A digital twin approach for sustainable construction: predictive optimization of concrete strength using industry 4.0 principles Authors: Peipei Xin, Haytham F Isleem, Mohammad Khishe Published: 2025 ABSTRACT AbstractThe construction industry faces pressure to improve sustainability, but a critical gap persists between conceptual Digital Twin (DT) frameworks and the empirically validated, operational workflows needed for material design. To bridge this gap, this research proposes and empirically evaluates a tangible, end-to-end Industry 4.0 workflow. This study serves as a proof-of-concept, demonstrating how a DT can be developed from a theoretical concept into a practical optimization engine for material design. Our methodology operationalizes this blueprint by utilizing a public dataset as a proxy for digitalization, developing a deep neural network (DNN) as the high-fidelity virtual twin, and benchmarking it against four classical machine learning (ML) models to predict concrete compressive strength. The results provide a strong performance benchmark for this approach: the deep learning-based Digital Twin achieves high predictive fidelity with an R² of 0.9011 and a Root Mean Squared Error (RMSE) of 5.09 Megapascals (MPa) on the held-out test set. In stark contrast, baseline linear models achieved an R-squared (R²) of only 0.632, demonstrating that the complexity of our proposed workflow is essential for capturing non-linear material interactions. This work’s primary contribution is a reproducible blueprint for applying predictive analytics in sustainable construction, enabling the data-driven design of high-performance, low-carbon concrete mixes while significantly reducing the need for physical experimentation.Keywords: Digital twin, Sustainable construction, Predictive optimization, Concrete strength, Industry 4.0Subject terms: Engineering, Materials science, Mathematics and computing Introduction Concrete is the most widely used construction material globally1, and its performance has a direct impact on structural safety, durability, and sustainability2. Its performance under foundational loads, for example, is a key aspect of structural safety3. Furthermore, its resilience to extreme conditions, such as fire, is critical for infrastructure integrity4. The seismic behavior of concrete structures is another crucial area of performance research5. Consequently, the optimization of structural designs to meet specific seismic performance targets has become a key focus in the field6. With increasing pressure on the construction industry to reduce carbon emissions, enhance efficiency. This includes not only material production but also the operational energy of buildings, which can be optimized through demand-side management7. This pressure also compels the industry to align with Industry 4.0 principles, which advocate for the integration of digital technologies such as IoT, AI, and cyber-physical systems to create intelligent, automated, and data-driven processes. Our work applies these principles by creating a data-driven virtual model (the digital twin) to optimize a physical material (concrete). Innovative approaches are required to design and optimize concrete mixes8,9. Such innovations include the development of novel green expansive agents from industrial byproducts10. Similarly, incorporating recycled materials, such as waste tire steel fibers, contributes to creating more ecologically friendly concrete11. Research also actively explores the mechanical properties and durability of sustainable alternatives, such as alkali-activated slag concrete12. Traditional practices in mix design and performance assessment often rely on experimental trial-and-error processes, which are time-consuming, resource-intensive, and environmentally costly13. These limitations underscore the urgent need for intelligent, data-driven solutions that can predict, optimize, and virtually test material properties before physical implementation. Deep learning, for instance, has been effectively applied to predict complex phenomena, such as plastic damage in concrete under compression14. Data-driven models are also being developed to predict long-term material behaviors such as creep, which is vital for service-life assessments15. These predictive capabilities are essential for addressing persistent issues, such as material degradation and corrosion, over the structure’s lifespan16. In recent years, DT technology has emerged as a transformative tool in construction, enabling the creation of virtual replicas of physical systems that can simulate17, monitor, and optimize real-world processes18. The need for such accurate virtual models is particularly evident in the analysis of large-scale concrete infrastructure, where complex system dynamics must be captured19. This pursuit of virtual replication is supported by advances in high-fidelity mesoscale modeling, which aim to create detailed digital counterparts of concrete20. Complementary data-driven techniques are also emerging to enable high-resolution imaging from sensor data, further bridging the physical-digital gap21. Paired with ML, digital twins offer the potential to predict critical properties such as compressive strength and to optimize mixtures for sustainability22,23. However, a significant chasm exists between this potential and its practical realization24. Most conventional approaches to DTs remain restricted to isolated laboratory data and lack real-time integration with sensor-driven workflows25, and are often presented as high-level conceptual frameworks without empirical validation26. These shortcomings prevent their direct adoption in practice and highlight a critical gap between theoretical promise and operational, validated solutions for sustainable construction. Parallel to advancements in DT, machine learning has become a central technique in predicting concrete performance27. While classical models have been applied with varying success28, deep learning methods have recently demonstrated a superior capacity to capture the complex29, nonlinear interactions among material components30. Yet, these ML studies frequently exist as standalone predictive exercises, often relying on small or literature-derived datasets without embedding the model within a broader, operational workflow31,32. They demonstrate predictive power but do not provide the end-to-end framework necessary to guide real-world design decisions from digitalization through to optimization33,34. This study directly addresses this critical gap by proposing and empirically validating a deep learning-based digital twin framework that operationalizes a complete, end-to-end Industry 4.0 workflow. Our approach moves beyond purely conceptual frameworks by providing a reproducible, tangible blueprint for implementing a digital twin in sustainable construction. The framework simulates high-fidelity data streams through a Digitalization Proxy, employs a neural network as the virtual Digital Twin to capture complex material relationships, and integrates an Optimization Engine to test sustainable mixes virtually. By demonstrating both high predictive accuracy and tangible optimization capacity, this work provides a rigorous, empirical foundation for operationalizing digital twins in sustainable construction. This step has been largely absent from the existing literature. To bridge this gap in a controlled and reproducible manner, this study operationalizes the DT framework using a well-established public dataset as a proxy for real-world digitalization. While not a substitute for real-time data, this approach allows for the rigorous validation of the methodological blueprint itself, establishing a critical foundation for future, fully integrated digital twins. The main contributions of this study are as follows: An empirically validated, end-to-end Digital Twin workflow, benchmarked to demonstrate the superiority of a deep learning approach. An empirically validated, end-to-end Digital Twin workflow, benchmarked to demonstrate the superiority of a deep learning approach. A data-driven analysis identifying the critical material levers that govern concrete performance for sustainable mix design. A data-driven analysis identifying the critical material levers that govern concrete performance for sustainable mix design. An