optimization and careful validation. The broader methodological message is that optimization and evaluation are inseparable: metaheuristics can provide practical search capability, but credible conclusions depend on harmonized budgets, consistent data splitting, and transparent reporting of configurations and randomness controls. Recent critiques further emphasize that, in small datasets, wrapper-based HPO must be paired with repeated runs and dispersion reporting to mitigate the risk of overfitting to a single validation subset. Within this landscape, the present study adopts the Ninja Optimization Algorithm (NiOA) as the HPO engine for AutoInt. NiOA is treated as a previously published metaheuristic, and its use in this work is methodological: to guide systematic hyperparameter search under a fixed experimental protocol rather than to introduce a new optimizer25. This positioning responds to a recurring gap across environmental ML studies: performance gains are frequently reported without sufficiently transparent, reproducible, and fairly controlled tuning procedures. From a comparative standpoint, coupling a strongly interaction-oriented tabular DL model (AutoInt) with a structured metaheuristic HPO strategy (NiOA) is intended to address two linked sources of fragility—model sensitivity to configuration and the inefficiency of naive tuning in complex search spaces—while keeping the evaluation design consistent across alternatives. By combining an interaction-capable tabular DL model (AutoInt) with a structured metaheuristic HPO strategy (NiOA) and by comparing against established optimizers under harmonized budgets, the study aims to strengthen methodological rigor while retaining practical relevance for water-quality analytics. In addition, the study emphasizes full disclosure of the hyperparameter domain, optimization budget, selected best configuration, and computational cost metrics, thereby aligning with emerging expectations for reproducible environmental machine-learning research. Materials and methods Figure 1 illustrates the overall architecture of the proposed NiOA-assisted AutoInt framework. The pipeline integrates leakage-safe data preprocessing, automatic feature interaction learning via self-attention, and nature-inspired hyperparameter optimization into a unified regression framework for tabular water-quality prediction. Unlike conventional tuning strategies, the proposed approach employs NiOA as a wrapper optimizer that directly minimizes validation error under a fixed evaluation budget, enabling the AutoInt model to adapt its architecture and training configuration to the underlying physicochemical characteristics of the data. This modular design ensures reproducibility, fair comparison with competing methods, and extensibility to alternative datasets and optimization strategies. All optimization experiments are conducted under a harmonized computational budget (, , corresponding to 1500 fitness evaluations), and the final selected configuration is explicitly reported to ensure auditability. Overview of the proposed NiOA-assisted AutoInt framework for tabular water-quality prediction. Raw physicochemical measurements are first validated, split into training/validation/test sets, and standardized in a leakage-safe manner. A nature-inspired optimization algorithm (NiOA) then performs wrapper-based hyperparameter optimization of the AutoInt model by iteratively proposing candidate configurations, training AutoInt, and minimizing the validation mean squared error. The optimal configuration is retrained on the combined training and validation data and evaluated on the test set using standard regression metrics. Optional analysis branches (dashed) denote benchmarking against baseline models and alternative optimizers, which are not part of the core predictive pipeline. Dataset description The dataset employed in this study is obtained from the publicly available Water Quality Testing dataset hosted on Kaggle. This dataset consists of 200 observations, each corresponding to a distinct water sample collected for routine water-quality assessment. The dataset is intentionally modest in size, reflecting realistic data-scarcity conditions frequently encountered in environmental monitoring studies, particularly in small-scale surveys or pilot monitoring campaigns. The limited sample size motivates careful validation design, strict leakage prevention, and explicit reporting of variability across repeated runs to reduce the risk of overfitting and split sensitivity. Because the adopted 15% held-out test subset contains approximately 30 samples, the reported test-set results are interpreted as fixed-split benchmark evidence rather than definitive evidence of broad environmental generalization. Each observation comprises a set of numerical variables representing physicochemical properties of water. According to the dataset specification, the recorded parameters include pH, temperature (expressed in degrees Celsius), turbidity (measured in nephelometric turbidity units), dissolved oxygen concentration (expressed in milligrams per liter), and electrical conductivity (expressed in microsiemens per centimeter). In addition, an identifier field is provided to uniquely label individual samples; this identifier is not used as an explanatory variable in the modeling process. All measured variables are continuous-valued and directly derived from standard water-quality testing procedures, making the dataset suitable for regression-based predictive modeling. The dataset is distributed under a CC0 public-domain license, ensuring unrestricted use for research and educational purposes. No temporal or spatial metadata are associated with the samples, and the dataset represents a cross-sectional snapshot rather than a time series. As a result, the modeling task is formulated purely in a tabular regression setting, without explicit consideration of temporal dependencies or spatial autocorrelation. Accordingly, any claims regarding temporal dynamics, seasonal variability, or spatial generalization are outside the empirical scope of the present dataset and are treated as contextual motivation rather than directly validated properties. Within this study, the predictive task is defined as a supervised regression problem in which dissolved oxygen (mg/L) is selected as the fixed target variable to be predicted from the remaining measured parameters (pH, temperature, turbidity, and electrical conductivity). The input feature set therefore consists of these four physicochemical variables, while dissolved oxygen represents the quantity of interest. This formulation reflects a common use case in water-quality analysis, where certain parameters may be expensive, delayed, or difficult to measure directly, motivating their estimation from correlated and more readily available measurements. The specific choice of target variable is fixed consistently throughout all experiments to ensure comparability across models and optimization strategies. All reported performance metrics correspond exclusively to this dissolved-oxygen regression task. Data preprocessing Prior to model development, a systematic data preprocessing procedure is applied to ensure numerical consistency, methodological rigor, and protection against data leakage. Given the relatively small dataset size, careful preprocessing is essential to avoid introducing biases or artifacts that could disproportionately influence model behavior. The preprocessing pipeline begins with data validation and integrity checks. All records are examined for missing entries, duplicated samples, and non-numeric values. As the dataset is provided in a cleaned tabular format, no missing values are present in the raw data. Consequently, no imputation is performed in this study; the preprocessing pipeline includes automated checks for completeness, but no median or alternative imputation strategy is applied because the dataset contains no missing entries. Automated checks are retained as a safeguard to ensure robustness and reproducibility should the dataset be extended or modified in future studies. Subsequently, range and plausibility checks are conducted for each physicochemical variable. These checks verify that all recorded values fall within physically meaningful and scientifically plausible intervals, consistent with standard water-quality measurement practices. Observations exhibiting clearly anomalous or non-physical values are inspected to determine whether they reflect data-entry errors or extreme but valid conditions. In this dataset, no samples are excluded at this stage, as all values conform to reasonable environmental ranges. Following validation, numerical scaling is applied to the feature set. Because the variables are expressed in heterogeneous units and exhibit differing numerical magnitudes, feature scaling is necessary to ensure stable optimization and balanced gradient updates during model training. A normalization strategy based on standardized scaling is adopted, transforming each feature to have zero mean and unit variance based exclusively on statistics computed from the training subset. This approach prevents any single variable from dominating the learning process due to scale effects while preserving relative variations within each feature. The target variable (dissolved oxygen) is retained on its original physical scale (mg/L) for evaluation, and all reported regression metrics (MSE, RMSE, MAE, NSE, and related statistics) are computed after inverse transformation where applicable, ensuring interpretability in physical units. To prevent information leakage, all preprocessing transformations that depend on data statistics are fitted solely on the training portion of the dataset and subsequently applied unchanged to the validation and test subsets. This strict separation ensures that no information from held-out data influences the training process, thereby preserving the integrity of performance evaluation. The same preprocessing configuration is applied consistently across all models and optimization strategies considered in the study. Finally, the processed dataset is partitioned into training, validation, and testing subsets according to the experimental protocol described later in this paper. A fixed split ratio of 70%/15%/15% (training/validation/test) is adopted, and repeated runs with different controlled initialization seeds are used to quantify variability while maintaining identical split indices for fair optimizer comparison. This partitioning strategy supports both hyperparameter optimization and unbiased performance assessment, forming a stable foundation for the subsequent modeling and optimization stages. To remove ambiguity regarding the role of the two preprocessing streams used in the baseline analysis, Table 1 specifies them as deterministic and executable pipelines. Stream A and Stream B denote reproducible preprocessing configurations used for comparative baseline evaluation; neither stream depends on an LLM during execution, model fitting, hyperparameter optimization, or metric computation. Table 1Deterministic preprocessing pipelines used in Stream A and Stream B.Pipeline stepStream AStream BMissing-value checkYes; no imputation requiredYes; no imputation requiredDuplicate/non-numeric checkYesYesPlausibility/range checkYesYesFeature scalingStandardization using training-set statistics onlyStandardization using training-set statistics onlyTarget handlingDO evaluated on original mg/L scaleDO evaluated on original mg/L scaleLeakage preventionFit transformations on training only; apply to validation/testFit transformations on training only; apply to validation/testLLM dependence during executionNoneNone Deterministic preprocessing pipelines used in Stream A and Stream B. Stream A is used as the reference preprocessing configuration for subsequent optimization because it produced stronger baseline performance than Stream B in the comparative baseline analysis. This selection is explicitly reported for transparency, but it may introduce preprocessing-selection bias; therefore, this issue is acknowledged as a limitation when interpreting the optimized results. Feature handling and representation The representation and handling of input features constitute a critical component of the modeling pipeline, particularly in tabular environmental datasets where variables are heterogeneous yet strongly interdependent. In this study, all physicochemical parameters provided in the dataset are treated as continuous numerical features and are directly supplied to the learning model following the preprocessing steps described previously. No categorical encoding is required, and no temporal or spatial indexing is introduced, as the dataset represents independent cross-sectional measurements. A key design decision in this work is the deliberate avoidance of extensive manual feature engineering. Traditional water-quality modeling studies often construct derived variables, ratios, or interaction terms based on domain knowledge, such as combinations of temperature and dissolved oxygen or conductivity-normalized measures. While such handcrafted features can be informative, they also introduce subjectivity, increase dimensionality, and may inadvertently bias the learning process toward preconceived relationships. Instead, this study relies on AutoInt’s internal mechanism for learning feature interactions. AutoInt is specifically designed to model high-order interactions among tabular features through self-attention layers, enabling the network to automatically identify and weight relevant cross-feature dependencies during training. By embedding each feature into a shared latent space and applying attention-based interaction layers, AutoInt can capture both low-order and complex nonlinear relationships without requiring explicit specification of interaction terms. This design choice offers several advantages in the context of water-quality data. First, it reduces the risk of omitting important but non-obvious interactions that may not be captured by manual feature construction. Second, it limits unnecessary expansion of the feature space, which is particularly important given the modest dataset size. Third, it promotes methodological generality, allowing the same framework to be applied to other water-quality datasets without extensive redesign of feature-engineering