Predictive modeling of strength characteristics of sustainable concrete using artificial neural networks with luffa fiber and marble dust Authors: S Anandaraj, S Shaniya, S Swaminathan, S Suganya, G Sindhu Priya, L Balaji, A Dhanalakshmi, Dawit Tafesse Gebreyohannes Published: 2026 ABSTRACT AbstractThis study involves an integrated approach to predict the mechanical properties of luffa fiber and marble dust- based concrete. It employs Artificial Neural Network (ANN) for mechanical properties prediction aiming for a higher accuracy than currently available models. The composite material used marble dust in the proportion 0–40% as fine aggregate replacement and luffa fiber in the proportion 0–2% as the natural reinforcement. Experimental results implied that the composite containing 20% marble dust and 1% luffa fiber exhibited greatest mechanical characteristics- Compressive strength of 34.5 MPa, flexural strength of 6.2 MPa and split tensile strength of 4.25 MPa. This improvement was attributed to enhanced particle packing by marble dust and effective crack bridging by treated luffa fiber. A single multi-output feedforward multilayer perceptron (MLP) ANN consisting of two hidden layers of 64 and 32 neurons with ReLU activation functions and a three-neuron linear output layer was developed for simultaneously modelling the nonlinear interactions between the input variables and strength outputs. The model was trained on 70% of the dataset, with 15% for validation and 15% for testing. The ANN model was able to predict all three mechanical strength properties simultaneously with a high degree of accuracy as demonstrated by R² values of 0.89 (compressive strength), 0.94 (flexural strength), and 0.96 (split tensile strength) for the training data sets and small root mean square error (RMSE) values and negligible bias. The 100% a20 score indicated that all the samples from the predictions fell within ± 20% of the actual experimental values, demonstrating good robustness or generality. This research basically aims to address a major gap in the existing works by exploring the limited application of ANN in predicting the performance of hybrid sustainable concrete mixes incorporating both marble dust and plant-based fibers by developing a predictive model which is capable of capturing the complex interactions between multiple factors and enhancing the prediction efficiency, minimizing the dependency on extensive experimental investigations thereby promoting data driven evaluation of strength characteristics in sustainable concrete composites.Keywords: Strength prediction, Luffa fiber, Marble dust, ANNSubject terms: Engineering, Materials science Introduction The rate of concrete production has quadrupled over the last three decades, reaching a rate of 26 Gt annually. Additionally, the demand for aggregates, estimated to be around 20 Gt annually, has surpassed the total amount of fossil fuel extracted globally1. However, cement and concrete contribute to a total of 7–9% of the total CO2 emissions, which comes from the manufacturing process and fuel usage2,3. Sand and aggregate mining has also been linked to the destruction of ecosystems, fueling conflicts and the emergence of a “sand crisis”4. Given that concrete is difficult to replace functionally, the approach is to reduce the content of cement, make use of by-products and wastes, and make the mix more efficient instead of replacing concrete altogether5,6. In this case, the use of industrial by-products and construction wastes is beneficial to the circular economy and can reduce the potential for global warming by 20–40%7. Of these materials, the waste generated in the processing of marble is in large quantities and is non-biodegradable; using it in concrete can reduce the waste disposed of in landfills and can save natural sand and cement8. Research suggests that the addition of waste marble powder up to 5–10% can enhance the compressive strength by 5–12% using the filler effect, durability can be improved in some cases. However, higher percentages can reduce the strength of the concrete in an optimized manner9,10. Similarly, partial replacement of sand by marble aggregates can improve the compressive strength, flexural strength, and impact resistance of self-compacting concrete. Life cycle assessments indicate that there can be a reduction in global warming potential up to 20%, improving sustainability indices in the case of marble waste used appropriately11. In parallel, natural fibres have emerged as a sustainable alternative for synthetic or steel fibres for crack control and toughness in cement and geopolymer concrete mixes. Literature surveys have confirmed that natural fibres such as cotton, sisal, jute, bamboo, and many more can be used to reinforce concrete successfully for eco-friendly and waste-free composites12,13. In particular, marble dust contains very fine particles of CaCO₃, which are inert fillers used to refine pores and particles for improved packing efficiency, resulting in reduced porosity and water absorption for increased density and compactness14–16. The marble dust at Optimal dosages increases the compressive strength by 10–26%, and the split tensile strength increases by 10–18% compared to control mixes17,18. Marble dust concretes also have reduced permeability, better sulphate resistance, better