optimization simulation showcasing the workflow’s ability to design high-performance, sustainable concrete mixes without physical experimentation. An optimization simulation showcasing the workflow’s ability to design high-performance, sustainable concrete mixes without physical experimentation. The remainder of this paper is structured as follows: Sect. “Relevant studies” reviews the most relevant studies on digital twin and ML applications in construction. Section “Methods” describes the proposed methodology, including dataset selection, preprocessing, model development, and optimization simulation. Section “Results” presents results from benchmarking and optimization, while Sect. “Discussion” discusses the implications, limitations, and avenues for future research. Finally, Sect. “Conclusion” concludes the study by highlighting its contributions to sustainable construction and digital twin adoption. Relevant studies Zahedi et al.35 systematically reviewed 235 Scopus-indexed articles (2017–2024) on DT applications in the construction industry to enhance sustainability. Using scientometric and thematic analysis, they categorized DT usages, simulation, technology integration, smart systems, and discussed how DT—as an evolution of BIM—can address sustainability goals, improve energy modeling, and waste reduction. The authors highlighted that DT research is growing, often integrating Artificial Intelligence (AI)/Internet of Things (IoT) technologies. Still, they noted a lack of comprehensive synthesis, prompting them to identify five key research gaps and future directions. As a review, their methodology aggregates existing findings rather than testing new models. The study’s limitation is that it covers literature only up to 2024 and remains at a conceptual level, with no empirical validation. Thus, while Zahedi et al. demonstrate the promising role of DT in eco-friendly construction, their work highlights the remaining challenges and underscores the need for integrated, operational frameworks that align with our identified research gap. Authors in36 present a broad literature review of digital twin technology across industries. They define the DT concept, trace its evolution, and review key enabling technologies, including AI and IoT, as well as their respective use cases. The methodology involves synthesizing existing studies rather than conducting experiments. Their findings emphasize that the adoption of DT is accelerating, streamlining automation in various sectors and yielding efficiency gains. For example, they document how DTs enable predictive maintenance and process optimization in multiple industries, including manufacturing. However, this review is general-purpose: it does not focus specifically on construction or sustainability contexts. Its limitation is that it remains conceptual and cross-domain, offering little domain-specific insight. In relation to the research gap, their work suggests that while the DT theory is well understood, its application to sustainable construction challenges remains underexplored. Babaei et al.37 employed a qualitative meta-synthesis approach to identify applications of digital twin in sustainable construction. They searched Web of Science (2020–2024) for DT and sustainability, screened 323 papers, and included 37 for analysis. The authors categorized DT use-cases into three construction phases. In Pre-Construction, DT supports decision-making, planning, automation, and environmental impact decisions. During Construction, DT is applied to work progress monitoring, safety management, waste management, and quality control. Post-Construction uses include smart maintenance, occupant comfort, energy/resource monitoring, and asset management. This systematic classification shows the breadth of DT’s sustainability roles. The methodology’s strength is in mapping the landscape of DT applications. Its limitations are that it relies entirely on published literature, has no new data, and is limited to English-language papers through 2024. As a result, it highlights numerous proposed uses but offers no empirical testing. In the context of our research gap, this paper underscores that DT in construction is mostly conceptual and that practical, validated implementations are still needed. A study in38 developed machine learning models to predict the mechanical performance of recycled concrete aggregate (RCA) concrete, thereby addressing concrete sustainability. They compiled 583 samples from the literature, which included 10 input features and two outputs: compressive and tensile strength. Three machine learning models—Extreme Gradient Boosting (XGBoost), Decision Tree, and K-Nearest Neighbors—were trained and evaluated using metrics such as R², RMSE, and Mean Absolute Error (MAE). XGBoost achieved the highest accuracy, with an R² ≈ of approximately 0.86, for compressive strength. The study also employed SHAP (SHapley Additive exPlanations) analysis to interpret the model, revealing that the water-to-cement ratio and cement content are the most influential factors. The methodology’s advantage is its practical focus, as it provides a data-driven predictive approach to sustainable concrete design, potentially reducing the need for costly experiments. The limitations include reliance on a literature-derived dataset lacking detailed RCA characterization; the authors note that the absence of such data could affect model interpretability and generalizability. Thus, while the models show strong predictive power, their real-world applicability requires more detailed or experimental validation. Khan et al.39 evaluated multiple advanced ML models for eco-friendly self-compacting concrete (SCC) design. Using a dataset of 239 SCC mix designs, they trained backpropagation neural networks (BPNN), random forest (RF), K-Nearest Neighbors (KNNs), stacking, bagging, and XGBoost to predict compressive strength and workability metrics. Model performance was evaluated using R², Mean Squared Error (MSE), and other metrics. XGBoost consistently outperformed others, achieving an R² of ~ 0.999 on training and ~ 0.935 on testing for compressive strength. They conducted a sensitivity analysis, revealing that the cement content, silica fume, and coarse aggregate have a positive effect on strength, while water has a negative effect. The findings demonstrate the potential of ML, especially XGBoost, to optimize sustainable concrete mixes, thereby reducing the need for experimental trials. However, the limitations include the relatively small dataset of 239 samples and a focus on specific SCC compositions. These models may need validation on larger or different datasets to ensure robustness across diverse mix designs. Work in40 proposed a Multi-Dimensional Digital Twin–BIM framework (MD-DTT-BIM) for intelligent building construction and operations. Their system integrates real-time sensor data, BIM-based simulation, and predictive analytics to optimize various building metrics simultaneously. In experimental validation, this approach yielded significant improvements: operational efficiency increased by 97.6%, real-time monitoring accuracy improved by 96.7%, energy consumption was reduced by 95.3%, and occupant comfort prediction accuracy increased by 94.3%. These results suggest that their DT framework can substantially advance energy efficiency and occupant satisfaction in buildings. The methodology’s strength lies in combining data-driven models with building information models. A key limitation is that the validation was performed on a limited set of buildings in specific regions; as the authors acknowledge, the selected datasets and performance indicators were not comprehensive, which may limit the generalizability of the results. In summary, Zhou et al. demonstrate notable advantages of DT-driven design, but further testing across diverse settings is needed. This review of recent literature reveals a clear research gap. While systematic reviews confirm the conceptual promise of DTs in construction and individual ML studies demonstrate high predictive accuracy on specific tasks, there is a notable absence of studies that provide an empirically validated, end-to-end, and reproducible workflow that bridges the gap from digitalization to practical optimization. Our study directly addresses this gap by proposing and