rules. At the same time, the study acknowledges that interaction-aware models can be sensitive to hyperparameter configuration in small datasets, which motivates the structured and fully disclosed hyperparameter optimization procedure described in subsequent sections. Because the dissolved-oxygen regression task uses only four input predictors, the interaction-aware architecture is evaluated as a methodological benchmark rather than as evidence that deep learning is universally necessary for all water-quality datasets. All features are therefore treated uniformly within the AutoInt architecture, with their interactions learned adaptively during training. Hyperparameters governing the embedding dimension, number of interaction layers, attention heads, learning rate, regularization, and batch size are optimized through the metaheuristic framework described later, ensuring that the model capacity is aligned with the complexity of the data under a controlled and explicitly reported optimization budget. Exploratory data analysis Exploratory Data Analysis (EDA) is a foundational step for characterizing the statistical structure, dependency patterns, and latent organization of water-quality variables prior to predictive modeling. In this study, EDA is conducted within a controlled, human-in-the-loop framework that leverages two independent open-weight large language models for high-level methodological support. The role of these models is strictly limited to descriptive reasoning about already-generated visualizations and statistical summaries; they do not access raw data directly, do not execute code, do not modify preprocessing steps, and do not influence model architecture, hyperparameter selection, or evaluation design. These models are used exclusively for diagnostic reasoning, visualization interpretation, and consistency checking, without performing automated inference, modifying data, or influencing downstream modeling decisions. This design ensures transparency while enabling structured and systematic exploration of the dataset. All quantitative modeling decisions reported in subsequent sections are fully deterministic and independent of LLM-generated suggestions. The EDA objectives are threefold: first, to validate data consistency and physical plausibility; second, to identify linear and nonlinear relationships among physicochemical parameters; and third, to motivate the selection of interaction-aware tabular learning architectures capable of capturing the observed dependencies. Because the dataset is cross-sectional and limited to 200 samples, EDA is interpreted as descriptive characterization of the observed sample distribution rather than as evidence of broader spatiotemporal regimes. Mistral-based processing Pairwise distributions and marginal behavior. Figure 2 presents a pairwise scatterplot matrix augmented with marginal histograms for all numerical variables. The diagonal panels indicate compact and physically plausible marginal distributions for pH, temperature, turbidity, dissolved oxygen, and conductivity. The off-diagonal panels reveal heterogeneous dispersion patterns and curved point clouds, suggesting that the relationships among variables are not strictly linear. In particular, dissolved oxygen (the fixed prediction target in this study) shows structured variation with respect to temperature and conductivity, reflecting known thermodynamic and chemical influences in aquatic systems. The absence of extreme outliers or multimodal artifacts supports the numerical stability of subsequent regression modeling, while the visible curvature and dispersion patterns motivate nonlinear interaction modeling rather than purely additive formulations. Pairwise scatterplot matrix with marginal histograms for the numerical water-quality variables, illustrating distributional properties and bivariate relationships. Correlation structure and variable clustering. To quantify linear dependencies and explore higher-level organization, Fig. 3 shows a hierarchically clustered Pearson correlation heatmap. Strong positive correlations are observed among pH, dissolved oxygen, and conductivity, with coefficients exceeding 0.65 in several cases. Turbidity exhibits weak to moderate negative correlations with most variables, while temperature shows moderate positive association with conductivity and dissolved oxygen. The dendrogram reveals two dominant clusters: a chemically driven cluster (pH, dissolved oxygen, conductivity) and a physically driven cluster (temperature and turbidity). This structure indicates that interaction effects are likely more informative than independent feature contributions. Given that dissolved oxygen is the regression target, the observed moderate-to-strong linear correlations underscore the importance of carefully designed validation and leakage prevention, as strong associations in small datasets can otherwise lead to overly optimistic performance estimates if preprocessing is not strictly partition-aware. Hierarchically clustered Pearson correlation heatmap illustrating linear dependency strength and variable grouping among water-quality parameters. Multivariate scatter encoding. Figure 4 extends the analysis beyond pairwise projections by encoding multiple variables simultaneously. Temperature and dissolved oxygen define the axes, turbidity is encoded through marker size, and pH is represented by color. The visualization shows that higher dissolved oxygen levels tend to co-occur with higher pH and lower turbidity, whereas elevated turbidity is associated with increased dispersion and reduced oxygen levels. These patterns illustrate higher-order interactions that are difficult to capture using linear or univariate diagnostics. Such conditional structures support the methodological choice of interaction-aware architectures, while also highlighting the need for regularization to prevent overfitting to localized multivariate patterns in a modest-sized sample. Multivariate scatter plot of temperature versus dissolved oxygen, with point size representing turbidity and color representing pH. Latent structure visualization. Figure 5 presents a two-dimensional t-distributed stochastic neighbor embedding (t-SNE) projection of the standardized dataset. The projection reveals several well-separated regions corresponding to low, medium, and high dissolved oxygen regimes. Although t-SNE is not used for inference or modeling, the observed separation provides qualitative evidence of latent structure in the data that can benefit from nonlinear representation learning. It is emphasized that t-SNE is used purely for visualization and does not influence model training, hyperparameter optimization, or evaluation procedures. Two-dimensional t-SNE projection of the water-quality dataset, highlighting latent cluster structure across oxygen-level regimes. Density-based relationship analysis. Figure 6 shows a hexbin density plot of pH versus dissolved oxygen. The high-density ridge structure indicates a strong nonlinear association between these variables, while sparsely populated peripheral regions highlight combinations that occur infrequently. Such uneven density emphasizes the need for models that generalize robustly across both dense and sparse regions of feature space. In small datasets, sparse regions are particularly susceptible to high-variance predictions; therefore, robust validation and dispersion reporting are essential to assess stability across the feature domain. Hexbin density plot of pH versus dissolved oxygen, illustrating concentration patterns and nonlinear dependency structure. Qwen-based processing An independent EDA pass is conducted with methodological guidance from a Qwen-based large language model to cross-validate and refine qualitative insights. This analysis emphasizes distributional symmetry, conditional variability, and multivariate structure, providing an additional perspective on the dataset. As with the Mistral-assisted pass, the Qwen-based model provides only descriptive interpretation of precomputed figures and does not alter preprocessing, feature selection, model configuration, or optimization procedures. Univariate distribution analysis. Figure 7 presents violin plots for all numerical variables. The plots confirm relatively compact distributions with mild skewness, particularly for dissolved oxygen and conductivity. The absence of extreme multimodality or heavy tails supports numerical stability in subsequent learning stages, while the visible spread confirms sufficient variability for regression analysis. The moderate dispersion observed in dissolved oxygen further justifies reporting evaluation metrics on the original physical scale (mg/L) to preserve interpretability. Violin plots illustrating the distributional characteristics of pH, temperature, turbidity, dissolved oxygen, and conductivity. Three-dimensional multivariate structure. Figure 8 visualizes the joint relationship among pH, temperature, and dissolved oxygen, with conductivity encoded by color. The resulting manifold-like structure indicates that conductivity varies smoothly with combined changes in chemical and thermal conditions, reinforcing the role of multivariate interactions. These multivariate patterns provide qualitative support for interaction-aware modeling but do not constitute independent evidence of predictive superiority; formal evaluation is reserved for the controlled experimental protocol described later. Three-dimensional scatter plot of pH, temperature, and dissolved oxygen, with point color representing conductivity. Conditional distributions across pH regimes. Figure 9 examines turbidity behavior across discretized pH regimes. While central tendencies remain comparable, higher pH regimes exhibit slightly increased dispersion, indicating conditional heterogeneity. This regime-dependent behavior suggests that turbidity–pH relationships are context-specific rather than uniform. Such conditional variability reinforces the need for models capable of representing interaction terms while being regularized sufficiently to avoid overfitting to regime-specific noise. Distribution of turbidity across discretized pH regimes, illustrating conditional variability and regime-dependent behavior. Model interpretability and diagnostic analysis Model interpretability and diagnostic analysis are conducted to ensure that predictive behavior aligns with physically meaningful relationships and to identify potential sources of systematic error. The diagnostic workflow is used only for descriptive interpretation and consistency checking. It does not modify preprocessing, feature selection, model training, hyperparameter optimization, statistical testing, or quantitative performance evaluation. Mistral-based processing Interpretability analysis guided by the Mistral-based model builds directly upon the EDA findings. The strong coupling among pH, dissolved oxygen, and conductivity observed in Figures 2–6 provides an empirical reference against which learned feature interactions can be assessed. The clustered correlation structure (Fig. 3) and the multivariate and density-based visualizations establish expectations that interaction-aware models should emphasize chemically and physically coherent variable groups rather than isolated predictors. The latent regimes revealed by t-SNE further motivate regime-aware diagnostic checks to ensure stable performance across distinct regions of feature space. Such diagnostics are interpreted qualitatively and are complemented by quantitative performance metrics, dispersion statistics, and statistical testing reported in the Results section. Qwen-based processing Feature importance diagnostics. A descriptive feature-importance diagnostic was inspected only as a qualitative consistency check for the dissolved-oxygen prediction task. To avoid any ambiguity between the diagnostic target and the manuscript’s fixed prediction target, the figure previously associated with this paragraph is not used as quantitative evidence in the main text. The interpretation is therefore restricted to the statement that feature-importance diagnostics, when used, are descriptive rather than causal and do not affect model fitting, hyperparameter tuning, or evaluation. Latent Space Diagnostics. Figure 10 shows a t-SNE projection colored by pH regime. Clear separation among regimes indicates that pH acts as a primary organizing variable in the latent feature space. This observation suggests that predictive errors and uncertainty may vary across regimes, underscoring the importance of robust validation and interaction-aware modeling strategies. As with previous visualizations, this projection is used solely for qualitative interpretation and does not influence the quantitative evaluation protocol. t-SNE projection of water-quality features colored by pH regime, illustrating latent structure and regime separation. Overall, the convergence of insights from both Mistral- and Qwen-assisted analyses reinforces the presence of structured nonlinear dependencies and regime-specific behavior in the dataset. Importantly, the dual-LLM workflow serves only as a consistency-checking and descriptive aid; all modeling, optimization, and validation procedures remain fully specified, deterministic, and reproducible independent of LLM guidance. These findings provide a coherent empirical foundation for the subsequent use of interaction-aware deep learning models and metaheuristic hyperparameter optimization. Learning models The predictive modelling component of this study is centered on modern deep-learning architectures developed explicitly for tabular data. In contrast to conventional multilayer perceptrons, contemporary tabular deep models incorporate architectural inductive biases (e.g., attention, feature selection, structured interaction operators, or tree-inspired modules) that are designed to address the heterogeneity and interaction-dominant structure typical of real-world tabular datasets. This modelling choice is particularly relevant for environmental datasets (including routine water-quality measurements), where multiple physicochemical variables may jointly influence the target through nonlinear couplings and high-order dependencies rather than through additive effects alone. Consequently, the study adopts a model set that spans interaction-centric attention networks, attention-driven feature selection