resistance to freeze-thaw actions, and better salt crystallization resistance due to the denser microstructure19. On the contrary, Luffa fibres have high porosity, large internal spaces, and high-water absorption capacity, which is above 200%, thus they can function as internal curing agents, thereby increasing hydration time, reducing autogenous shrinkage up to 57%, and increasing hydration products and matrix compactness in high-strength mortar20. Alkali (NaOH) and benzoyl chloride treatments involve surface treatments that remove hemicellulose/lignin, making the surface rough and increasing the tensile strength and modulus by about 140–150%. This enhances the bond properties of the fibres, thus improving the composite properties by increasing flexural strength and tensile modulus21,22. In this context, Artificial Neural Networks have come to be commonly employed in civil and concrete engineering as data-driven tools for modeling complex material behavior in a more efficient way than traditional regression or empirical approaches. This is because the complex properties of concrete are affected by numerous variables in nonlinear ways. Earlier work has demonstrated the effectiveness of intelligent prediction frameworks for cementitious materials; for example, nearest-neighbourhood clustering with PSO optimization has been applied to predict 28-day compressive strength of pozzolanic concrete23, fuzzy logic has been integrated with experimental push-out data to model bond strength in fiber-reinforced self-compacting concrete (FRSCC)24, and slant-shear fuzzy logic models have been used to evaluate repair-interface bond in SCC and FRSCC25. More recently, AutoML combined with DEM simulation has enabled prediction of macroscopic behaviour in cementitious composites with variable frictional parameters26. The ANN based frameworks control the cement content, SCMs, aggregates, water, admixtures, and curing to attain the desired strength while cutting down on cost and CO2 and limiting the trial-and-error processes. ANN and other ML algorithms are used to predict the modulus of elasticity, slump, shrinkage, stiffness, and durability or service life of RC structures27,28. The objectives of the present study are to develop an ANN model for predicting the mechanical properties of marble dust and luffa fibre concrete. To develop and use an experimental data set for concrete mixes with varying levels of marble dust replacement and luffa fibre content as input to the ANN model, as used in earlier ANN models for predicting concrete strength29. The model is developed to predict three important mechanical properties of concrete simultaneously. These are Compressive strength, split tensile strength, and Flexural strength using a single multi-output model, as adopted in earlier studies on lightweight concrete, metakaolin concrete, fibre-reinforced concrete, high-performance concrete, etc30. The predicted results are compared with the corresponding experimental results for compressive, split tensile, and flexural strengths to assess the accuracy and error range in the prediction, similar to the previous ANN concrete research in which the predicted and measured values were found to be in close conformity. For the evaluation of the developed ANN model using statistical parameters such as R, R², RMSE, MAE, and MAPE for training, validation, and testing datasets, similar to the ANN-based concrete strength prediction research31,32. Nevertheless, inspite of the extensive application of ANN and other machine learning models in concrete materials, the majority of the existing research is focused on conventional or singly modified concretes, while the specific mix of marble dust and luffa fibres is not taken into account. The majority of the research carried out with the application of ANN/ML is focused on the following types of concretes: Conventional or steel fibre concretes, often without waste mineral additions33. Or modified with marble dust or sludge with polypropylene fibres34. Or with stone or quarry dust with one type of synthetic or natural fibres35. While research on natural fibers such as jute, bamboo, coconut, and sisal with SCMs has been conducted, luffa fibers have rarely been used and the combinations are limited36. Also, there is limited research carried out using the combined input of marble dust and natural fibres under a unified framework; the investigations carried out using the input of marble dust and fibres have used synthetic fibres not bio-fibers10. Existing models rarely incorporate fibre content, waste mineral replacement level, and curing age as parameters in the prediction of sustainable fibre-reinforced mix designs; in fact, some models ignore the variables of fibre content or consider only non-waste mixes37. Consequently, there is an evident gap in terms of research, with regards to creating ANN-based models to forecast mechanical properties of marble dust concrete with luffa fibre, with consideration of combined effects of fibre volume, marble dust replacement percentage, and curing age in a predictive modelling framework. Materials and methods Materials In the current study, traditional concrete ingredients have been used alongside marble dust, which was employed to replace fine aggregates partially. The material selection process has been conducted in such a manner that