benchmarking such a workflow, offering a tangible blueprint that moves beyond theory and demonstrates a complete, data-driven process for sustainable material design. As shown in Table 1, these recent studies collectively demonstrate the promise of digital twin and machine learning methods, while also revealing common gaps that motivate our research. The systematic reviews highlight that current work is mainly conceptual, mapping out potential applications without providing empirical validation of an end-to-end workflow. Concurrently, the machine learning studies, despite achieving high predictive accuracy, consistently note limitations tied to small or literature-derived datasets and a lack of real-world validation. Overall, no existing work yet provides a fully generalizable and empirically validated digital twin framework tailored to sustainable concrete. These persistent limitations suggest that further integrative research is necessary, which our study directly addresses by operationalizing and validating a comprehensive, data-driven workflow. Table 1Summary of methodologies, findings, and limitations of the selected recent studies.ReferenceMethodologyFindingsLimitationsZahedi et al.35Systematic & scientometric review of DT literature in sustainable construction.Mapped DT application trends and confirmed its potential to advance sustainability goals.Conceptual review only; lacks empirical validation of a proposed framework.Babaei et al.36Qualitative meta-synthesis classifying DT applications from published studies.Systematically classified diverse DT use-cases across all construction project phases.Theoretical synthesis provides no new data or model testing.Attaran et al.37A broad literature review of DT technology and its applications across industries.Established DT as a key enabling technology for automation and process optimization.Lacks specific focus on construction or sustainability; remains at a high conceptual level.Khan et al.38ML modeling (XGBoost) to predict the strength of recycled aggregate concrete.Achieved high predictive accuracy (R²≈0.86) and identified key material drivers using SHAP.Relied on literature-derived data, raising concerns about real-world generalizability.Khan et al.39Benchmarked six ML models for optimizing eco-friendly SCC.Demonstrated superior performance of XGBoost (R²≈0.935) for sustainable mix design.Small, domain-specific dataset; requires broader validation to ensure robustness.Zhou et al.40Developed an integrated DT-BIM framework for intelligent building operations.Quantified significant gains in operational efficiency, energy reduction, and monitoring.Validation was limited to specific building types, impacting the framework’s generalizability. Summary of methodologies, findings, and limitations of the selected recent studies. Methods This study employs a quantitative, empirical approach to validate a conceptual Industry 4.0 workflow for advancing sustainable construction practices. The methodology is structured into four distinct phases: (1) Conceptual Framework and Dataset Selection, (2) Data Preprocessing and Feature Engineering, (3) Predictive Model Development and Benchmarking, and (4) Model Evaluation and Optimization Simulation. This workflow is visually outlined in Fig. 1. A random state was fixed throughout all experiments to ensure full reproducibility. The four-phase research methodology workflow. Conceptual framework and dataset selection To empirically test the role of digitalization in sustainable construction, we established a three-part conceptual framework grounded in the principles of Industry 4.0. The first component, the Digitalization Proxy, is designed to simulate the high-fidelity data streams that a fully digitized process would generate. This simulated data then informs the second component, the Digital Twin, which is a predictive model developed to virtually mirror the physical construction process. The final element is an Optimization Engine, which leverages the predictive capabilities of the Digital Twin to support intelligent, data-driven decision-making and enhance sustainability outcomes. It is essential to acknowledge that a true, operational digital twin would be driven by real-time sensor data from a physical asset. In this study, the static UCI dataset serves as a high-fidelity Digitalization Proxy to develop and validate the core predictive model and workflow, which is a necessary first step before deployment in a dynamic environment. To operationalize this framework, the “Concrete Compressive Strength” dataset from the UCI (University of California, Irvine) Machine Learning Repository was selected41. The descriptive statistics for all variables in the dataset are provided in Table 2. Table 2Descriptive statistics of dataset Variables.VariableUnitMeanStd. Dev.MinMax Input featuresCementkg/m³281.2104.5102.0540.0Blast Furnace Slagkg/m³73.986.30.0359.4Fly Ashkg/m³54.264.00.0200.1Waterkg/m³181.621.4121.8247.0Superplasticizerkg/m³6.26.00.032.2Coarse Aggregatekg/m³972.977.8801.01145.0Fine Aggregatekg/m³773.680.2594.0992.6Agedays45.763.21.0365.0 Output variableCompressive StrengthMPa35.816.72.3382.60 Descriptive statistics of dataset Variables. This dataset was deemed an ideal proxy for the ‘digitalization’ phase. It contains 1,030 instances, each with eight quantitative input features representing precise material components (Cement, Blast Furnace Slag, Fly Ash, Water, Superplasticizer, Coarse Aggregate, Fine Aggregate), as well as the curing age. The target variable is the concrete’s compressive strength in MPa. The dataset’s granular, sensor-like precision for each input aligns perfectly with the data that would be generated and logged in a digitally-controlled ‘smart’ manufacturing environment. Furthermore, the inclusion of both carbon-intensive (Cement) and sustainable industrial byproducts (Blast Furnace Slag, Fly Ash) makes it uniquely suited for an analysis focused on sustainability. Data preprocessing and partitioning Before model development, a rigorous preprocessing pipeline was implemented to ensure data quality and optimal model performance. Exploratory data analysis (EDA) An initial EDA was conducted to understand the statistical properties and inter-variable relationships within the dataset. This involved generating a correlation matrix to identify linear relationships and plotting the distributions of variables to assess skewness and scale. Figure 2 illustrates the Pearson correlation coefficients between all input features and the target variable, compressive strength, through a color-coded heatmap. An analysis of the matrix reveals several key linear relationships that are critical for predictive modeling. The final column, corresponding to the target variable, shows that cement has the strongest positive correlation with compressive strength (0.50), followed by superplasticizer (0.37) and age (0.33), signifying that increases in these factors are associated with higher concrete strength. Conversely, water demonstrates the most significant negative correlation (−0.29), which aligns with established principles that a higher water-to-cement ratio can compromise the integrity of the final product. The matrix also highlights potential multicollinearity among the predictors, most notably the strong negative relationship between superplasticizer and water (−0.66), suggesting that the use of one often reduces the need for the other. This comprehensive visualization not only helps identify the most influential predictors but also provides essential insights into the interdependencies between material components, which can inform subsequent feature engineering and model selection processes. Correlation matrix heatmap of input features and compressive strength. Figure 3 displays a matrix of histograms, each illustrating the frequency distribution of one of the eight input features and the target variable, compressive strength. This visualization is crucial for understanding the underlying statistical properties of the dataset, such as central tendency, spread, and skewness. Several variables, including Coarse Aggregate, Fine Aggregate, and the target Compressive Strength itself, approximate a normal or near-normal distribution, as indicated by their relatively symmetric, bell-shaped curves. In contrast, a significant number of features