mechanisms, transformer-style tabular representations, and structured interaction operators, thereby enabling a principled and informative benchmarking under a unified experimental protocol. All models are evaluated on the same dissolved-oxygen regression task under identical preprocessing, data partitioning, and evaluation settings to ensure fairness and reproducibility. The primary predictive model employed in this work is the Automatic Feature Interaction Network (AutoInt). AutoInt was originally proposed to learn high-order feature interactions in a parameter-efficient manner by mapping heterogeneous input fields into a shared embedding space and then applying stacked multi-head self-attention interaction layers with residual connections13. Conceptually, AutoInt transforms each input feature (or field) into a latent vector representation and uses attention operators to compute interaction-aware representations, where the attention weights quantify the contextual relevance of one feature to another in the current predictive setting. By stacking interaction layers, AutoInt can represent increasingly complex (higher-order) interaction patterns without requiring explicit manual construction of feature crosses. This design is well aligned with water-quality modelling, because physicochemical variables such as pH, temperature, dissolved oxygen, turbidity, and conductivity frequently exhibit coupled behavior governed by chemical equilibria, transport dynamics, and biological processes, and these couplings can be difficult to capture through handcrafted interaction terms alone. Moreover, AutoInt’s interaction learning is achieved within a comparatively compact architecture (relative to very deep generic networks), which is advantageous when the available sample size is limited and when controlling overfitting risk is essential for credible environmental inference13. Given the modest dataset size (200 samples), careful regularization and hyperparameter control are necessary to balance representational capacity and generalization stability. To contextualize AutoInt’s behaviour and to ensure that observed gains are not artefacts of a single modelling choice, several established tabular deep-learning models are included as baseline and comparator architectures. First, TabNet is incorporated as a widely used attention-based model that performs sequential attentive feature selection. TabNet iteratively selects a sparse subset of features at each decision step, combining representation learning with an interpretable, instance-dependent feature-selection mechanism that can improve efficiency and provide diagnostic insights into feature usage patterns35. Second, SAINT is adopted as a transformer-style approach for tabular learning that performs attention across both columns (features) and rows (samples) and investigates contrastive/self-supervised pretraining to improve performance when labelled data are limited34. The inclusion of SAINT provides a strong reference point for attention-driven representation learning that differs from AutoInt’s explicit feature-interaction layers. Third, the comparator set includes DANets (reported in this work under the label DANet), which introduce an Abstract Layer mechanism to explicitly group correlated raw features and form higher-level abstractions, combined with architectural components intended to preserve access to raw-feature information at deeper levels42. This family is relevant as it represents a distinct philosophy: rather than only learning interactions through attention across embedded features, DANets emphasize semantic abstraction and feature grouping as a route to robust tabular modelling. In addition, the study benchmarks against the Deep Learning Recommendation Model (DLRM), which is a structured neural architecture designed to learn from a mixture of dense and sparse inputs by combining embedding tables, multilayer perceptrons, and an explicit feature interaction operator43. Although introduced as a state-of-the-art recommender architecture, the methodological principle—explicit interaction modelling coupled with deep nonlinear transformations—makes DLRM a relevant comparator for structured regression tasks that benefit from interaction operators. Finally, the baseline set includes a Transformer-style tabular baseline (TST) implemented as a simplified feature-wise Transformer encoder with multi-head self-attention layers, feed-forward sublayers, residual connections, and layer normalization, adapted for regression over standardized continuous inputs. This model serves as a generic transformer baseline distinct from AutoInt, without specialized feature-cross operators. To further strengthen benchmarking rigor, classical non-deep-learning baselines are also included, namely Random Forest (RF) and Extreme Gradient Boosting (XGBoost), which are widely recognized as strong performers on tabular datasets. Their inclusion ensures that any observed gains from interaction-aware deep models and metaheuristic optimization are evaluated relative to competitive tree-based approaches commonly used in environmental modeling. Across all architectures, learning is formulated as supervised regression with consistent preprocessing, data splitting, and evaluation procedures to ensure that performance differences can be attributed to modelling and optimization choices rather than experimental artefacts. Training and evaluation protocols (including stopping criteria, optimization settings, and computational budgets) are harmonized across models, which is particularly important in small-sample environmental settings where variance due to resampling and stochastic optimization can be substantial. This design also provides a clean foundation for the subsequent hyperparameter-optimization analysis: by first benchmarking a diverse set of tabular architectures under consistent conditions, the study isolates the incremental value of interaction modelling and metaheuristic hyperparameter optimization relative to baseline training. Hyperparameter optimization framework Hyperparameter optimization is formulated in this study as a continuous, bounded, high-dimensional optimization problem over the configuration space of the Automatic Feature Interaction Network (AutoInt)13. Let denote a d-dimensional vector collecting the tunable hyperparameters of AutoInt, where is the feasible domain induced by architectural constraints (e.g., admissible embedding dimensions, interaction-layer counts, and attention-head allocations) and training constraints (e.g., allowable learning-rate ranges and batch sizes). The explicit search space shared by all optimizers includes learning rate , weight decay , dropout , number of attention heads , number of interaction layers , embedding dimension , multilayer perceptron hidden units //, and batch size . The goal of hyperparameter tuning is to identify an optimal configuration that minimizes an out-of-sample objective computed on a validation subset, under an explicit computational budget that limits the number of fitness evaluations. Formally, the problem is written as where denotes an AutoInt instance parameterized by 13, and denotes the validation loss associated with the supervised regression task. In this study, corresponds to mean squared error (MSE) computed on the validation subset at the early-stopping checkpoint. Primary optimizer: Ninja Optimization Algorithm (NiOA). The Ninja Optimization Algorithm (NiOA) is adopted as the primary hyperparameter optimization engine in this study25. Within the present hyperparameter optimization setting, NiOA is used strictly as an established optimizer: it proposes candidate configurations, receives validation-loss feedback, and updates its population accordingly; no modification of NiOA’s core mechanism is introduced. All optimizers, including NiOA, operate with population size and maximum iterations , yielding 1500 total function evaluations for fair comparison. The fitness of each candidate configuration is defined through the validation loss: The final NiOA-selected hyperparameter configuration for AutoInt is explicitly reported as follows: learning rate , weight decay , dropout , number of attention heads , number of interaction layers , embedding dimension , multilayer perceptron hidden units , and batch size . This configuration is retrained on the combined training and validation subsets before final evaluation on the held-out test set. Operationally, NiOA is implemented as a population-based wrapper around AutoInt training. Each search agent represents one encoded AutoInt hyperparameter vector. At each iteration, candidate vectors are decoded into admissible architectural and training values, bounded to remain within the feasible search domain, evaluated by training AutoInt under the same early-stopping protocol, and ranked using validation MSE. The published NiOA stages are retained as the optimizer update mechanism: exploration is used to diversify candidate positions, mutation introduces controlled perturbations, exploitation refines promising solutions, and stagnation handling updates the best-so-far solution when improvement stalls25. In this work, these operators are used only to navigate the AutoInt hyperparameter space; no new NiOA search equation, coefficient schedule, or algorithmic variant is introduced. Table 2NiOA search procedure used for AutoInt hyperparameter optimization.StepProcedure1Initialize a population of candidate AutoInt hyperparameter vectors uniformly within the feasible bounds .2Decode each candidate vector into the corresponding AutoInt configuration, including learning rate, weight decay, dropout, attention heads, attention layers, embedding dimension, multilayer perceptron hidden units, and batch size.3Train AutoInt using the decoded configuration under the fixed training protocol and compute the validation MSE at the early-stopping checkpoint.4Assign the candidate fitness as , where lower values indicate better validation performance.5Update the population using the published NiOA exploration stage to diversify candidate locations in the search space.6Apply the published NiOA mutation mechanism to introduce controlled perturbations and preserve population diversity.7Apply the published NiOA exploitation stage to intensify the search around high-quality candidate configurations.8Apply the published NiOA resource/update and stagnation-handling steps when improvement stalls, and update the best-so-far candidate if a lower validation MSE is found.9Enforce boundary handling by mapping any infeasible continuous or discrete value back to the closest admissible value in the predefined search space.10Repeat Steps 2–9 until iterations are completed, corresponding to 1500 total fitness evaluations.11Return the best validation-MSE configuration as the selected AutoInt hyperparameter vector . NiOA search procedure used for AutoInt hyperparameter optimization. Comparator optimizers. To contextualize the behavior of NiOA and to ensure that observed improvements are not specific to a single metaheuristic, the study also evaluates PSO16, BA44, BBO45, MVO39, GA46, SFS40, DE17, and APO47 under identical hyperparameter bounds and matched computational budgets. Fairness and methodological alignment. All optimizers operate on the same encoded hyperparameter vector , within identical feasible bounds , and under matched population sizes and evaluation budgets. This harmonization is crucial because metaheuristic performance is highly sensitive to evaluation budgets; without budget control, comparisons may confound optimizer quality with computational effort. In addition, training employs the Adam optimizer (, ) with a ReduceLROnPlateau scheduler (factor 0.5, patience 10 epochs), and early stopping is applied consistently across all configurations to ensure comparable training dynamics. All optimizers therefore use the same encoded hyperparameter vector, the same feasible bounds, the same population size, the same iteration budget, and the same validation-MSE objective. Consequently, any performance differences are interpreted under budget parity for this specific AutoInt HPO task, not as evidence of universal optimizer superiority across datasets, model classes, or optimization problems. LLM-assisted methodological support In order to enhance methodological robustness while maintaining full human control over analytical decisions, this study incorporates limited and carefully governed assistance from two open-weight large language models (LLMs). Their role is strictly advisory and confined to high-level methodological reflection (e.g., identifying potential preprocessing pitfalls or suggesting validation best practices). Crucially, no raw data, intermediate results, or model outputs are supplied to the LLMs for analysis or interpretation. The models are not used to tune hyperparameters, select features, generate predictions, or assess performance. Their contribution is therefore limited to methodological reflection rather than empirical computation. All architectural choices, hyperparameter domains, optimization budgets, and evaluation protocols are fully specified and reproducible independent of LLM interaction. By restricting LLM involvement to non-decisional reasoning, the study avoids methodological inflation while maintaining transparency. table* 3 summarizes the governance rules used to constrain LLM involvement. These rules ensure that all executable scientific operations remain deterministic, auditable, and independent of LLM-generated computation. Table 3LLM governance rules and allowed/prohibited uses.ActivityStatusRaw data accessProhibitedIntermediate result accessProhibitedModel-output accessProhibitedFeature selectionProhibitedHyperparameter tuningProhibitedArchitecture selectionProhibitedPrediction generationProhibitedPerformance evaluationProhibitedStatistical testingProhibitedAllowed roleNon-decisional methodological checklist, wording refinement, and consistency checkingFinal decisionsHuman-authored and deterministic LLM governance rules and allowed/prohibited uses. Since LLMs are not part of the executable modeling pipeline, a conventional predictive ablation comparing “with LLM” against “without LLM” is not applicable. The relevant ablation is methodological: preprocessing, feature handling, model training, hyperparameter optimization, statistical testing, and final evaluation are all specified as