they become compatible and appropriate for M25 grade concrete. Prior to mixing, all ingredients were subjected to testing in compliance with standard procedures. Cement For the experiment, OPC of Grade 53 was used as the major binder. The OPC was obtained from one supplier to ensure consistency in all the properties of the cement throughout the experimentation process. Physical properties of the cement affect its performance during hydration reaction. The properties of cement were analyzed by conducting various tests including consistency test, setting time, and fineness is shown in Table 1. Table 1Physical properties of cement.PropertyValueSpecific gravity3.15Standard consistency (%)31Initial setting time (min)45Final setting time (min)600Fineness (% retained)3.8 Physical properties of cement. The relatively high fineness of the cement results in higher surface area, hence improved hydration and early gain in strength. The setting times recorded were satisfactory enough to allow proper mixing, placement, and compaction. Fine aggregate and marble dust River sand which meets the standards of Zone III grading was used as a fine aggregate. This acts as a filling agent whose role is to enhance the flow and cohesion of concrete. Marble dust was used as a partial substitute for the fine aggregate, and it was sourced from the marble industry. The use of marble dust not only tackles the problems related to disposal of waste but also enhances the microstructure of concrete because of its fine particles. The presence of high amounts of calcium carbonate in marble dust increases density without leaving any voids. Table 2 presents the chemical composition of marble dust. Table 2Chemical composition of marble dust.CompoundContent (%)CaO44.30SiO₂8.10Al₂O₃0.70Fe₂O₃0.70MgO5.60SO₃0.04Loss on ignition40.50 Chemical composition of marble dust. Since both values are high, they are due to calcium carbonate. Calcium carbonate is not a pozzolanic material, but a material that will work as filler (Table 3). Table 3Physical properties of fine aggregate and marble dust.PropertyFine aggregateMarble dustSpecific gravity2.732.86Bulk density (g/cm³)1.641.30Water absorption (%)0.841.26Fineness modulus3.00–Specific surface area (m²/kg)35335 Physical properties of fine aggregate and marble dust. As the marble dust has higher surface area than the natural sand, there will be better packing and densification, although it will cause increased water demand. Coarse aggregate The coarse aggregates that were used for construction were crushed stone aggregates of size 12.5 mm and 20 mm. The coarse aggregates are the main components of concrete and determine the strength and durability of the same. The aggregates were washed and saturated with water before usage. Table 4 represents the physical properties of coarse aggregate. The relatively low water absorption indicates good quality aggregates with minimal porosity, which contributes to improved durability of concrete. Table 4Physical properties of coarse aggregate.PropertyValueSpecific gravity2.64Water absorption (%)0.65Bulk density (kg/m³)1580Fineness modulus7.17 Physical properties of coarse aggregate. Luffa fibers As shown in Fig. 1, Luffa fibers are natural and bio-degradable fibers, have been employed for reinforcing concrete by virtue of its tensile strength and crack resistance qualities. Being hydrophilic, luffa fibers normally absorb water and deteriorate the adhesion between fiber and matrix. Luffa fiber. In order to address this drawback, the fibers were chemically treated with NaOH (sodium hydroxide). The fibers of diameter 0.4 mm were taken and initially cut into uniform pieces of length 20 mm each (l/d = 50) and immersed in an aqueous NaOH solution (5 wt%) for 24 h at ambient temperature (27 ± 2 °C). The fibers after soaking is removed and washed repeatedly using distilled water until the pH is neutralized which was verified with the help of a pH paper. The neutralized fibers were then oven-dried at 60 °C for 24 h to a constant mass before use. This procedure is consistent with the protocol reported by Zidi et al.22. The water absorption of the treated luffa fibers was measured by the 24-hour immersion method and found to be approximately 215%, which is significantly higher than that of conventional synthetic fibers but was managed by pre-soaking the fibers to a saturated surface-dry (SSD) condition prior to mixing in order to minimize undue water demand from the concrete matrix. The marble dust specific surface area of 335 m²/kg (Table 3) confirms its role as a micro-filler; X-ray diffraction (XRD) analysis of the marble dust confirmed a CaCO₃ content exceeding 96%, consistent with a predominantly inert filler behavior with minor pozzolanic contribution. Such treatment helps in removing any impurities on the surface, reduces the amount of hemicellulose and lignin, and smoothening the fiber surface, enhancing the ability of concrete to bridge cracks and the properties of luffa fiber is shown in Table 5. Table 5Properties of luffa fibers.PropertyValueDensity (g/cm³)1.30Tensile strength (MPa)180Elastic modulus (GPa)10Length (mm)20Diameter (mm)0.4Aspect ratio (l/d)50Water absorption (%)~ 215%(measured by 24 h immersion; treated fibers pre-soaked to SSD before mixing) Properties of luffa fibers. ~ 215% (measured by 24 h immersion; treated fibers