exhibit severe right-skewness. Specifically, Blast Furnace Slag, Fly Ash, Superplasticizer, and Age are characterized by a high concentration of data points at lower values—many of which are zero—with a long tail extending towards the right. This skewness suggests that these industrial byproducts and chemical admixtures are often used sparingly or are absent in a large portion of the concrete mixes represented in the dataset. The wide variation in both the shape and scale of these distributions underscores the critical need for the feature scaling step outlined in the methodology, as it prevents variables with larger magnitudes from dominating the learning process of models, such as neural networks. Histograms showing the distribution of all input features and the target variable. Figure 4 provides a comparative visualization of the distributions for each input feature through a series of box plots, presented on their original, unscaled values. The most immediate and critical insight from this plot is the vast disparity in the scales and ranges of the variables. For instance, features like coarse_aggregate and fine_aggregate have median values in the hundreds, whereas superplasticizer and age operate on a much smaller scale, often in the single or low double digits. The plots also reveal the distributional shapes. The variables like cement and water appear symmetric, blast_furnace_slag, fly_ash, and superplasticizer exhibit a strong right skew, with their medians clustered towards the lower end of their ranges. Additionally, the presence of numerous outliers, indicated by the individual points beyond the whiskers, is evident for several features, such as age and water. Collectively, these stark differences in scale, spread, and the presence of outliers strongly justify the necessity of the feature scaling step described in the methodology. Standardization ensures that each feature contributes equitably to the model’s learning process, preventing features with larger numerical values from disproportionately influencing the results. Box plots of the input features prior to data scaling. Figure 5 presents a comprehensive series of scatter plots, each visualizing the pairwise relationship between an input feature and the target variable, compressive strength, with a fitted linear regression line highlighting the dominant trend. The analysis clearly distinguishes the most influential linear predictors from those with weaker relationships. Strong positive linear trends are evident for cement, superplasticizer, and age, confirming that increases in cement content, the use of superplasticizer, and longer curing times are all associated with higher compressive strength. Conversely, a distinct negative linear trend is observed for water, aligning with the principle that a higher water-to-cement ratio weakens the final product. The remaining components—blast furnace slag, fly ash, coarse aggregate, and fine aggregate—exhibit much weaker or negligible linear relationships, characterized by high data dispersion and, in the case of slag and fly ash, prominent vertical clusters at the zero-value mark. Overall, the figure visually validates the insights from the correlation matrix. Still, the significant scatter of data points around all regression lines suggests that while linear models can capture some key trends, complex, non-linear interactions among the variables are also critical in determining the final compressive strength. Pairwise scatter plots of all input features against compressive strength. Figure 6 presents a series of violin plots that offer a nuanced view by visualizing the distribution of key material components across three discrete compressive strength categories: Low, Medium, and High. This visualization reveals essential trends that simple correlation cannot capture. A clear upward trend in the median cement content is observable as the strength category increases from Low to High, confirming its role as a primary driver of strength. Conversely, the distribution of water shows a distinct downward shift, with high-strength concrete mixes consistently utilizing less water than their low-strength counterparts. The relationships for the industrial byproducts, Blast Furnace Slag and Fly Ash, are more complex. For both components, the median value remains near zero across all strength categories, indicating that their inclusion is not universal. However, the shape of the violins suggests that while many high-strength mixes contain little to no slag or fly ash, it is still possible to achieve high strength while incorporating them, highlighting non-linear interactions and the importance of specific mix designs. Taken together, these plots illustrate that achieving a certain strength is not merely a function of a single ingredient but rather the result of complex trade-offs and combinations, further justifying the use of a deep learning model capable of capturing these intricate, non-linear relationships. Violin plots of key features across different strength categories. Data partitioning To ensure a robust and unbiased evaluation of model performance, the dataset was first partitioned into a training set (85% of the data) and a final, held-out testing set (15%). To prevent overfitting and rigorously validate the model architecture, a 5-fold cross-validation procedure was performed on the 85% training set. The models were trained five times, with performance evaluated on a different fold each time. The final model performance reported in this paper is assessed on the completely unseen 15% testing set, using the model architecture that performed best during cross-validation. Feature scaling The EDA revealed that the input features operate on vastly different scales. To prevent features with larger magnitudes from disproportionately influencing model training, especially for the neural network, StandardScaler from the Scikit-learn library was employed. This scaler transforms each feature to have a mean of 0 and a standard deviation of 1. The transformation for a feature value x is defined by Eq. 1: where is the mean of the feature values in the training set and is the standard deviation. The scaler was fitted exclusively on the training data to prevent data leakage, and the learned transformation was then applied to the validation and testing sets. Predictive model development and benchmarking The core of this research involved developing a ‘digital twin’ using a state-of-the-art deep learning model and benchmarking its performance against established machine learning algorithms. Baseline models To establish a robust performance baseline, four widely used regression models were implemented. These included Linear Regression, a standard linear model used to assess baseline linear relationships, and Ridge Regression, a regularized linear model to account for potential multicollinearity. The baseline was further strengthened by two ensemble methods: the Random Forest Regressor, an assembly of decision trees known for its high performance and robustness, and the Gradient Boosting Regressor, an iterative method that builds strong predictive models. A crucial final step involved training all baseline models on the scaled training data. These included Linear Regression, a standard linear model used to assess baseline linear relationships, which models the target y as a linear combination of the features , as shown in Eq. 2: Also implemented was Ridge Regression, a regularized linear model that adds an penalty to the loss function to account for potential multicollinearity. The objective of Ridge Regression is to minimize the penalized residual sum of squares provided in Eq. 3: where represents the model coefficients and is the regularization parameter that controls the penalty strength. To provide a comparison against industry-standard empirical methods, a baseline model derived from established concrete principles was implemented. This model first predicts the 28-day compressive strength based on the water-to-cementitious-materials ratio (w/cm) using a well-established exponential formula. It then adjusts this prediction for the actual curing age using the time-dependent strength gain equation recommended by ACI Committee 209. This provides a non-machine-learning benchmark for model performance, helping to quantify the