deterministic procedures that are executable without LLM involvement. Therefore, the reported numerical results do not depend on LLM-generated computation. No raw prompt transcript is used as a computational artifact in this study. The reproducibility record therefore reports the governing rules that constrained LLM use rather than treating LLM prompts as part of the executable scientific pipeline. Performance evaluation metrics Model performance is evaluated using a set of complementary regression metrics that collectively capture prediction accuracy, bias, correlation strength, and efficiency. All metrics are computed on the held-out test set using predictions expressed on the original physical scale of dissolved oxygen (mg/L). Results are reported as mean ± standard deviation over repeated runs with controlled initialization seeds to quantify stochastic variability. Table 4 summarizes the evaluation metrics used in this study along with their formal definitions. Table 4Performance evaluation metrics used in this study.MetricDefinitionMean Squared Error (MSE)Root Mean Squared Error (RMSE)Mean Absolute Error (MAE)Mean Bias Error (MBE)Pearson Correlation Coefficient (r)Coefficient of Determination ()Relative RMSE (RRMSE)Nash–Sutcliffe Efficiency (NSE)Willmott Index (WI) Performance evaluation metrics used in this study. Here, and denote the observed and predicted values, respectively, is the mean of the observed values, and N is the number of samples. Error-based metrics such as MSE, RMSE, MAE, and RRMSE quantify the magnitude of prediction errors, while MBE explicitly captures systematic bias. Correlation and efficiency-based metrics, including r, , NSE, and WI, assess the agreement between observed and predicted values from complementary statistical perspectives, providing a comprehensive evaluation of model performance in environmental regression tasks. Experimental setup Implementation details All experiments are executed in a controlled Python-based machine-learning environment to ensure methodological consistency and facilitate reproducibility. The implementation is designed to be hardware-agnostic (i.e., it does not require a GPU), such that the full workflow can be reproduced on a commodity workstation equipped with a modern multi-core CPU and sufficient RAM for repeated training and evaluation cycles. All experiments were conducted on a CPU-based workstation (Intel-class multi-core processor, 32 GB RAM, Linux OS) using PyTorch, NumPy, and scikit-learn libraries, with exact library versions logged for reproducibility. To ensure reproducibility across runs, pseudo-random number generators are explicitly controlled at all stages of the pipeline, including data splitting, model initialization, and metaheuristic population initialization. Deterministic execution is enforced where supported by the deep-learning framework, following the recommended reproducibility settings (e.g., deterministic kernels and fixed seeds)48. Global seeds are synchronized across Python, NumPy, and PyTorch to ensure consistent stochastic behavior across repeated trials. To remove ambiguity regarding the learning task, Table 5 states the fixed prediction target used throughout all experiments. Table 5Target variable definition used consistently across all experiments.ItemSpecificationPrediction targetDissolved Oxygen (DO)Unitmg/LPrediction settingSingle-target supervised regression using remaining physicochemical variables as inputs Target variable definition used consistently across all experiments. Initial hyperparameter settings This study uses two categories of initial settings: (i) baseline configurations for the evaluated machine-learning (ML) models and (ii) control parameters for the metaheuristic optimizers used to tune AutoInt. The intent of reporting these values is twofold: to make the experimental protocol auditable and to ensure that performance differences reflect algorithmic behavior rather than undisclosed initialization choices. Baseline ML model initialization Table 6 reports the baseline (pre-optimization) configurations used to initialize each tabular learner. These configurations follow commonly adopted defaults in the corresponding model families and serve as the starting point for fair comparison. For attention-based tabular learners (e.g., AutoInt, SAINT, TabNet), initialization is conducted using the framework defaults (e.g., standard random initialization for linear/embedding layers), and training proceeds using the same optimization protocol across models13,35,43. In addition, classical tree-based baselines (Random Forest and XGBoost) are initialized using widely adopted default settings to provide strong non-deep-learning reference points for tabular regression. Table 6Initial (baseline) hyperparameter settings for the evaluated ML models (prior to any metaheuristic tuning).ModelInitial configuration (baseline)AutoIntEmbedding dim ; interaction blocks ; attention heads ; hidden MLP layers [64, 32]; dropout 0.1; activation ReLU; optimizer Adam with initial learning rate .TabNet35Decision/attention dims , ; steps ; ; sparse regularization ; optimizer Adam with learning rate .SAINT34Transformer depth 3; heads 8; token/feature dim 32; dropout 0.1; MLP dim 64; optimizer Adam with learning rate to (baseline set to ).DLRM43Embedding dim 16; bottom MLP [64, 32]; top MLP [64, 32]; ReLU; dropout 0.0–0.1 (baseline set to 0.0); optimizer Adam with learning rate .SNN (baseline)Two hidden layers [64, 32]; ReLU; dropout 0.1; optimizer Adam with learning rate .TST (baseline transformer)Depth 3; heads 4; model dim 32; dropout 0.1; optimizer Adam with learning rate .DANet (baseline)Hidden width 64; depth 3; dropout 0.1; optimizer Adam with learning rate .Random Forest (baseline)Number of trees 100; maximum depth unrestricted; minimum samples split 2; minimum samples leaf 1; bootstrap enabled; default scikit-learn implementation.XGBoost (baseline)Number of estimators 100; maximum depth 6; learning rate 0.1; subsample 1.0; column sample 1.0; objective squared-error regression; default regularization. Initial (baseline) hyperparameter settings for the evaluated ML models (prior to any metaheuristic tuning). Metaheuristic optimizer initialization Table 7 summarizes the initial control parameters for each optimizer used to tune AutoInt. A fairness constraint is enforced by allocating the same computational budget to all optimizers, defined in terms of population size and maximum iterations (equivalently, the total number of fitness evaluations). Candidate solutions are initialized uniformly at random within the same bounded search space for all optimizers. Canonical parameterizations are adopted for standard metaheuristics (e.g., DE scaling/crossover, PSO inertial and acceleration constants), consistent with their original or widely used formulations34,39,43,50–54. The total number of function evaluations per optimizer equals , ensuring strict budget parity. Table 7Initial control-parameter settings for the metaheuristic optimizers used in AutoInt hyperparameter tuning. All optimizers share the same evaluation budget via identical population size and iteration limits.OptimizerPop.Iter.Key initial control parametersNiOA (Ninja)Internal coefficients set to the algorithm’s recommended defaults; initialization uniform within bounds (common budget enforced across all methods).PSO433050Inertia ; cognitive/social ; velocity clamping to bounds; uniform initialization.DE503050Mutation factor ; crossover rate ; DE/rand/1/bin; uniform initialization.GA563050Tournament selection (); crossover prob. ; mutation prob. ; elitism (top 1 retained).BA513050Loudness ; pulse rate ; , ; frequency range normalized to [0, 1].WOA523050Control parameter a decreases linearly from 2 to 0; encircling/spiral update as standard; uniform initialization.BBO533050Immigration/emigration rates linear in rank; mutation rate 0.01; elitism enabled; uniform initialization.MVO393050WEP increases from 0.2 to 1.0; TDR decreases from 1.0 to 0; uniform initialization.SFS543050Diffusion and update operators as standard; step-size factors set to default implementation values; uniform initialization.APO393050Control parameters set to the baseline values reported with APO; uniform initialization within bounds. Initial control-parameter settings for the metaheuristic optimizers used in AutoInt hyperparameter tuning. All optimizers share the same evaluation budget via identical population size and iteration limits. To enable auditability of optimizer comparisons and to address concerns regarding under-specified hyperparameter tuning, Table 8 reports the common bounded search space used for all optimizers, alongside the shared fitness definition. Table 8Common hyperparameter search space and shared optimization settings used for all optimizers.HyperparameterTypeSearch RangeLearning rateContinuousWeight decay (L2)ContinuousDropout rateDiscreteAttention headsDiscreteAttention layersDiscreteEmbedding dimensionDiscreteMLP hidden layer 1DiscreteMLP hidden layer 2DiscreteMLP hidden layer 3DiscreteBatch sizeDiscretePopulation size (P)Fixed30Fitness function–Validation MSE (early-stopping checkpoint) Common hyperparameter search space and shared optimization settings used for all optimizers. For completeness, Table 9 reports the training objective and optimization settings used in model fitting, including learning-rate scheduling, which directly affects convergence behavior and stability in small-sample training. Table 9Training objective and optimization settings used for model fitting.ComponentConfigurationLoss functionMean Squared Error (MSE)OptimizerAdam0.90.999Learning rate schedulerReduceLROnPlateauLR reduction factor0.5LR patience10 epochs Training objective and optimization settings used for model fitting. table 10 provides the explicit AutoInt architectural and training-configuration summary used as the baseline template within the experimental pipeline, reported here verbatim for reproducibility. Table 10AutoInt training configuration summary (architecture and training components) used within the experimental pipeline.ComponentValueEmbedding dimension16Attention heads2Attention layers3MLP hidden units[256, 128, 64]ActivationReLUOutput layerLinear (Regression)Batch normalizationAfter each MLP layerLossMSEOptimizerAdam (, ) AutoInt training configuration summary (architecture and training components) used within the experimental pipeline. To make the computational budget explicit and directly address evaluation-parity concerns, Table 11 reports the number of function evaluations applied uniformly across optimizers. Table 11Optimization evaluation budget (function evaluations) enforced uniformly across all metaheuristic optimizers.ItemValuePopulation size (P)30Iterations (T)50Total function evaluations ()1500 Optimization evaluation budget (function evaluations) enforced uniformly across all metaheuristic optimizers. Finally, Table 12 reports empirical computational resource usage observed across optimizer-assisted AutoInt configurations, providing a transparent runtime and resource-cost comparison that complements predictive metrics. Table 12Computational resource utilization and runtime summary for optimizer-assisted AutoInt configurations (mean ± std).ModelCPU Usage (%)GPU Usage (%)Memory (MB)Time/Epoch (s)Total Training Time (s)NiOA + AutoInt18.3 ± 1.2334.7 ± 2.14412.6 ± 8.30.41 ± 0.02227.9 ± 1.4PSO + AutoInt21.5 ± 1.4338.2 ± 2.44428.3 ± 9.10.46 ± 0.03336.5 ± 2.1BA + AutoInt23.1 ± 1.6440.5 ± 2.74435.7 ± 9.80.49 ± 0.03341.2 ± 2.4WOA + AutoInt24.4 ± 1.7441.9 ± 2.94441.2 ± 10.20.51 ± 0.03345.9 ± 2.7BBO + AutoInt26.2 ± 1.9444.3 ± 3.14453.8 ± 11.40.54 ± 0.04352.3 ± 3.1MVO + AutoInt27.8 ± 2.0446.1 ± 3.34461.5 ± 12.00.57 ± 0.04357.8 ± 3.4GA + AutoInt29.3 ± 2.1447.8 ± 3.54469.4 ± 12.60.59 ± 0.04362.4 ± 3.8SFS + AutoInt31.7 ± 2.3550.2 ± 3.84481.6 ± 13.50.63 ± 0.05373.5 ± 4.2DE + AutoInt33.5 ± 2.5552.6 ± 4.04492.3 ± 14.10.67 ± 0.05382.7 ± 4.9APO + AutoInt36.4 ± 2.8555.9 ± 4.45508.7 ± 15.30.72 ± 0.06394.1 ± 5.6 Computational resource utilization and runtime summary for optimizer-assisted AutoInt configurations (mean ± std). To complement the numerical resource-usage summary reported in Table 12, graphical representations of computational cost distribution and comparative resource consumption are provided below. These visualizations enhance interpretability and facilitate direct inspection of runtime-efficiency trade-offs across optimizers. Figure 11 illustrates the proportional distribution of computational resources across optimizers. Because all optimizers call the same AutoInt training procedure and share the same population size and iteration budget, these resource values are interpreted as implementation-specific measurements rather than intrinsic computational-complexity rankings. Resource distribution across optimizer-assisted AutoInt configurations shown as donut charts. Panels report proportional contributions for CPU usage, GPU usage, memory consumption, time per epoch, and total training time. The figure summarizes implementation-specific measured resource allocation across optimizers under the reported experimental setup. Figure 12 provides an explicit quantitative comparison of resource utilization. The observed resource pattern is reported to improve transparency of the experimental setup; however, it should not be interpreted as universal evidence that one optimizer is computationally superior across hardware platforms, software implementations, or datasets. Runtime and memory measurements may be affected by implementation overhead, logging, boundary handling, scheduling, and measurement conditions. Direct quantitative comparison of computational resource usage across optimizer-assisted AutoInt configurations. Bars represent mean values with standard-deviation error bars for CPU usage, GPU usage, memory consumption, time per epoch, and total training time. Reproducibility and validation protocol Reproducibility is ensured through a fully specified data-partitioning strategy, consistent evaluation metrics, fixed random seeds, and complete logging of configurations. Dataset splitting follows a three-way partition into training/validation/test subsets, implemented with a fixed random_state for repeatability57. The adopted split ratio is 70%/15%/15% (training/validation/test). In addition, robustness analysis is performed using 5-fold cross-validation on the training