pre-soaked to SSD before mixing) Water Potable water free from any contaminants was utilized in the process of mixing and curing of concrete. Water quality is an important factor that affects concrete hydration and its general durability. Mix proportions The mix design for concrete was done following the requirement of Grade M25. The water cement ratio used was restricted to 0.46(186 L) and the ratio of cement to fine aggregate to coarse aggregate was obtained as 1:1.38:2.45. Thus, the derived quantity of cement in the study is 465 kg/m3, fine aggregate is 684 kg/m3 and coarse aggregate is 1108 kg/m3. The marble dust was used to partially replace the fine aggregate material in different proportions (0%, 10%, 20%, 30%, 40%) while luffa fiber was introduced in the proportion 0%, 0.5%, 1.0%, 1.5%, 2% by mass of cement. Fiber dosage is expressed as a percentage by mass of total cementitious content. Table 6 presents the complete mix-design matrix showing the exact constituent quantities (cement, water, fine aggregate, coarse aggregate, marble dust, and luffa fiber content in kg/m³) for each of the 25 individual mixtures tested. All quantities are reported per cubic metre of concrete. Fiber content is expressed both as a percentage by mass of cement and in absolute kg/m³ to avoid ambiguity. Table 6Complete mix-design matrix for all 25 concrete mixtures.Mix IDMD (%)LF (%)Cement (kg/m³)Water (L/m³)FA (kg/m³)CA (kg/m³)MD (kg/m³)LF (kg/m³)Slump (mm)M10%0%465186684.011080.00.082M20%0.5%465186684.011080.02.3372M30%1.0%465186684.011080.04.6561M40%1.5%465186684.011080.06.9745M50%2.0%465186684.011080.09.330M610%0%465186615.6110868.40.078M710%0.5%465186615.6110868.42.3368M810%1.0%465186615.6110868.44.6556M910%1.5%465186615.6110868.46.9738M1010%2.0%465186615.6110868.49.325M1120%0%465186547.21108136.80.068M1220%0.5%465186547.21108136.82.3358 M13 20% 1.0% 465 186 547.2 1108 136.8 4.65 44 M1420%1.5%465186547.21108136.86.9728M1520%2.0%465186547.21108136.89.318M1630%0%465186478.81108205.20.061M1730%0.5%465186478.81108205.22.3350M1830%1.0%465186478.81108205.24.6537M1930%1.5%465186478.81108205.26.9722M2030%2.0%465186478.81108205.29.312M2140%0%465186410.41108273.60.054M2240%0.5%465186410.41108273.62.3343M2340%1.0%465186410.41108273.64.6531M2440%1.5%465186410.41108273.66.9718M2540%2.0%465186410.41108273.69.38Mix M13 (bold) represents the optimal combination (20% MD, 1.0% LF). MD = Marble Dust; LF = Luffa Fiber; FA = Fine Aggregate; CA = Coarse Aggregate. W/C ratio = 0.46 (constant). Fiber dosage expressed as % by mass of cement. Complete mix-design matrix for all 25 concrete mixtures. Mix M13 (bold) represents the optimal combination (20% MD, 1.0% LF). MD = Marble Dust; LF = Luffa Fiber; FA = Fine Aggregate; CA = Coarse Aggregate. W/C ratio = 0.46 (constant). Fiber dosage expressed as % by mass of cement. Experimental procedure Mixing and casting Concrete mixing was carried out using a mechanical mixer to ensure uniform distribution of materials. Initially, dry components including cement, fine aggregate, coarse aggregate, and marble dust were mixed thoroughly. Water was then added gradually to achieve the desired consistency. Luffa fibers were gradually added into the mix towards the final part of the mixing process to ensure proper dispersion and to avoid fiber agglomeration. Fresh-state properties were measured immediately after mixing in accordance with IS 1199:1959. Workability was assessed by the slump cone test. The control mix (0% marble dust, 0% fiber) exhibited a slump of 82 mm. Progressive substitution of fine aggregate by marble dust up to 40% reduced the slump to 78 mm, 68 mm, 61 mm and 54 mm owing to the increased specific surface area of marble dust. Addition of 0.5%, 1.0%, 1.5% and 2% luffa fiber reduced the slump to 72 mm, 61 mm, 45 mm, and 30 mm respectively, confirming the workability-loss trend attributed to fiber water absorption and physical obstruction of particle flow. Fresh density was measured for all mixes using the unit weight container method and ranged from 2,381 to 2,412 kg/m³. Air content was not measured separately but is implicitly reflected in the fresh density values. Slump values for all 25 mixtures are tabulated in Table 6(complete mix-design matrix). Concrete specimens were cast in standard moulds: Cubes (150 mm × 150 mm × 150 mm) for compressive strength. Cubes (150 mm × 150 mm × 150 mm) for compressive strength. Cylinders (150 mm × 300 mm) for split tensile strength. Cylinders (150 mm × 300 mm) for split tensile strength. Prisms (100 mm × 100 mm × 500 mm) for flexural strength. Prisms (100 mm × 100 mm × 500 mm) for flexural strength. The cast specimens were kept undisturbed for a period of 24 h, demoulded, and then subjected to curing in water with a temperature range of 27 ± 2 °C until the day of testing. Mechanical testing Mechanical properties were evaluated at curing ages of 7, 14, and 28 days. For each mix and curing age, 3 specimens were casted and tested. The average for the 3 replicate samples was reported as the indicative value for each condition. Standard deviation (SD) and coefficient of variation (CV) for each mix combination at 28 days are presented in Table 7. The optimum at 20% marble dust and 1% luffa fiber (M13) was confirmed statistically; the CV for compressive, flexural, and split tensile strengths at this optimum mix did not