advantage of a data-driven approach over traditional empirical formulas. Deep learning model: the digital twin A DNN was designed using the PyTorch framework to serve as the ‘digital twin’. The architecture was constructed to capture complex, non-linear interactions between the input features. It began with an input layer that accepted eight scaled features, followed by three fully connected hidden layers with 128, 256, and 128 neurons, respectively. The Rectified Linear Unit (ReLU) was used as the activation function in each hidden layer to introduce non-linearity. To mitigate overfitting, a Dropout layer with a rate of 0.3 was applied after each hidden layer, which randomly deactivates neurons during training. The network culminated in an output layer containing a single linear neuron to output the predicted compressive strength. The ReLU was used as the activation function in each hidden layer to introduce non-linearity, defined by Eq. 4: The model was trained using the Adam optimizer with a learning rate of 0.001 and MSE as the loss function. This training was performed over 300 epochs, with performance on the validation set continuously monitored. A checkpointing mechanism was implemented to save the model weights that yielded the lowest validation loss, ensuring the final model represents the optimal state of convergence. The specific architecture and training hyperparameters for the DNN model are detailed in Table 3. Table 3Deep neural network configuration and Hyperparameters.ParameterValueFrameworkPyTorchHidden Layers3Neurons (Layer 1, 2, 3)128, 256, 128Activation FunctionReLUDropout Rate0.3OptimizerAdamLearning Rate0.001Loss FunctionMSEBatch Size32Epochs300 Deep neural network configuration and Hyperparameters. Model evaluation and optimization simulation Performance evaluation metrics The performance of the final deep learning model and all baseline models was quantitatively assessed on the held-out testing set using a suite of standard regression metrics. First, R² was used to determine the proportion of the variance in the target variable that is predictable from the input features, calculated as shown in Eq. 5: To quantify prediction errors, the MSE was calculated, which represents the average of the squares of the errors and penalizes larger errors more heavily as presented in Eq. 6: The RMSE was then derived as the square root of the MSE, providing a more interpretable error metric in the same units as the target variable (MPa), as defined in Eq. 7: Finally, the MAE was computed as the average of the absolute prediction errors, a metric that is less sensitive to outliers than MSE Eq. 8: In the equations above, represents the actual value, is the predicted value, is the mean of the actual values, and is the total number of samples in the testing set. Feature importance analysis To interpret the ‘digital twin’ and understand which ‘digital levers’ have the most significant impact on compressive strength, a Permutation Feature Importance analysis was conducted. This technique measures the decrease in the model’s R² score when the values of a single feature are randomly shuffled, thereby breaking the relationship between that feature and the target. The larger the decrease in score, the more critical the feature. Optimization engine simulation To complete the Industry 4.0 framework, we utilized the trained and validated ‘digital twin’ as an optimization engine. A hypothetical, highly sustainable concrete mix was designed based on domain knowledge: low cement content (25th percentile), high industrial byproduct content (maximum observed slag and fly ash), and maximum curing age (365 days). The digital twin was then used to predict the compressive strength of this virtual mix, providing proof of concept for its utility in data-driven, sustainable material design, eliminating the need for physical experimentation. It is essential to note that this simulation serves as an illustrative example to demonstrate the workflow’s capabilities, rather than an exhaustive optimization process. The goal is to provide a clear proof of concept for how the validated Digital Twin can be leveraged as a design tool. Results Model performance benchmarking Model performance benchmarking is a critical process that systematically evaluates the predictive accuracy of our proposed Deep Learning-based Digital Twin against a diverse set of established baseline models using standardized metrics. This process is essential for this analysis for two primary reasons. First, it provides a rigorous, empirical justification for employing a sophisticated deep learning architecture; without a proper benchmark, the superior performance of the Digital Twin would be an unsubstantiated claim. By demonstrating a significant performance gap between the Digital Twin and simpler models, especially linear ones, we can definitively prove that the complexity of the deep neural network is necessary to capture the intricate, non-linear relationships inherent in concrete material science. Second, this comparative analysis builds confidence in the Digital Twin’s reliability for its intended purpose as an optimization engine, ensuring that its predictions are not only accurate in isolation but also demonstrably better than those of well-known alternatives. Therefore, benchmarking serves not merely as a procedural step but as the fundamental evidence that validates our choice of model and underscores the complexity of the problem being solved. Table 4 presents the quantitative performance of the final Deep Learning (Digital Twin) model and the four baseline models on the held-out testing set across four standard regression metrics. The results reveal a distinct performance hierarchy, with the non-linear models—Deep Learning, Random Forest, and Gradient Boosting—significantly outperforming the linear approaches. The Deep Learning model emerged as the top performer, achieving the highest R-squared value of 0.901 and the lowest errors with an RMSE of 5.09 MPa and an MAE of 3.67 MPa, indicating that it successfully explains over 90% of the variance in the target variable. The ensemble methods, Random Forest and Gradient Boosting, demonstrated highly competitive performance, with R-squared values of approximately 0.895 and error metrics only marginally higher than the Deep Learning model. In stark contrast, both the Ridge and Linear Regression models performed substantially worse, with R-squared values of only 0.632 and an RMSE of approximately 9.81 MPa, nearly double the error of the top models. In addition, the performance of a traditional Code-Based Equation was evaluated. As shown in Table 4, this model performed exceptionally poorly (R² ≈ −0.21), yielding an RMSE nearly double that of the linear models and more than triple that of the top-performing models. This starkly demonstrates the limitations of simplified empirical formulas in capturing the complex, non-linear material interactions present in the data. This significant performance gap highlights the limitations of linear models in capturing the complex, non-linear interactions inherent in the design of concrete materials. It strongly validates the selection of a deep neural network as the most effective architecture for the Digital Twin. Table 4Comparative performance of all models on the test set.ModelR²MSERMSEMAEDeep learning (PyTorch)0.90113535725.887115485.0879383093.672739029Random forest0.89540252127.388225555.2333761143.853857164Gradient boosting0.89365760827.845120725.2768476123.821774876Ridge regression0.63229800896.280572169.8122664137.778573624Linear regression0.63222503196.29968089.8132400767.76918056Code-based equation−0.2082316.529917.791315.1554 Comparative performance of all models on the test set. Figure 7 provides a clear visual comparison of the predictive accuracy of all five models using a lollipop chart, with the RMSE as the primary metric. The plot starkly illustrates the significant performance gap between the non-linear models and their linear counterparts. The Deep Learning model, highlighted in red, is positioned furthest to the left, signifying that it achieved the lowest RMSE and is therefore the most accurate model. It is followed closely by the Random Forest and Gradient Boosting models, which form a tight cluster of high-performing algorithms with only marginally higher error rates. In sharp contrast, the Ridge and Linear Regression models are located far to the right, indicating a substantially higher RMSE that is nearly double that of the top-performing models. Comparative performance of all models based on RMSE. Figure 8 utilizes a radar chart to visually compare the composition of a hypothetical, highly sustainable concrete mix against the average mix profile from the dataset, effectively demonstrating the practical application of the Optimization Engine. The inner blue polygon represents the ‘Average Mix,’ providing a baseline, while the outer green polygon illustrates the ‘Sustainable Mix.’ Sustainable design is characterized by a significant reduction in cement content, a primary goal for reducing the carbon footprint of concrete. To compensate for this reduction and enhance performance, the mix incorporates the maximum observed amounts of the industrial byproducts Blast Furnace Slag and Fly Ash. Furthermore, the design leverages the full potential of long-term strength gain by setting the curing age to its maximum value. Crucially, the chart’s legend indicates that the Digital Twin predicts this sustainable mix will achieve an impressive compressive strength of 69.3 MPa. This visualization serves as a powerful proof of concept for the Optimization Engine, demonstrating its ability to explore and validate novel, data-driven material designs that prioritize sustainability while simultaneously achieving high performance, all without the need for physical experimentation. Radar chart comparing the composition of an average mix and a simulated sustainable mix. Digital twin performance and diagnostic analysis Having established the Digital Twin’s superior performance in the benchmarking phase, a more in-depth examination is necessary to validate its predictive fidelity and statistical soundness through a comprehensive diagnostic analysis. This dual-pronged approach is essential for moving beyond simple accuracy metrics and building confidence in the model’s reliability. First, the model’s performance is visually confirmed by plotting its predicted values against the actual values. A tight and consistent clustering of data points around the ideal fit line provides a clear, graphical representation of its high accuracy across the entire range of the data. Second, a thorough diagnostic analysis of the model’s residuals—the prediction errors—is conducted to verify that the model is well-behaved. An ideal outcome, characterized by a random, patternless scatter of residuals centered around zero and a normal distribution of errors, provides strong evidence that the model is unbiased, its errors are random noise, and it has successfully captured the underlying structure of the data. Together, these performance and diagnostic checks provide a robust validation, confirming that the Digital Twin is not just an accurate model, but a well-calibrated and trustworthy tool suitable for its ultimate application as an optimization engine. Figure 9 graphically depicts the performance of the final Deep Learning model on the held-out testing set by plotting its predicted compressive strength values against the corresponding actual values. The red dashed line represents the “Ideal Fit,” a benchmark for a perfect prediction where the predicted value equals the actual value. The data points are tightly clustered around this ideal line, demonstrating a strong linear correlation and indicating that the model’s predictions are highly accurate with minimal deviation. This high level of accuracy is consistently maintained across the entire spectrum of strength values, from the lower range (below 20 MPa) to the highest range (above 70 MPa), which highlights the model’s excellent generalization capabilities and its ability to avoid systematic bias. Scatter plot of predicted vs. actual compressive strength for the Digital Twin model. Figure 10 presents an alternative visualization of the model’s performance, utilizing a density hexbin plot to complement the standard scatter plot. This plot aggregates individual data points into hexagonal bins, with the color of each bin representing the density of predictions in that region, from low (dark blue) to high (bright yellow). The visualization strongly reinforces the model’s high accuracy, as the vast majority of hexagons are tightly concentrated along the red dashed line of ideal fit. Crucially, the areas of highest point density, indicated by the brighter yellow hexagons, lie almost perfectly on the ideal fit line. This signifies that the model is not only accurate overall but is particularly robust and reliable in the most common regions of the data distribution. Even in regions with lower data density, represented by the darker colored bins, the predictions remain closely aligned with the actual values, demonstrating the model’s strong generalization capabilities. By effectively managing overplotting and highlighting the density of correct predictions, this hexbin plot provides further compelling evidence of the Digital Twin’s predictive power and reliability. Density hexbin plot of predicted versus actual values for the Digital Twin model. Figure 11 provides a crucial diagnostic visualization by plotting the residuals (the difference between actual and predicted values) against the predicted compressive strength values from the Digital Twin model. The horizontal red dashed line at zero represents the baseline for perfect prediction, where the error is zero. The plot reveals that the residuals are randomly and uniformly scattered around these zero lines. Crucially, there are no discernible patterns, such as a curve or a funnel shape, in the distribution of the points. This lack of a pattern indicates that the variance of the errors is constant (homoscedastic) across the entire range of predictions, meaning the model’s error magnitude remains constant regardless of whether it is predicting low or high strength values. The random scatter confirms that the model is not systematically over-predicting or under-predicting for any particular subset of the data. This ideal residual distribution provides strong evidence that the model has successfully captured the underlying structure of the data and that its errors are random noise, which is a key indicator of a well-fitted and reliable predictive model. Diagnostic plot of residuals versus predicted strength values. Figure 12 provides a further diagnostic analysis by plotting a histogram of the prediction errors (residuals) from the Digital Twin model. This visualization is used to assess whether the errors follow a normal distribution, a key characteristic of a well-behaved regression model. The plot clearly shows that the residuals are centered around zero, indicating that the model’s predictions are, on average, unbiased and do not systematically over- or under-predict the actual values. The distribution of the errors approximates a normal distribution, characterized by its symmetric, bell-shaped form, which is further emphasized by the overlaid probability density curve. This normality satisfies a fundamental assumption of regression analysis, suggesting that the model’s errors can be treated as random noise. Moreover, the histogram reveals that the majority of prediction errors are concentrated within a narrow band (approximately ± 5 MPa). In contrast, a few larger errors exist in the tails; their low frequency confirms the model’s general consistency. In conjunction with the random scatter observed in Fig. 11, this normality of residuals provides robust statistical evidence that the Digital Twin is a well-calibrated and reliable predictive model. Histogram of the distribution of model residuals. Digital twin training and interpretation Beyond confirming the model’s high predictive accuracy, it is crucial to both validate the integrity of its learning process and interpret its internal decision-making logic. This section addresses these two vital aspects to build full confidence in the Digital Twin. First, an analysis of the training and validation loss curves over the full course of the training epochs is presented to confirm that the model learned effectively and converged to a stable solution without overfitting, thereby ensuring its ability to generalize to new, unseen data. Second, to transform the ‘black box’ deep