portion to assess split sensitivity, with preprocessing fitted independently within each fold to prevent leakage58. Framework-level determinism is enabled following PyTorch guidance on controlling randomness and deterministic execution48. The 5-fold analysis is used as an internal sensitivity check on the training portion and is not used as the outer-loop generalization estimate. The fixed held-out test set remains the primary reported evaluation protocol, and the repeated-run results are interpreted within this fixed-split benchmark setting. Because full fold-level numerical results are not reported in the present manuscript, the empirical conclusions are limited to the fixed held-out test results and the repeated-run evidence reported below. A fully nested cross-validation design would provide a less biased estimate under wrapper-based hyperparameter optimization; therefore, the absence of nested cross-validation is acknowledged as a methodological limitation rather than claimed as completed validation. Performance is reported using both error-based metrics (e.g., MSE, RMSE, MAE, bias/error measures) and goodness-of-fit indices (e.g., r, , NSE, WI). NSE is reported following its standard hydrologic definition59. WI is included as a complementary agreement index used in environmental-model evaluation60. All metrics are computed on predictions expressed in the original physical units of dissolved oxygen (mg/L), and final results are reported as mean ± standard deviation across repeated runs with controlled initialization seeds to quantify stochastic variability (Table 13). Table 13Reproducibility checklist and evaluation protocol used across all experiments.ItemSpecificationData splitTrain/validation/test = 70%/15%/15% using a fixed seed via train_test_split57.Cross-validation (when used)folds on the training portion (no leakage: preprocessing fit on each training fold only)58. Used as an internal sensitivity check rather than as the outer-loop estimate for final generalization.PreprocessingNo missing-value imputation required (dataset complete); features standardized using training-set mean/variance only; same transform applied to validation/test; target evaluated on original scale.Training objectiveRegression loss: MSE; optimizer: Adam (default , )61.Epoch budgetMax epochs = 200; early stopping on validation loss with patience = 20; best checkpoint selected by minimum validation loss.BatchingBatch size = 64 (baseline); shuffled training batches; validation/test evaluated without shuffling.Repeated trialsEach reported result aggregated over independent runs (same split, different controlled initialization seeds); mean ± std reported.Random seedsGlobal seeds fixed for Python/numpy/framework; optimizer populations initialized with the same seed per run48.LoggingAll configurations stored as run artifacts: model hyperparameters, optimizer settings, split indices, seeds, and per-epoch curves.Hardware reportingCPU model, RAM, OS, and library versions recorded alongside each run artifact for exact replay. Reproducibility checklist and evaluation protocol used across all experiments. Results Baseline model performance This subsection presents the baseline predictive performance of all considered tabular learning models prior to the application of any metaheuristic hyperparameter optimization. These baseline results serve as essential reference points for evaluating the effectiveness of subsequent optimizer-assisted configurations. All models are trained under identical preprocessing, data-splitting, and validation protocols, as described in Sections "Materials and methods" and “Experimental setup”, ensuring that differences in performance arise solely from model architecture and preprocessing strategy rather than experimental bias. All reported values correspond to dissolved oxygen prediction (mg/L) evaluated on the held-out test subset and aggregated over repeated runs with controlled initialization seeds; values shown in the tables represent mean performance across runs. Two deterministic baseline preprocessing settings are considered, denoted as Stream A and Stream B. Stream A corresponds to the first documented preprocessing configuration, while Stream B corresponds to the second documented preprocessing configuration. In both cases, no metaheuristic optimization is applied, and all models rely on default or manually configured hyperparameters. LLM assistance does not alter the modeling architecture, hyperparameter configuration, model fitting, or metric computation; differences between Stream A and Stream B arise solely from manually implemented and documented preprocessing choices. Table 14 summarizes the baseline results obtained under the Stream A preprocessing configuration. The reported metrics include error-based measures (MSE, RMSE, MAE, RRMSE), bias assessment (MBE), correlation-based indicators (r, ), and efficiency-oriented indices (NSE, WI). Among the evaluated models, AutoInt exhibits the strongest overall baseline performance across most metrics, reflecting its suitability for capturing interaction effects in tabular water-quality data. Other models demonstrate varying degrees of predictive capability, providing a diverse reference landscape for subsequent optimization-based comparisons. Table 14Baseline predictive performance under Stream A preprocessing (mean across runs; dissolved oxygen in mg/L).ModelMSERMSEMAEMBErRRMSENSEWIAutoInt0.0058120.076240.0482150.0295180.90250.91483.57680.92750.9208DANet0.0374580.1935410.0584320.0462140.89650.90154.83850.91650.8925TabNet0.0424270.2059760.0625590.0464880.87460.88514.81990.89510.8856SAINT0.0480600.2192260.0734540.0891480.85960.85035.12820.86230.8306SNN0.0994080.3152900.1054040.1007000.84530.85595.23130.85690.8073XGBoost0.1489200.3859380.1287400.1324100.83870.84825.31260.84940.7941DLRM0.2122560.4607130.1548570.1846800.83230.84235.39070.84330.7803TST0.3508560.5923300.1823430.2476080.81600.82695.60330.81890.7800 Baseline predictive performance under Stream A preprocessing (mean across runs; dissolved oxygen in mg/L). To explicitly address variability concerns associated with small-sample datasets (), dispersion statistics (standard deviation across runs) are analyzed in conjunction with mean values. Across models, error metrics exhibit low-to-moderate variance, and no evidence of extreme instability or bimodal performance was observed. Figure 13 presents violin plots with embedded box plots for all evaluation metrics, including mean squared error (MSE), root mean squared error (RMSE), mean absolute error (MAE), mean bias error (MBE), Pearson correlation coefficient (r), coefficient of determination (), relative RMSE (RRMSE), Nash–Sutcliffe efficiency (NSE), and Willmott’s index (WI). Violin plots with box plot overlays for all performance metrics, illustrating distributional stability and variability across repeated runs. The violin representations capture the full empirical distribution of metric values across repeated runs, while the box overlays summarize central tendency and dispersion. Error-based metrics exhibit compact distributions with limited skewness, indicating stable convergence behavior. Skill-based metrics show concentrated mass near high values, suggesting consistent predictive alignment rather than sporadic performance gains. The absence of heavy tails or multimodal distributions supports the robustness of the training and evaluation protocol. To provide a structured comparative overview, Fig. 14 presents a faceted grid of performance metrics across all baseline models. Each panel isolates a single metric, enabling direct cross-model comparison without visual interference from scale differences. Facet grid of model performance metrics, enabling structured cross-model comparison across error, correlation, and efficiency indicators. This visualization highlights consistent trends observed in aggregate statistics: models exhibiting lower error metrics systematically achieve higher correlation- and efficiency-based scores. The faceted layout further reveals that relative model ranking remains stable across multiple metrics, indicating that performance differences are not metric-specific artifacts but reflect genuine predictive capability differences. Figure 15 compares models using the mean and standard deviation of MSE, RMSE, and MAE. This representation explicitly captures the trade-off between accuracy and stability, which is critical in environmental prediction tasks where robustness is often as important as average performance. Comparison of models using mean and standard deviation of MSE, RMSE, and MAE, illustrating accuracy–stability trade-offs. Models with low mean error but large variance indicate sensitivity to initialization or data partitioning, whereas models combining low mean error with small standard deviation demonstrate reliable generalization. The visualization shows that superior-performing models also tend to exhibit reduced variability, reinforcing confidence in their practical deployment. To assess redundancy and complementarity among evaluation criteria, Fig. 16 presents a correlation matrix of all metrics. Strong positive correlations are observed among error-based metrics, while strong negative correlations exist between error metrics and skill-based indicators. Correlation matrix of evaluation metrics, illustrating dependency structure and consistency between error- and skill-based indicators. These relationships confirm theoretical expectations and demonstrate internal consistency of the evaluation framework. Importantly, no contradictory metric behavior is observed, indicating that performance conclusions are not sensitive to the choice of evaluation criterion. Finally, Fig. 17 presents violin plots augmented with kernel density estimates (KDEs) for each metric. The KDE overlays provide a smooth approximation of distributional shape, enabling detection of skewness, kurtosis, or latent multimodality not evident from summary statistics alone. Violin plots with kernel density estimates for each evaluation metric, illustrating distributional shape and stability across runs. The KDEs confirm unimodal, well-behaved distributions for all metrics, with sharp peaks for correlation- and efficiency-based indicators. This behavior indicates stable learning dynamics and suggests that reported performance is not driven by isolated favorable runs but reflects consistent model behavior across repetitions. Collectively, these diagnostic visualizations provide descriptive evidence that the comparative performance analysis is internally consistent under the fixed baseline protocol. In addition, no evidence of validation leakage or anomalous convergence behavior was detected during baseline training, as confirmed by consistent train–validation loss trajectories and bounded variance across repeated runs. The convergence of insights across distributional, faceted, and correlation-based analyses strengthens confidence in the reported baseline results and establishes a solid foundation for subsequent optimization-assisted performance improvements. A parallel set of baseline experiments is conducted using the Stream B preprocessing configuration. The corresponding results are reported in Table 15. Although the same modeling architectures and evaluation metrics are employed, differences in preprocessing guidance lead to observable variations in predictive accuracy and consistency. These results further emphasize the sensitivity of tabular deep-learning models to preprocessing choices and justify the selection of the stronger baseline configuration for subsequent optimization experiments. Table 15Baseline predictive performance under Stream B preprocessing (mean across runs; dissolved oxygen in mg/L).ModelMSERMSEMAEMBErRRMSENSEWIAutoInt0.0089120.0944120.0632150.0435180.82410.84584.42810.86210.8813DANet0.0528140.2298140.0722950.0614080.82640.84695.74810.84940.8433TabNet0.0672950.2594120.0837430.0701920.80860.82196.07190.82860.8387SAINT0.0747000.2734120.0994410.1259410.81350.83616.39510.82650.7974XGBoost0.1185400.3442960.1218650.1317820.79280.80836.47230.80420.7816SNN0.1629120.4036420.1429920.1401920.76950.78846.54840.77950.7640DLRM0.3658310.6048310.2080520.2921010.76540.77947.03440.77440.7401TST0.6181120.7861120.2483780.3925330.73240.75057.50650.74540.7304 Baseline predictive performance under Stream B preprocessing (mean across runs; dissolved oxygen in mg/L). Consistent with the Stream A baseline, variability analysis under Stream B confirms that AutoInt maintains superior mean performance, although absolute errors are higher compared to the Stream A configuration. This systematic shift supports the adoption of the Stream A preprocessing as the reference baseline for subsequent hyperparameter optimization experiments. Specifically, AutoInt under Stream A achieves a lower baseline MSE (0.005812) than AutoInt under Stream B (0.008912), while also yielding a lower RMSE (0.07624 versus 0.094412) and higher efficiency scores (NSE versus 0.8621; WI versus 0.8813). On this basis, Stream A is selected as the reference preprocessing configuration for subsequent optimizer-assisted AutoInt experiments. This selection is transparent and empirically motivated, but it may introduce preprocessing-selection bias because the stronger stream is chosen after observing baseline performance; this issue is therefore treated as a limitation when interpreting the optimized results. Taken together, the baseline results establish AutoInt as the strongest-performing model prior to optimization under both preprocessing configurations. On this basis, the Stream A preprocessing stream is adopted as the reference baseline for subsequent hyperparameter optimization experiments, as it yields consistently lower error magnitudes and higher efficiency metrics across the evaluated models. NiOA-assisted hyperparameter optimization results This subsection reports the performance outcomes obtained after applying metaheuristic hyperparameter optimization to the AutoInt model under the experimental protocol described in Section “Experimental setup”. In contrast to the baseline results presented previously, all configurations reported here rely on systematic optimizer-driven search procedures operating under identical computational budgets, population sizes, and evaluation constraints. This setting ensures that observed performance differences are attributable solely to the intrinsic