exceed 4.2%, indicating acceptable repeatability consistent with IS 456:2000 and IS 516:1959 guidelines. Table 7Mean strength, standard deviation and coefficient of variation at 28 days for all 25 mixtures.MixMD%LF%fc Mean (MPa)fc SDfc CV%ff Mean (MPa)ff SDff CV%ft Mean (MPa)ft SDft CV%M10%0%28.21.073.84.80.173.53.10.124.0M20%0.5%28.90.923.24.90.153.03.180.113.5M30%1.0%29.41.033.55.00.142.93.250.123.8M40%1.5%28.71.184.14.850.183.83.150.134.2M50%2.0%27.61.244.54.650.24.22.980.144.8M610%0%29.80.893.05.20.152.83.450.113.2M710%0.5%30.60.862.85.40.142.53.550.113.0M810%1.0%31.50.983.15.60.152.73.70.123.3M910%1.5%30.81.113.65.450.173.23.580.143.8M1010%2.0%29.51.184.05.150.193.73.350.144.2M1120%0%31.20.92.95.60.152.63.80.123.1M1220%0.5%33.10.892.75.90.142.44.050.122.9 M13 20% 1.0% 34.5 1.1 3.2 6.2 0.17 2.8 4.25 0.18 4.2 M1420%1.5%33.21.133.46.00.183.04.10.153.6M1520%2.0%31.81.344.25.70.223.93.880.174.5M1630%0%30.51.013.35.30.163.03.60.133.5M1730%0.5%31.40.943.05.50.152.73.720.123.2M1830%1.0%32.11.093.45.70.183.13.850.143.7M1930%1.5%31.31.193.85.50.193.53.70.154.0M2030%2.0%29.91.324.45.20.214.03.50.164.6M2140%0%28.81.073.75.00.173.43.30.133.9M2240%0.5%29.71.013.45.10.163.13.420.123.6M2340%1.0%30.31.093.65.30.173.33.550.143.9M2440%1.5%29.51.184.05.10.183.63.40.154.3M2540%2.0%28.11.324.74.80.214.33.150.165.0M13 = optimal mix. fc = compressive strength; ff = flexural strength; ft = split tensile strength. n = 3 replicates per mix. SD = standard deviation; CV = coefficient of variation (%). All values at 28-day curing age. Mean strength, standard deviation and coefficient of variation at 28 days for all 25 mixtures. M13 = optimal mix. fc = compressive strength; ff = flexural strength; ft = split tensile strength. n = 3 replicates per mix. SD = standard deviation; CV = coefficient of variation (%). All values at 28-day curing age. Compressive strength: Split tensile strength: Flexural strength: Such tests help determine the bearing load properties of concrete. However, the results from the 28th day test results were used as the primary dataset for ANN training. The dataset used for ANN training was assembled in two stages to ensure both experimental fidelity and adequate sample density for reliable model generalisation. In the first stage, three replicate specimens were cast and tested for each of the 25 mix combinations at 28 days, yielding three independently measured strength values (compressive, flexural, and split tensile) per mix. Each replicate constitutes a distinct data point, preserving specimen-level variability rather than collapsing replicates into a single averaged value. In the second stage, the experimentally bounded range of strength responses for each mix was systematically expanded through linear interpolation between the minimum and maximum replicate values at uniform intermediate intervals. This interpolation was performed strictly within the experimentally established upper and lower bounds of each mix’s three measured replicates, so no extrapolation beyond the observed data envelope was introduced. The interpolated points capture the natural within-batch variability that would be expected from additional replicate testing and are physically consistent with the measured scatter quantified in Table 7. This two-stage approach resulted in a substantially enlarged dataset that retains the statistical distribution of the experimental observations while providing the sample volume necessary for stable partitioning into training, validation, and testing subsets. The coefficient of variation across replicates at 28 days is reported in Table 7 and confirms the tight repeatability (CV not exceeding 5%) that underpins the reliability of the interpolation bounds. Multi-age (7- and 14-day) results were not included in the primary ANN model because the objective was to develop a mix-design prediction tool for 28-day design strength, consistent with standard structural design practice; curing age could be added as an additional input in future extended models. ANN-based strength prediction For modelling the non-linear relation between the constituent materials and their strengths in case of concrete mixed with luffa fibers and marble dust, the ANN technique was employed. There is non-linearity associated with the interaction of fiber reinforcing and matrix densification because of the presence of marble dust, which cannot be modelled using linear regression. One feature of ANN models is their ability to approximate nonlinear functions, which makes them suitable for modelling multivariate data. A single multi-output ANN model was developed to simultaneously predict all three mechanical strength properties – compressive strength, flexural strength, and split tensile strength – from the same input variables. A shared-representation multi-output architecture is adopted because the three strength properties are governed by the same underlying mix design variables and their interactions; training a unified model exploits the correlations among outputs, improves generalisation, and reduces modelling complexity compared with maintaining separate independent models. The definitive input vector for the multi-output model comprises exactly 5 variables: (1) marble dust replacement level (% by mass of fine aggregate), (2) luffa fiber content (% by mass of cement), (3) cement content (kg/m³), (4) fine aggregate content (kg/m³), and (5) coarse aggregate content (kg/m³). Curing age was fixed at 28 days for all observations and is therefore a constant, not an independent input variable. Marble dust content in kg/m³ and fiber content in kg/m³ are directly derivable from variables (1)–(3) and are therefore not entered as separate inputs to avoid redundancy. To increase numerical stability and training efficiency, all input attributes were scaled using Min-Max Scaling method as: where and refer to the minima and maxima of each attribute. Such a scaling process would ensure the same scale for all the attributes and avoid dominance of higher magnitude attributes38. The dataset is split into training, validation, and testing datasets in the ratio of 70%, 15%, and 15%, respectively. The complete dataset (comprising 75 original replicate measurements plus interpolated points, all at 28 days) was first assembled and Min-Max normalised using statistics computed exclusively from the training subset; no statistics derived from the validation or test subsets were used in scaling, thereby preventing data leakage. The data were then split randomly at the mixture level: all data points belonging to a given mix design were assigned exclusively to one of the three partitions (training, validation, test), ensuring that information from any single mixture did not appear in more than one partition. The test set was kept entirely withheld throughout the training and hyperparameter-selection process, and was only used once -for final performance evaluation and after the model architecture and early-stopping criterion had been fixed. The convergence plots shown in Figs. 4, 8, and 11 display training loss and validation loss per epoch; the test set metric reported in Tables 9, 10, 11 represents this single held-out evaluation and was not used at any point during model selection or training. This protocol ensures the test set remained fully independent and that the reported test R², RMSE, and other statistics are unbiased estimates of out-of-sample predictive performance. Compressive strength analysis: (a) Experimental vs. predicted values, (b) histogram plot, (c) Taylor diagram and (d) Model performance Radar chart. Flexural strength analysis: (a) Experimental vs. predicted values, (b) histogram plot, (c) Taylor diagram and (d) Model performance Radar chart. Regression plots for split tensile strength. Table 9ANN performance metrics – flexural strength.R2RMSEWINSEPBIASa20Train0.9422860.1257550.9842770.942286− 0.01652100Validation0.9428950.1319730.9842340.942895− 0.00716100Test0.9326480.1277070.981470.9326480.321298100All0.9411490.1270.9839070.9411490.035301100 ANN performance metrics – flexural strength. Table 10ANN performance metrics – split tensile strength.R2RMSEWINSEPBIASa20Train0.9572410.0802310.9884910.957241-0.11485100Validation0.955910.0865610.987950.95591-0.09663100Test0.9534550.07980.9872610.9534550.203117100All0.9565310.0811480.9882410.956531-0.06463100 ANN performance metrics – split tensile strength. A Feedforward MLP structure was used having one input layer (five neurons, one per input variable), two hidden layers, and one output layer with three neurons corresponding to the three strength outputs (compressive strength, flexural strength, and split tensile strength). Each unit carries out a weighted sum and then applies a non-linear operation as follows: where represents the connection weights, is the bias and is the activation function. The hidden layers employed the Rectified Linear Unit (ReLU) activation function: and it was quite effective in dealing with nonlinearities as well as in solving the vanishing gradients problem38. The activation function used by the output layer was a linear one. The optimization process was carried out by backpropagation, while the loss function optimized was the mean squared error (MSE): The weights were iteratively updated as: where η is the learning rate, while ∇L denotes the gradient of the loss function. To avoid overfitting and enhance the generalization performance of the model, early stopping using the validation loss was applied. The training of the network took place within a few epochs38. Pseudo Code: The overall ANN modelling procedure adopted in this study is summarized below: Input: Marble dust replacement %. Marble dust replacement %. Luffa Fiber content (%). Luffa Fiber content (%). Cement content. Cement content. Fine aggregate content. Fine aggregate content. Coarse aggregate content. Coarse aggregate content. Output: Compressive Strength (MPa). Compressive Strength (MPa). Split Tensile Strength (MPa). Split Tensile Strength (MPa). Flexural Strength (MPa). Flexural Strength (MPa). Step 1: Load dataset D. Step 2: Remove irrelevant identifiers (e.g. Mix ID). Step 3: Separate input features X and output Y. Step 4: Normalize input features using Min-Max scaling. Step 5: Split dataset: Training (70%), Validation (15%), Testing (15%). Step 6: Define ANN architecture: Input layer. Hidden layer (64 neurons, ReLU). Hidden layer (32 neurons, ReLU). Output layer (3 neurons, linear activation) [Compressive Strength, Flexural Strength, Split Tensile Strength]. Step 7: Compile model: Loss = MSE. Optimizer = Adam (learning rate = 0.001). Step 8: Train model. For each epoch: Forward propagation. Compute loss. Backpropagation and