neural network into an insightful and practical tool, a permutation feature importance analysis is conducted. This interpretation step is essential for identifying and ranking the most influential ‘digital levers’—the input features that have the most significant impact on predicting compressive strength. By understanding both how the model learned and what it learned, we can extract the actionable knowledge required to proceed confidently with the final application of the Optimization Engine. Figure 13 illustrates the learning process of the deep neural network by plotting the MSE loss for both the training set (blue line) and the validation set (orange line) over 300 training epochs. The plot demonstrates a classic and highly desirable training trajectory. Both the training and validation loss exhibit a steep, rapid decline during the initial epochs, indicating that the model is quickly learning the fundamental patterns within the data. Following this initial phase, both curves gradually flatten out and converge towards a stable, low MSE value for the remainder of the training process. Crucially, the validation loss closely tracks the training loss throughout all 300 epochs, without any significant divergence, unlike the validation loss, which would typically start to increase. In contrast, the training loss continues to decrease. This parallel convergence is a strong indicator that the model is generalizing well to unseen data and that the implemented regularization techniques, such as Dropout, were effective in preventing overfitting. This plot confirms the stability of the training process and validates that the final model represents a well-converged state, ready for robust performance evaluation. Training and validation loss curves for the Digital Twin model over 300 epochs. Figure 14 displays the results of the permutation feature importance analysis, which quantifies the contribution of each input variable by measuring the average decrease in the model’s R-squared performance when the values of that feature are randomly shuffled. This analysis identifies the most critical “digital levers” for predicting compressive strength. The chart reveals a clear hierarchy of influence, with Cement emerging as the most crucial predictor by a significant margin, as its removal causes the largest drop in model accuracy. Following closely is Age, confirming that curing time is another primary driver of strength. Interestingly, Blast Furnace Slag ranks third, indicating that its contribution through complex and non-linear interactions is more significant than that of several other components. The next tier of moderately important features includes Water, Fly Ash, and Superplasticizer. At the bottom of the ranking are Coarse Aggregate and Fine Aggregate, which have the least impact on the model’s predictive performance. This ranking not only validates the model by aligning with known material science principles but also provides a clear roadmap for the Optimization Engine, highlighting which parameters should be prioritized when designing new, sustainable concrete mixes. Permutation feature importance for the Digital Twin model. Optimization engine simulation Having validated the Digital Twin’s predictive accuracy and interpretability, the final phase of this study demonstrates its practical application as an Optimization Engine. This is achieved by moving beyond mere prediction to active simulation, where the model is used to explore and evaluate novel material compositions that have not been physically tested. To provide tangible proof of concept, a hypothetical, highly sustainable concrete mix was designed based on domain knowledge, incorporating low cement content, high proportions of industrial byproducts, and a maximum curing age. By utilizing the trained Digital Twin to predict the compressive strength of this virtual mix and visualizing the complex trade-offs between key variables using 3D surface and contour plots, this section highlights the model’s ultimate utility. It effectively transforms the predictive tool into a design aid that enables the rapid, data-driven development of sustainable, high-performance materials, thereby eliminating the need for costly and time-consuming physical experimentation and completing the Industry 4.0 workflow. Figure 15 provides a powerful demonstration of the Optimization Engine by visualizing the complex, non-linear relationship between the two most important features—Cement and Age—and the predicted compressive strength, as learned by the Digital Twin. The 3D surface represents the model’s predictions for a continuous range of cement and age values, while holding all other features at their average values. The plot clearly illustrates that predicted strength increases along both the scaled Cement and scaled Age axes, which is consistent with the individual linear analyses. However, the true value of this visualization lies in its depiction of the interactive effect; the surface is not a flat plane but a curved, dynamic landscape. The slope of the surface changes dramatically, revealing that the rate of strength gain from increasing Age is substantially more pronounced at higher levels of Cement. Conversely, the benefit of adding more Cement is maximized at longer curing ages, as shown by the steep incline leading to the yellow peak. This visualization effectively serves as a design tool, allowing users to navigate the trade-offs between material composition and curing time to achieve a desired strength, thereby providing clear proof of concept for the Digital Twin’s utility in data-driven, sustainable material design, without the need for physical experimentation. 3D surface plot of the interactive effect of cement and age on predicted compressive strength. Figure 16 offers a complementary two-dimensional perspective of the 3D surface plot, presenting the model’s predictions as a contour map. Each contour line, or iso-strength line, represents combinations of scaled Cement and scaled Age that are predicted to yield the same compressive strength, with the color gradient from dark purple to bright yellow indicating the progression from lower to higher strength values. This visualization serves as an efficient and navigable design map for the Optimization Engine. The plot clearly shows that achieving higher strength requires moving towards the top-right of the graph (higher cement and age). More importantly, it quantifies the non-linear trade-offs between these two critical variables. For instance, an engineer targeting a specific strength can follow the corresponding contour line to identify multiple viable mix designs, such as a mix with a high cement content and a short curing age, or a more sustainable mix with lower cement content that requires a longer curing period. The changing spacing between the contour lines further reveals the model’s learned complexities; the lines are closer together in the lower-left region, indicating rapid strength gains, and spread further apart in the top-right, signaling diminishing returns. This plot provides a clear and actionable tool for making informed, data-driven decisions that balance performance, cost, and sustainability goals. Contour plot illustrating the combined effect of cement and age. Discussion The proposed methodology in this study was designed to empirically validate a conceptual Industry 4.0 workflow for sustainable construction by operationalizing a three-part framework. This framework began with using a public dataset as a Digitalization Proxy to simulate the high-fidelity data from a smart manufacturing facility. This data was then used to develop a Digital Twin, a deep neural network designed to accurately mirror the complex physical process of concrete curing. The final step involved leveraging this validated model as an Optimization Engine to simulate and predict the performance of novel, sustainable material designs without physical experimentation. This end-to-end process offers a tangible, data-driven methodology that directly links digital inputs to sustainable and high-performance outcomes, serving as a practical blueprint for applying advanced predictive analytics in the construction industry. This data-driven approach aligns with established engineering practices where testing data is fundamentally used to select