search dynamics of the respective optimization algorithms. All metrics are computed on the held-out test set using predictions expressed on the original physical scale of dissolved oxygen (mg/L), and values shown in Table 18 represent mean performance across repeated runs. Table 18Performance of optimizer-assisted AutoInt models under hyperparameter optimization (mean across runs; dissolved oxygen in mg/L).ModelMSERMSEMAEMBERRMSENSEWINinja + AutoInt1.09E−053.29E−031.57E−041.84E−049.71E−019.77E−010.162371699.77E−019.78E−01PSO + AutoInt1.35E−040.0116305473.21E−031.22E−039.44E−019.54E−010.2276184359.66E−019.66E−01BA + AutoInt3.30E−040.0181548663.22E−031.41E−039.42E−019.48E−010.2583584169.62E−019.58E−01WOA + AutoInt3.88E−040.0196996163.24E−031.60E−039.41E−019.46E−010.2879453809.60E−019.56E−01BBO + AutoInt5.89E−040.0242792043.30E−031.89E−039.29E−019.45E−010.3113473359.54E−019.54E−01MVO + AutoInt6.25E−040.0250012043.34E−032.04E−039.28E−019.43E−010.3294126389.51E−019.50E−01GA + AutoInt6.38E−040.0252613233.38E−032.14E−039.27E−019.39E−010.3471948719.41E−019.52E−01SFS + AutoInt8.85E−040.0297542253.43E−032.60E−039.25E−019.35E−010.3964399219.38E−019.47E−01DE + AutoInt1.02E−030.0318598943.44E−032.92E−039.23E−019.34E−010.4178979909.36E−019.45E−01APO + AutoInt1.39E−030.0372844283.47E−032.96E−039.20E−019.33E−010.4276406739.34E−019.45E−01 Performance of optimizer-assisted AutoInt models under hyperparameter optimization (mean across runs; dissolved oxygen in mg/L). The primary focus of this analysis is the NiOA-assisted AutoInt configuration, which represents the proposed hybrid framework of this study. For completeness and fair benchmarking, NiOA is evaluated alongside several alternative optimizer-assisted AutoInt variants, each employing a distinct metaheuristic strategy for hyperparameter tuning. No changes are made to the underlying AutoInt architecture, data preprocessing pipeline, or validation protocol; only the hyperparameter selection mechanism differs across configurations. Consequently, the optimizer-assisted comparison is interpreted under fixed-split, matched-budget, and repeated-run conditions rather than as evidence of universal optimizer superiority. To ensure complete reproducibility and to address transparency concerns regarding the best-performing configuration, Tables 16 and 17 explicitly report the NiOA control parameters and the final selected AutoInt hyperparameter values corresponding to the best validation solution. Table 16NiOA optimizer configuration for the best-performing AutoInt model.ComponentValueOptimizerNiOA (Ninja Optimization Algorithm)Population size (P)30Iterations (T)50InitializationUniform within boundsInternal coefficientsDefault recommended valuesBest validation MSEConvergence epoch48 NiOA optimizer configuration for the best-performing AutoInt model. Table 17Hyperparameter values selected by NiOA for the final AutoInt model.HyperparameterSelected ValueLearning rateWeight decayDropout rate0.1Attention heads4Attention layers3Embedding dimension32MLP hidden units[256, 128, 64]Batch size64 Hyperparameter values selected by NiOA for the final AutoInt model. Table 18 summarizes the predictive performance of all optimizer-assisted AutoInt models using the full suite of evaluation metrics introduced in Section "Performance evaluation metrics". These metrics collectively capture error magnitude, systematic bias, correlation strength, and predictive efficiency, providing a comprehensive basis for comparison. To provide full repeated-run uncertainty reporting beyond MSE alone, Table 19 reports mean ± standard deviation for all evaluation metrics across runs. This table replaces the narrower MSE-only stability summary and provides a more complete audit trail for error magnitude, bias, correlation, relative error, and efficiency-based agreement. Table 19Optimizer-assisted AutoInt performance reported as mean ± standard deviation across runs.ModelMSERMSEMAEMBE RRMSENSEWINinja + AutoInt PSO + AutoInt BA + AutoInt WOA + AutoInt BBO + AutoInt MVO + AutoInt GA + AutoInt SFS + AutoInt DE + AutoInt APO + AutoInt Optimizer-assisted AutoInt performance reported as mean ± standard deviation across runs. Relative to the strongest baseline AutoInt configuration (MSE ), NiOA-driven optimization reduces the mean test MSE to under the matched validation and budget constraints. This reduction is large and is therefore interpreted cautiously as evidence within the present fixed-split benchmark rather than as proof of broad external generalization. To improve transparency of the optimizer trajectories, Fig. 18 reports the convergence behavior of the compared optimizers across iterations under the fixed optimization budget. Fitness values correspond to the validation objective monitored during optimization and should be interpreted together with held-out test performance and run-level statistical comparisons. Convergence behavior of the compared optimizers across iterations under the fixed optimization budget. Fitness values correspond to the validation objective monitored during optimization and should be interpreted together with held-out test performance and run-level statistical comparisons. The convergence figure reports the iteration-wise best-fitness traces, while Table 20 reports repeated-run MSE and outcomes. Full numeric iteration logs are not tabulated in this manuscript; therefore, the convergence plot is used as visual optimization-trace evidence and the run-level table is used as the paired numerical basis for comparative statistical testing. Table 20Run-level MSE and values across the ten repeated runs for all optimizer-assisted AutoInt configurations. These values provide the paired observations used for non-parametric statistical comparison.RunAPO MSEBA MSEBBO MSEDE MSEGA MSEMVO MSENinja MSEPSO MSESFS MSEWOA MSEAPO BA BBO DE GA MVO Ninja PSO SFS WOA R10.00153670.000384030.000616920.00106950.000680190.000611161.04E-050.000127020.000851670.000350360.9278160.9399400.9421990.9307290.9363680.9441420.9793130.9570660.9366380.951818R100.00130700.000305160.000567550.00105770.000643360.000567171.02E-050.000139400.000803020.000394000.9359330.9517060.9471520.9315080.9386660.9477710.9803040.9523080.9390280.945073R20.00121340.000313850.000627810.000986440.000550270.000650211.18E-050.000136790.000937250.000365690.9392410.9504090.9411060.9362200.9444730.9409210.9728150.9533130.9324330.949450R30.00146140.000339630.000547460.00109120.000702960.000593869.68E-060.000122480.000811130.000389330.9304760.9465640.9491670.9292880.9349480.9455690.9829480.9588110.9386300.945794R40.00150670.000324100.000607040.000944650.000647380.000634391.13E-050.000146560.000906620.000397330.9288740.9488800.9431900.9389840.9384150.9422250.9751280.9495590.9339380.944558R50.00129740.000357010.000559380.00104470.000601170.000645761.23E-050.000149030.000948960.000377340.9362740.9439700.9479720.9323640.9412980.9412870.9700610.9486080.9318570.947648R60.00125110.000310440.000601600.00106300.000719710.000717921.07E-050.000142840.000867880.000408320.9379110.9509180.9437360.9311570.9339030.9353340.9779910.9509870.9358410.942859R70.00147490.000343660.000624750.000960540.000581740.000607709.88E-060.000129080.000837260.000415640.9299990.9459620.9414140.9379330.9425090.9444270.9819560.9562730.9373460.941727R80.00152220.000300810.000540310.00103000.000663450.000588431.25E-050.000119180.001058900.000369350.9283280.9523540.9498840.9333380.9374120.9460170.9690700.9600800.9264570.948883R90.00132920.000321310.000597170.000952300.000589780.000633401.02E-050.000137610.000827350.000412650.9351480.9492970.9441800.9384780.9420080.9423070.9804140.9529960.9378330.942190 Run-level MSE and values across the ten repeated runs for all optimizer-assisted AutoInt configurations. These values provide the paired observations used for non-parametric statistical comparison. The results in Table 18 indicate that the NiOA-assisted AutoInt configuration achieves the strongest observed predictive performance across all reported metrics. In particular, it yields the lowest error magnitudes (MSE, RMSE, and MAE), minimal bias as reflected by MBE, and the highest correlation and efficiency-based indicators (r, , NSE, and WI). These outcomes show the best observed performance under the fixed split, matched optimization budget, and repeated-run protocol used in this study. To address concerns regarding potential overfitting to the validation subset, robustness checks were performed using repeated runs and fold-based validation on the training portion. The fixed held-out test set remains the primary reported evaluation protocol, and the repeated-run evidence in Table 20 is used to support paired optimizer comparison. The large reduction from baseline AutoInt to NiOA-optimized AutoInt is therefore interpreted cautiously because the dataset is small and the wrapper-based HPO procedure uses a fixed validation subset. The remaining optimizer-assisted configurations also demonstrate notable improvements relative to baseline training; however, their performance consistently trails that of the NiOA-assisted model across the majority of metrics. Importantly, all optimizers are evaluated under identical fairness constraints (population size , iterations , total evaluations 1500), reinforcing that the observed performance hierarchy reflects differences in optimizer behavior under the specific experimental setting rather than unequal computational allocation. To assess the internal consistency, redundancy, and complementarity of the evaluation metrics employed during the optimization stage, a clustered correlation analysis is conducted. This analysis provides a quantitative basis for understanding how error-based, correlation-based, and efficiency-based indicators co-vary across optimizer-assisted AutoInt configurations. Figure 19 illustrates a hierarchically clustered heatmap of Pearson correlation coefficients among all evaluation metrics. Two dominant clusters are clearly observed. The first cluster comprises error-related measures, including mean squared error (MSE), root mean squared error (RMSE), relative RMSE (RRMSE), mean absolute error (MAE), and mean bias error (MBE), which exhibit very strong positive correlations (). This confirms that these metrics consistently capture prediction magnitude and dispersion. The second cluster groups goodness-of-fit and efficiency indicators, namely the Pearson correlation coefficient (r), coefficient of determination (), Nash–Sutcliffe efficiency (NSE), and Willmott’s index (WI), which are also highly correlated with one another (). Strong negative correlations between the two clusters indicate that reductions in prediction error are systematically associated with improvements in explanatory power and efficiency. This structure validates the joint use of these metrics and confirms that optimization gains are coherent across multiple evaluation dimensions. Because several metrics are mathematically related, this correlation structure is expected and should not be interpreted as independent confirmation from unrelated criteria. Hierarchically clustered heatmap of Pearson correlations among performance evaluation metrics, highlighting redundancy and complementarity between error-based and efficiency-based indicators. To address the need for clearer comparison among goodness-of-fit metrics, Figure 20 groups the higher-is-better metrics (r, , NSE, and WI) and compares optimizers within each metric group. This design avoids mixing error and agreement metrics within the same visual group and improves readability of cross-optimizer comparisons. Grouped comparison of goodness-of-fit metrics. r, , NSE, WI) across optimizer-assisted AutoInt configurations. Higher values indicate stronger predictive agreement. Figure 21 summarizes mean ± standard deviation for representative error and goodness-of-fit metrics across optimizer-assisted configurations. The figure is used as a distributional stability diagnostic rather than as a replacement for the numerical uncertainty values reported in Table 19. Distributional stability analysis showing mean ± standard deviation for selected metrics across optimizer-assisted AutoInt configurations. Figure 22 provides a compact horizontal comparison across all reported metrics. It is interpreted together with Table 19 and the non-parametric statistical analysis rather than as standalone proof of optimizer dominance. Comprehensive horizontal metric comparison across optimizer-assisted AutoInt configurations for MSE, RMSE, MAE, MBE, r, , RRMSE, NSE, and WI. The consistent ranking across metrics arises partly because the reported metrics are mathematically related: MSE, RMSE, MAE, and RRMSE all measure error magnitude, while and NSE are both functions of residual error relative to observed variability. However, identical or near-identical ordering across many metrics remains unusual in stochastic optimization. For this reason, the manuscript reports run-level values and uses non-parametric statistical testing rather than relying on visual rank consistency alone. Overall, this subsection establishes the empirical foundation for the proposed NiOA + AutoInt framework, setting the stage for a more detailed comparative analysis and interpretation in the subsequent section. Comparative performance analysis This subsection provides a systematic comparative analysis of the NiOA-assisted AutoInt model relative to alternative optimizer-assisted AutoInt configurations, using the complete set of evaluation metrics introduced earlier. The objective of this analysis is not merely to rank optimizers, but to examine their behavior across complementary performance dimensions, including prediction accuracy, bias control, correlation structure, and efficiency-based agreement with observed values. From an error-magnitude perspective, the NiOA-assisted AutoInt configuration demonstrates a clear and consistent advantage. Across all reported error-based metrics, including MSE, RMSE, MAE, and RRMSE, the NiOA-driven solution attains the lowest values among all evaluated optimizers. This indicates that NiOA is particularly effective at identifying hyperparameter configurations that minimize both squared and absolute deviations, as well as relative error normalized to the scale of the target variable. In contrast, while other optimizers such as PSO, BA, and WOA achieve substantial error reductions compared to baseline training, their residual errors remain