weight update. Validate model. Apply early stopping if required. Step 9: Generate predictions on all datasets. Step 10: Evaluate model using R², RMSE, NSE, WI, PBIAS, a20. Step 11: Visualize results (actual vs. predicted, error plots). Step 12: Save trained model. The performance of the designed ANN model was measured using several statistical parameters for a thorough assessment of its precision, dependability, and forecasting ability. No single parameter could be sufficient to estimate model efficiency, and hence several parameters were considered for the same purpose. The coefficient of determination (R2) was taken to estimate the percentage of variance in experimental data that is accounted for by the model. A value of R2 closer to one shows a high correlation between the predicted and actual data points: where are the experimental values, denotes the predicted values, and and is the average value of experimental values. The statistic is commonly used in assessing prediction models based on regression. For determining the size of the error in the prediction, the Root Mean Square Error (RMSE) was determined. The statistic determines the average error between the predicted and actual values and is especially sensitive to larger errors: The lower the value of RMSE is, the more accurate the model prediction is. To assess the effectiveness of our model against the mean observed value, we used the Nash-Sutcliffe Efficiency (NSE). Its values can vary within the range between -∞ and 1, where 1 shows maximum efficiency: This index is frequently employed in the validation process of models using environmental and engineering data sets39. In order to evaluate the degree of similarity between the predicted and observed data, Willmott’s Index of Agreement (WI) was calculated. WI is a normalized index of error ranging from 0 to 1, with higher values indicating better agreement. The Percent Bias (PBIAS) was used to find the mean tendency of the predicted values to be larger or smaller than the observed ones: Prediction is considered to be unbiased when PBIAS is near zero; with the formula as written (numerator = Σ(yi − i)), a positive PBIAS indicates that the model systematically under-predicts (observed exceeds predicted on average), while a negative PBIAS indicates systematic over-prediction. The reported PBIAS values near zero across all three outputs confirm negligible systematic bias40. Apart from these measures that have been discussed previously, a20 index has been included in order to measure robustness in terms of the following: Where is the number of predictions that comply with 0.8y_i ≤ _i ≤ 1.2y_i. The larger the a20, the greater the predictability of the model. The combination of these two statistics enabled us to provide a complete assessment of the performance of the ANN model. Results and discussion Compressive strength test results Figure 2 depicts the plot for compressive strength results. The highest peak is observed at Luffa fiber =1% and Marble dust =20% with compressive strength 34.5 MPa. This region appears as the bright yellow peak on the surface indicating the optimal mix combination. The strength on addition of luffa fiber increased initially from 0 to 1% but beyond 1% it decreased gradually. This showed that small amounts of luffa fiber seemed to improve the bonding and crack resistance. However, increased dosage led to poor workability, fiber clustering and weak interfacial bonding. Similarly, Marble dust beyond 20% addition showed decreased strength and reduced structural integrity. Observed experimental result for compressive strength test. The ANN model shows excellent predictive power and does an excellent job of capturing the nonlinear nature of the relationship between the 5 input variables and compressive strength as one of its three simultaneous outputs. Figures 3 and 4 represents, when applying in regression analysis, predicted and experimental values are reasonably close together as evidenced by high correlation coefficients, R = 0.944 (training), 0.948 (validation), and 0.942 (testing). R² = 0.88 overall indicates strong agreement and satisfies acceptable reliability criteria is shown in Table 8. Regression plots for compressive strength. Table 8ANN performance metrics – compressive strength.R2RMSEWINSEPBIASa20Train0.8895480.4634240.9709430.8895480.167267100Validation0.8976410.4679850.9722450.8976410.205192100Test0.8811130.4673210.9680010.8811130.333843100All0.8897370.4646980.9707530.8897370.197915100 ANN performance metrics – compressive strength. Through evaluation of Mean Squared Error (MSE) in Fig. 5, the training performance of the ANN model shows that MSE values quickly decrease from high (over 100) to below 1 over approximately 70 epochs, demonstrating effective learning and convergence without any overfitting of the model. In addition, evaluation with other statistical performance measures (i.e., RMSE, NSE, Willmott’s Index (WI)), all support the excellent agreement between predicted and measured values. The a20 index, approaching 100%, also supports that most predictions fall within ± 20% of measured results. The data point concentration around the ideal prediction line and the consistently decreasing error trend confirm that the ANN model is robust, precise, and reliable in predicting compressive strength and in capturing increasing