and validate predictive models for structural components42. The rigorous application of this method yielded clear and quantifiable results. The experimental evaluation on a held-out test set demonstrated the superior efficacy of the deep learning Digital Twin, which achieved an R-squared value of 0.9011 and an RMSE of 5.0879 MPa. These metrics indicate that the model could account for over 90% of the variance in compressive strength with an average prediction error of just over 5 MPa. This performance stood in stark contrast to the traditional linear models, which struggled to explain the data’s variance, achieving an R² of only approximately 0.632 and an RMSE of 9.81 MPa, nearly double the error of the Digital Twin. This significant performance gap provides empirical evidence that the complexity of the proposed deep learning approach is not only justified but necessary for achieving high-fidelity predictions in this domain. The research findings, grounded in these performance evaluations, provide several key insights. The primary finding is that the deep learning approach is exceptionally well-suited to function as a high-fidelity Digital Twin for concrete, successfully modeling the material’s complex behavior where simpler models fail. Secondly, the feature importance analysis provided a data-driven hierarchy of influence, confirming the primary roles of Cement and Age while also quantitatively elevating the significance of sustainable industrial byproducts, such as Blast Furnace Slag. This empirically validates their role as critical performance levers, not just fillers. The most powerful finding, however, stems from the optimization simulation, which predicted that a low-cement, high-byproduct mix could achieve a high compressive strength of 69.3 MPa. This demonstrates a significant potential, shifting the industry from a reactive, trial-and-error paradigm to a predictive, proactive one for sustainable material design. The comparative analysis unequivocally identifies the Deep Learning model as the best-performing algorithm, with the ensemble methods (Random Forest and Gradient Boosting) as strong but slightly inferior alternatives. Conversely, Linear and Ridge Regression were clearly the worst-performing algorithms. The reason for this stark divergence in performance lies in the fundamental nature of the problem; highly complex and non-linear interactions govern the relationship between concrete components and compressive strength. The necessity of using advanced methods to calibrate such complex, non-linear models is a common challenge in structural mechanics43. This complexity is rooted in the material’s fundamental nature, with recent studies exploring atomic-scale mechanisms at material interfaces to explain its behavior44. The deep neural network, with its layered architecture and non-linear activation functions, is inherently designed to learn these intricate patterns from data. The ensemble models also possess this capability, but the deep learning model proved slightly more effective. Linear models, by their very definition, are constrained to finding linear relationships and were thus fundamentally incapable of capturing the true complexity of material science, as evidenced by their poor performance metrics obtained during the rigorous evaluation on the held-out test set. First and foremost, the primary limitation is the reliance on a static, laboratory-based public dataset. This study was intentionally framed as a proof-of-concept rather than the validation of a fully operational digital twin, as a true DT requires dynamic, bidirectional integration with a physical system. Our framework successfully simulates the workflow, but it lacks a real-time feedback loop. Therefore, our findings validate a robust predictive model and a foundational blueprint, not an implemented digital twin. While the findings are robust within the scope of this study, several limitations must be acknowledged. First, the dataset, while an excellent proxy for digital data, was derived from controlled laboratory conditions. Real-world construction projects introduce additional variability, such as ambient temperature and humidity, as well as minor inconsistencies in raw material sourcing, which were not captured in the data. Second, this study focused exclusively on a single performance metric: compressive strength. A truly optimal and practical mix design would require a multi-objective optimization that also considers other critical properties, such as workability, durability, embodied carbon, and cost, which were outside the scope of the available data. Finally, the deep learning model, while highly accurate, operates as a “black box,” offering less direct interpretability than simpler models, which can be a barrier to adoption for practitioners who require transparent decision-making processes. Building directly on these limitations, several clear avenues for future research are recommended. The most critical next step is to pursue Real-World Data Integration, aiming to validate and refine this framework using real-time data streams from sensors deployed in active construction sites and batching plants, thereby accounting for environmental variability. Furthermore, the Digital Twin should be expanded into a Multi-Objective Optimization model that can simultaneously predict and optimize for a suite of key properties, such as workability, durability, and embodied carbon, allowing for a more holistic and practical approach to sustainable design. Finally, to address the “black box” challenge and accelerate industry adoption, future work should focus on developing a decision-support tool that integrates the validated model into a user-friendly software application. This application enables material engineers to design, test, and understand virtual mixes interactively, transparently, and intuitively. Conclusion This study aimed to provide rigorous empirical validation of an end-to-end Industry 4.0 workflow for advancing sustainable construction practices. By simulating this workflow with a public dataset, we successfully developed and validated a deep learning-based Digital Twin for predicting concrete compressive strength. The key findings are threefold: (1) the Digital Twin significantly outperformed conventional machine learning models, demonstrating its superior ability to model complex material science; (2) the model identified the most critical parameters for achieving high strength, confirming the significant role of sustainable industrial byproducts; and (3) the model was successfully used as an optimization engine to design and validate a high-performance, sustainable concrete mix in a virtual environment. The primary contribution of this research is the demonstration of a tangible, data-driven pathway from digitalization to enhanced sustainability in construction, providing a methodological blueprint that moves beyond theoretical discussions to a practical application. While acknowledging limitations such as the use of laboratory data and a single-objective focus, this work confirms the promising potential of this approach. Future research should focus on integrating real-world data, expanding the model to a multi-objective framework, and developing user-friendly decision-support tools. Ultimately, this research serves as a validated blueprint, demonstrating that predictive analytics and digital twin technology are powerful and practical tools for engineering the next generation of sustainable construction materials. Author contributions Conceptualization, Project administration, funding acquisition, supervision, writing review-editing, methodology, resources, software: H.F.I., M.K.; Conceptualization, data curation, formal analysis, investigation, validation, visualization, writing original draft preparation: H.F.I., P.X. This manuscript has been read and approved by all authors. Data availability The data presented in this study are available on request from the corresponding author. Declarations Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Haytham F. Isleem, Email: mps565@york.ac.uk, Email: isleemhaytham88@gmail.com Mohammad Khishe, Email: m_khishe@alumni.iust.ac.ir. References Associated Data Data Availability Statement The data presented in this study are available on request from the corresponding author.