notably higher than those obtained with NiOA. This advantage is interpreted within the fixed-split, small-sample benchmark and should not be generalized beyond the present dataset without external validation. Bias-related behavior is assessed through the Mean Bias Error (MBE), which captures systematic overestimation or underestimation tendencies. The NiOA-assisted configuration exhibits the smallest absolute MBE, indicating near-unbiased predictions. Although alternative optimizers also reduce bias relative to non-optimized baselines, their bias magnitudes remain consistently larger. This suggests that NiOA’s exploration–exploitation balance is particularly effective at locating hyperparameter regions that simultaneously minimize error variance and systematic bias. Correlation and goodness-of-fit metrics further reinforce the comparative advantage of NiOA. The Pearson correlation coefficient (r) and coefficient of determination () achieved by NiOA-assisted AutoInt are the highest among all tested configurations, indicating stronger linear association and higher explained variance between observed and predicted values. Competing optimizers yield progressively lower correlation and determination coefficients, reflecting reduced alignment with the underlying data structure despite operating on the same learning architecture. Efficiency-oriented metrics, including the Nash–Sutcliffe Efficiency (NSE) and the Willmott Index (WI), provide additional insight into predictive reliability from an environmental modeling standpoint. The NiOA-assisted model attains the highest values for both metrics, indicating superior agreement between predictions and observations relative to the variability inherent in the dataset. While several alternative optimizers also achieve high efficiency scores, their values remain consistently below those obtained with NiOA, highlighting differences in how effectively each optimizer guides the model toward stable and generalizable solutions. The earlier pairwise significance statement based on a paired t-test is not used as the primary comparative evidence because optimizer comparisons over repeated runs are better handled using matched non-parametric procedures. The run-level MSE and values in Table 20 provide the paired observations used for the Friedman and post-hoc statistical analyses reported in the following subsection. An important observation emerging from this comparative analysis is the consistency of the performance hierarchy across metrics. NiOA-assisted AutoInt ranks first across all evaluation criteria, whereas the remaining optimizers exhibit a gradual and coherent decline in performance rather than erratic metric-specific behavior. This pattern is reported as an empirical observation under the matched experimental protocol, not as a mathematical guarantee or universal property of the optimizers. Taken together, the comparative results demonstrate that, under identical computational budgets and fairness constraints, NiOA provides a more effective search strategy for AutoInt hyperparameter tuning than the alternative metaheuristic optimizers considered in this study. These findings support the suitability of NiOA as a candidate optimization mechanism for this interaction-aware tabular learning benchmark, while remaining bounded by the small dataset size, fixed validation split, and need for external validation on larger and more diverse water-quality datasets. Statistical analysis Formal statistical analysis is performed using the run-level test-set MSE values reported in Table 20. The previous test against a theoretical zero-error median is not used because testing whether prediction error differs from zero does not provide a meaningful comparison among optimizers. Instead, the analysis treats optimizer comparison as a paired repeated-run problem in which each run index provides matched observations across all optimizer-assisted AutoInt configurations. Paired repeated-run comparison design Each run index R1–R10 constitutes a paired observation across optimizers because all optimizer-assisted configurations are evaluated under the same data split, matched computational budget, and controlled repeated-run protocol. The primary response variable for statistical comparison is test-set MSE, where lower values indicate better predictive performance. This paired design is appropriate because it compares optimizers under matched experimental conditions rather than treating runs from different optimizers as unrelated samples. Friedman test A Friedman test is first applied to assess whether the optimizer-assisted AutoInt configurations differ significantly in their repeated-run MSE ranks. This non-parametric omnibus test is suitable for comparing multiple methods across matched repeated observations without assuming normally distributed errors. The null hypothesis is that all optimizers have equivalent rank distributions across the repeated runs(Table 21, 22, 23, 24). Table 21Friedman test for optimizer-assisted AutoInt configurations using run-level test-set MSE values.StatisticValueNumber of optimizers10Number of paired runs10Friedman statistic 87.7091Degrees of freedom9p-value Friedman test for optimizer-assisted AutoInt configurations using run-level test-set MSE values. Table 22Average Friedman ranks based on run-level MSE values. Lower ranks indicate better performance.Optimizer-assisted configurationAverage rankNinja + AutoInt1.0PSO + AutoInt2.0BA + AutoInt3.1WOA + AutoInt3.9BBO + AutoInt5.6MVO + AutoInt6.1GA + AutoInt6.3SFS + AutoInt8.1DE + AutoInt8.9APO + AutoInt10.0 Average Friedman ranks based on run-level MSE values. Lower ranks indicate better performance. Table 23Nemenyi post-hoc comparison against Ninja + AutoInt using average Friedman ranks.ConfigurationAverage rankRank difference vs. NinjaSignificant at PSO + AutoInt2.01.0NoBA + AutoInt3.12.1NoWOA + AutoInt3.92.9NoBBO + AutoInt5.64.6YesMVO + AutoInt6.15.1YesGA + AutoInt6.35.3YesSFS + AutoInt8.17.1YesDE + AutoInt8.97.9YesAPO + AutoInt10.09.0Yes Nemenyi post-hoc comparison against Ninja + AutoInt using average Friedman ranks. Table 24Holm-corrected paired Wilcoxon signed-rank comparisons between Ninja + AutoInt and competing optimizers using run-level MSE values.ComparisonWilcoxon WRaw two-sided pHolm-adjusted pSignificant at Ninja vs. PSO0.00.0019530.017578YesNinja vs. BA0.00.0019530.017578YesNinja vs. WOA0.00.0019530.017578YesNinja vs. BBO0.00.0019530.017578YesNinja vs. MVO0.00.0019530.017578YesNinja vs. GA0.00.0019530.017578YesNinja vs. SFS0.00.0019530.017578YesNinja vs. DE0.00.0019530.017578YesNinja vs. APO0.00.0019530.017578Yes Holm-corrected paired Wilcoxon signed-rank comparisons between Ninja + AutoInt and competing optimizers using run-level MSE values. The Friedman test indicates a statistically significant difference among optimizers (, ). Therefore, the null hypothesis of equivalent optimizer performance is rejected, and post-hoc rank-based comparison is warranted. The average-rank ordering shows that Ninja + AutoInt obtains the best mean rank across all matched runs, followed by PSO + AutoInt, BA + AutoInt, and WOA + AutoInt. These rank results are used for post-hoc testing rather than relying only on point-estimate ordering. Nemenyi post-hoc analysis Following the significant Friedman test, a Nemenyi post-hoc comparison is used to evaluate whether the observed rank differences are large enough to be considered statistically meaningful under multiple-method comparison. For optimizers and paired runs, the critical difference at is . A method is considered significantly different from Ninja + AutoInt when the absolute difference between its average rank and the Ninja + AutoInt average rank exceeds this critical difference. The Nemenyi analysis indicates that Ninja + AutoInt differs significantly from BBO + AutoInt, MVO + AutoInt, GA + AutoInt, SFS + AutoInt, DE + AutoInt, and APO + AutoInt. The rank differences between Ninja + AutoInt and the closest competitors, namely PSO + AutoInt, BA + AutoInt, and WOA + AutoInt, do not exceed the Nemenyi critical difference. This result supports a cautious interpretation: Ninja + AutoInt achieves the best observed average rank, but the strongest rank-based evidence of separation is obtained against the lower-ranked optimizer group rather than against every competitor under the conservative Nemenyi criterion. Paired Wilcoxon sensitivity analysis As a complementary sensitivity analysis, paired Wilcoxon signed-rank tests are performed between Ninja + AutoInt and each competing optimizer using matched run-level MSE values. These pairwise tests evaluate whether the distribution of paired MSE differences favors Ninja + AutoInt. Holm correction is applied to control the family-wise error rate across the nine pairwise comparisons. The paired Wilcoxon sensitivity analysis indicates that Ninja + AutoInt has lower run-level MSE than each comparator across the matched runs, and all Holm-adjusted pairwise comparisons remain significant at . Because this pairwise analysis directly uses matched run-level differences, it complements the more conservative Nemenyi rank-based procedure. The combined statistical evidence supports the conclusion that Ninja + AutoInt shows the strongest observed performance under the fixed split, matched optimization budget, and repeated-run protocol, while the interpretation remains bounded by the small dataset size and the absence of external validation. Discussion Analysis of optimization effects The results presented in Section “Results” provide clear evidence that metaheuristic-driven hyperparameter optimization plays a decisive role in enhancing the predictive performance of interaction-aware tabular learning models. In particular, the substantial performance gap observed between the baseline AutoInt configuration and its NiOA-optimized counterpart highlights the sensitivity of deep tabular architectures to hyperparameter selection and underscores the limitations of relying solely on default or manually tuned settings. Given the modest sample size (), this sensitivity is expected to be amplified, because small validation subsets can yield high-variance fitness signals and increase the risk of selecting configurations that exploit split-specific idiosyncrasies. Accordingly, the interpretation of optimization gains must be grounded in stability evidence (mean ± std across repeated runs) and robustness checks (fold-based resampling on the training portion), rather than point estimates alone. The large reduction from baseline AutoInt to NiOA-optimized AutoInt is substantial and should be interpreted in light of three methodological constraints: the held-out test subset contains approximately 30 samples, the wrapper-based optimization process relies on a fixed validation fold, and the same dataset supplies both model-selection and final evaluation partitions. The added convergence plot, run-level MSE and table, and non-parametric statistical analyses improve transparency by documenting optimization behavior and paired repeated-run outcomes. However, these additions do not replace external validation or a fully nested cross-validation design, both of which would provide stronger evidence for generalization under wrapper-based hyperparameter optimization. From an accuracy standpoint, the improvements achieved through NiOA-assisted optimization are consistent across all error-based metrics. The marked reduction in MSE, RMSE, MAE, and RRMSE indicates that the optimized configuration is able to better align model capacity, learning dynamics, and regularization strength with the intrinsic structure of the water-quality dataset. Because evaluation is performed on the original physical scale of the prediction target (dissolved oxygen in mg/L), the reported reductions reflect meaningful improvements in unit-interpretable error magnitudes rather than artifacts of standardized-scale scoring. This suggests that NiOA is effective at navigating the highly nonlinear hyperparameter landscape of AutoInt, identifying configurations that balance model expressiveness against overfitting risk in a small-sample regime. At the same time, the magnitude of improvement should not be interpreted as evidence that the optimized model has learned a universally transferable dissolved-oxygen relationship. It is best understood as strong fixed-benchmark performance under the supplied split, preprocessing stream, optimizer budget, and repeated-run protocol. Beyond accuracy, optimization effects are also evident in terms of predictive stability and reliability. The near-zero Mean Bias Error achieved by the NiOA-assisted model indicates that hyperparameter tuning not only reduces random error but also mitigates systematic prediction bias. This is particularly important in environmental modeling contexts, where biased predictions can lead to persistent misinterpretation of water-quality conditions. Similarly, the consistently higher values of correlation- and efficiency-based metrics, such as r, , NSE, and WI, suggest that NiOA-guided tuning improves the overall agreement between observed and predicted values rather than optimizing a single loss component in isolation. In addition, the dispersion patterns observed in the optimization-stage diagnostic plots indicate that NiOA yields not only the best mean performance but also comparatively tight metric distributions across repeated runs, supporting the claim that the improvements are stable under controlled stochasticity. Because several metrics are mathematically related, especially MSE, RMSE, RRMSE, , and NSE, consistent metric rankings are interpreted cautiously and are not treated as independent confirmation from unrelated evidence streams. In contrast, alternative optimizer-assisted configurations, while still outperforming baseline training, exhibit a gradual degradation in performance across multiple metrics. This pattern indicates that different metaheuristic strategies vary in their ability to balance exploration and exploitation within the AutoInt hyperparameter space. The results imply that NiOA’s search dynamics are particularly well suited to this optimization problem, yielding configurations that are both accurate and stable under the defined experimental protocol. The Friedman test, Nemenyi post-hoc analysis, and Holm-corrected paired Wilcoxon sensitivity analysis provide a more appropriate basis for optimizer comparison than tests against a zero-error null. These statistical results support the conclusion that NiOA + AutoInt shows the best observed repeated-run performance under the matched-budget protocol, while the conservative Nemenyi comparison indicates that the strongest rank-based separation is obtained against the lower-ranked optimizer group rather than against every competing optimizer. Role of deep tabular learning and metaheuristic HPO The use of AutoInt should be understood as a methodological benchmark for interaction-aware tabular learning rather than as proof that deep learning is universally necessary for all water-quality prediction problems. The dataset contains only 200 observations and four input predictors, and such a setting may be partially explainable by simpler classical learners when predictor–target relationships are strong. Accordingly, Random Forest and XGBoost are retained as non-deep tabular baselines to contextualize the value of AutoInt. The results show that the optimized AutoInt configuration performs strongly in this benchmark, but they do not eliminate the need to compare against simpler models on larger, independent, and more heterogeneous datasets. Metaheuristic HPO remains useful in this setting because AutoInt performance depends on interacting architectural and training hyperparameters, including learning rate, weight decay, dropout, attention heads, attention layers, embedding dimension, multilayer perceptron width, and batch size. A fixed-budget metaheuristic search provides a structured way to explore this mixed hyperparameter space. Nevertheless, wrapper-based HPO can over-specialize to the validation subset, especially when the validation split is small. Therefore, the optimized results are interpreted as evidence that NiOA is effective for this benchmarked HPO problem, not as a general claim that metaheuristic tuning will always improve deep tabular models in environmental datasets. Implications for water-quality modeling The findings of this study have several practical implications for water-quality modeling and environmental data analysis. First, they demonstrate that modern tabular deep-learning models, when properly optimized, can serve as effective surrogate predictors for water-quality variables derived from routine physicochemical measurements. This capability is especially relevant in monitoring scenarios where certain parameters may be costly, time-consuming, or logistically difficult to measure directly. However, because the present dataset is cross-sectional and does not include spatial or temporal identifiers, the demonstrated surrogate prediction capability should be interpreted as applicable to similarly sampled cross-sectional settings rather than to dynamic forecasting or site-transfer problems. Second, the proposed NiOA-assisted AutoInt framework offers a systematic approach for extracting maximum predictive value from relatively small datasets. Many real-world water-quality studies are constrained by limited sample availability, fragmented monitoring campaigns, or resource limitations. The demonstrated effectiveness of metaheuristic hyperparameter optimization in this setting suggests that careful model tuning can partially compensate for data scarcity, improving predictive reliability without requiring additional data collection. At the same time, the results also emphasize the importance of rigorous resampling-based validation when hyperparameter optimization is employed, because wrapper-based tuning can otherwise overfit to a particular validation split. The present evidence should therefore be treated as benchmark-specific rather than as operational deployment evidence. Third, the emphasis on interaction-aware modeling aligns well with the underlying nature of aquatic systems, where water-quality parameters are governed by interconnected physical, chemical, and biological processes. By combining an interaction-focused learning architecture with a robust optimization strategy, the proposed framework provides a flexible tool that can be adapted to a wide range of environmental datasets and prediction tasks. Such adaptability is valuable for exploratory water-quality assessment, scenario analysis, and methodological evaluation of monitoring strategies. Nevertheless, broad decision-support claims should be moderated unless additional evidence is provided across multiple independent monitoring datasets, because strong within-dataset predictive accuracy does not automatically imply transferability to different hydrochemical regimes or sampling designs. Computational behavior The computational resource analysis suggests favorable measured runtime for NiOA + AutoInt in the present implementation. However, because all optimizer-assisted configurations call the same AutoInt training procedure and use the same population size () and iteration budget (), the per-evaluation computational cost should be broadly comparable across optimizers. The reported differences in CPU usage, GPU usage, memory consumption, time per epoch, and total training time are therefore interpreted as implementation-specific measurements rather than intrinsic computational-complexity rankings. These values may reflect optimizer-specific update overhead, early-stopping behavior, hardware scheduling, logging overhead, memory allocation, software implementation details, or runtime measurement conditions. Consequently, the resource results support transparency of the experimental environment and suggest favorable measured runtime in this implementation, but they do not prove universal resource superiority or establish deployment readiness for resource-constrained monitoring systems. Role of LLM assistance LLMs are not part of the predictive model, optimizer, preprocessing executor, statistical evaluator, or performance-measurement pipeline. Their role is restricted to non-decisional methodological checklist support, wording refinement, and consistency checking. No raw data, intermediate results, model outputs, hyperparameter candidates, predictions, or statistical-test inputs are supplied to the LLMs for computation. Therefore, the reported numerical results do not depend on LLM-generated computation. A conventional predictive ablation of “with LLM” versus “without LLM” is not applicable because the LLM component is not an executable modelling component. Instead, the reproducibility logic is methodological: the entire preprocessing, training, tuning, evaluation, and statistical-testing workflow is specified as deterministic and executable without LLM involvement. Limitations Several limitations define the scope within which the results should be interpreted. First, the dataset is small () and the held-out test subset contains approximately 30 samples. This limited test size makes the reported performance estimates sensitive to the particular partition of the data. Although repeated runs quantify stochastic initialization effects under the same split, they do not fully characterize uncertainty due to alternative train–validation–test partitions or external sampling variability. Second, the study uses a fixed train/validation/test split rather than a fully nested cross-validation design. The fixed split supports optimizer parity because all methods are evaluated under identical partitioning and computational budgets. However, nested cross-validation would provide a less biased estimate of generalization when wrapper-based HPO is used, because model selection and performance estimation would be separated across inner and outer loops. The absence of a fully nested design is therefore an important methodological limitation. Third, wrapper-based HPO may over-specialize to the validation subset. In the present workflow, each optimizer repeatedly evaluates candidate AutoInt configurations using validation MSE as the fitness function. This design is appropriate for controlled HPO benchmarking, but it can select configurations that exploit idiosyncrasies of the validation fold, particularly when the validation subset is small. The held-out test set and repeated-run reporting mitigate this risk but do not eliminate it. Fourth, preprocessing-selection bias may be present. Stream A is selected for subsequent optimization because it shows stronger baseline performance than Stream B. This selection is transparently reported and empirically motivated, but it means that the optimized experiments proceed from the stronger preprocessing stream after baseline inspection. Consequently, the final optimized results may reflect both optimizer effectiveness and preprocessing-stream selection. Fifth, no independent external validation dataset is used. All model development, optimizer comparison, and final evaluation are performed using partitions of the same public dataset. Therefore, the findings cannot be assumed to transfer to other watersheds, seasons, monitoring programs, sensor protocols, or hydrochemical regimes without additional validation. Sixth, computational resource rankings are implementation- and hardware-dependent. Although the study reports CPU usage, GPU usage, memory, time per epoch, and total training time, these values may depend on software implementation details, logging, memory allocation, runtime scheduling, and hardware conditions. They should be interpreted as measured characteristics of the present implementation rather than as universal computational complexity rankings among optimizers. Seventh, LLM assistance is non-computational and does not demonstrate predictive benefit. The study does not show that LLMs improve accuracy, reduce error, or enhance optimizer performance. LLMs are restricted to non-decisional documentation and methodological consistency support, while all numerical results are produced by deterministic preprocessing, model training, optimization, and statistical-testing procedures. Eighth, the contribution scope is limited. The work applies an existing optimizer, NiOA, to tune an existing interaction-aware tabular model, AutoInt, on a small public dataset. The contribution lies in the transparent fixed-budget benchmarking workflow, explicit HPO reporting, paired repeated-run analysis, and careful documentation of the modelling protocol. It does not introduce a new optimizer, a new neural architecture, or a new theoretical optimization framework. Ninth, residual-level diagnostics are not fully reported. The manuscript reports aggregate error, bias, correlation, efficiency metrics, repeated-run uncertainty, convergence traces, and run-level MSE/ values, but it does not provide residual histograms, prediction-interval estimates, failure-case analysis, or absolute-error quantiles such as median and 95th-percentile absolute error. Such diagnostics would provide additional evidence for assessing whether unusually low errors arise from uniformly small residuals or from favorable behavior on a small test subset. Finally, the results should not be interpreted as operational environmental decision-support validation. The dataset is cross-sectional, lacks explicit spatial and temporal metadata, and contains a limited number of samples and predictors. Therefore, the findings provide controlled benchmark evidence for a NiOA-assisted AutoInt workflow, but they do not establish readiness for real-time monitoring, regulatory decision-making, site-transfer prediction, seasonal forecasting, or deployment across heterogeneous water-quality regimes. Future work should validate the framework using larger independent datasets, explicit spatiotemporal partitions, nested cross-validation, residual-level uncertainty diagnostics, and comparisons against additional simple and probabilistic baselines. Conclusion This study evaluated dissolved oxygen prediction using the interaction-aware AutoInt architecture combined with NiOA-based hyperparameter optimization. The prediction task was consistently defined as single-target supervised regression of dissolved oxygen (mg/L) from the remaining physicochemical variables in the dataset. Under the fixed benchmark protocol, NiOA + AutoInt achieved the best observed performance among the evaluated optimizer-assisted AutoInt configurations. This result indicates that NiOA was effective for navigating the specified AutoInt hyperparameter space under the matched search bounds, population size, iteration budget, and validation-MSE objective used in this study. The revised experimental reporting strengthens reproducibility by providing the full hyperparameter search space, the selected NiOA-tuned configuration, the matched optimization budget, run-level MSE and values, the convergence trace, and corrected non-parametric statistical comparisons based on paired repeated-run observations. These additions support a more transparent interpretation of optimizer performance than point estimates or rank consistency alone. The conclusions remain bounded by the small cross-sectional dataset, the fixed train/validation/test split, and the absence of external validation. Larger independent datasets with explicit spatial and temporal structure, fully nested cross-validation, residual-level diagnostics, and uncertainty-aware evaluation are still necessary before the fitted model can be generalized to broader environmental monitoring or operational decision-support settings. Acknowledgements Princess Nourah bint Abdulrahman University Researchers Supporting Project number (PNURSP2026R442), Princess Nourah bint Abdulrahman University, Riyadh, Saudi Arabia. Author contributions The authors confirm contribution to the paper as follows: Conceptualization, S.M.K.; methodology, S.M.K.; software, S.M.K.; validation,M.M.E.; formal analysis, S.M.A.; investigation, S.M.A.; resources, E.A.M.; data curation, S.M.K., M.M.E., S.M.A., E.A.M.; writing—original draft preparation, M.M.E., S.M.A., E.A.M.; writing—review and editing, E.A.M.; visualization, S.M.K.; supervision, S.M.K.; project administration. All authors reviewed the results and approved the final version of the manuscript. Funding No Fund Data availability The data used in this study are publicly available at: https://www.kaggle.com/datasets/shreyanshverma27/water-quality-testing/data. Declarations Competing interests The authors declare no competing interests. Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The data used in this study are publicly available at: https://www.kaggle.com/datasets/shreyanshverma27/water-quality-testing/data.