or decreasing trends for different amounts of luffa fiber and marble dust. Analysis of MSE and epoch convergence under compression. Flexural strength test results Figure 6 depicts the change in flexural strength with respect to how much of each luffa fiber and marble dust mix was used. The greatest recorded flexural strength was 6.2 MPa and occurred when using 1% luffa fiber combined with 20% marble dust, which can be seen as a bright yellow peak (high value) on the response surface. This flexural strength illustrates that adding small amounts of luffa fiber will provide an increase in flexural strength through the bridging of microcracks, as well increasing the weight-carrying capacity of these samples when bending rather than breaking. As the amount of luffa fiber increases to greater than 1%, the flexural strength begins to decrease because the fibers become less well dispersed through the concrete mixture, resulting in clusters of fibers forming at different places in the resulting sample, leading to less workable concrete and the weakening of the bond between the fiber and the concrete matrix. Figure 7 shows the regression analysis for flexural strength. Observed experimental for flexural strength test. Regression plots for flexural strength. The ANN model accurately predicts flexural strength, successfully modeling the nonlinear relationship between predictor variables (luffa fiber and marble dust incorporated into concrete) and response variable (flexural strength). Figure 8 represents the results showed a strong coefficient of determination, R² = 0.942 (training), 0.943 (validation), and 0.933 (testing), meaning that experimental results and predicted values were highly correlated and it is shown in Table 9. Figure 9 represents the analysis of MSE and Epoch Convergence under flexural loading. The RMSE values ranged from 0.125 to 0.132, indicating low prediction error, and the Nash–Sutcliffe Efficiency (NSE) values further confirmed reliability in the model. The Willmott’s Index (WI) value was > 0.98; thus, there was excellent agreement between observed values and predicted values. Furthermore, PBIAS values ranged from − 0.0165 to 0.3213, indicating low overall systematic bias. The a20 index of 100% shows that predicted values fell within ± 20% of experimental results. In summary, this ANN model is very accurate, robust, and provides good generalization capabilities; therefore, the model can be used reliably to predict flexural strengths with different percentages of luffa fiber and marble dust. Analysis of MSE and epoch convergence under flexural loading. Split tensile strength results The range of split tensile strength values based on the different amounts of luffa fiber and marble dust are shown in Fig. 10. The peak value of 4.25 MPa occurs at 1% of luffa fiber and 20% of marble dust on the response surface where it represents an optimum mix ratio. In addition to providing an optimal value, the observed values indicate that as more luffa fiber is incorporated up to 1%, there will be an increase in split tensile strength; afterwards, a decrease is experienced. Therefore, adding a limited quantity of luffa will increase the tensile resistance of the concrete by improving crack bridging as well as delaying crack propagation. With high percentages of luffa fiber present, the effectiveness of the concrete will be reduced because of poor workability characteristics, fibre segregation, and lack of strong interfacial bond at the matrix-fiber interface. Observed experimental results for split tensile strength test. The ANN model shows excellent predictive power and does an excellent job of capturing the nonlinear nature of the relationship between material composition and split tensile strength as one of its three simultaneous outputs. When applying regression analysis in Figs. 11 and 12, predicted and experimental values are reasonably close together as evidenced by high correlation coefficients, R = 0.979 (training), 0.979 (validation), and 0.978 (testing). R² = 0.957 overall indicates strong agreement and satisfies acceptable reliability criteria and it is shown in Table 10. Split tensile strength analysis: (a) experimental vs. predicted values, (b) histogram plot, (c) Taylor diagram and (d) model performance radar chart. Through evaluation of Mean Squared Error (MSE), the training performance of the ANN model shows that MSE values in Fig. 13 quickly decrease from high (over 100) to below 1 over approximately 20 epochs, demonstrating effective learning and convergence without any overfitting of the model. In addition, evaluation with other statistical performance measures (i.e., RMSE, NSE, Willmott’s Index (WI)), all support the excellent agreement between predicted and measured values. The a20 index, approaching 100%, also supports that most predictions fall within ± 20% of measured results. The data point concentration around the ideal prediction line and the consistently decreasing error trend confirm that the ANN model is robust, precise, and reliable in predicting split tensile strength and in capturing increasing or decreasing trends for different amounts of luffa fiber and marble dust. Analysis of MSE and epoch convergence under tensile loading. Conclusion When optimal proportions of marble dust and luffa fibers are used in concrete, it contributes to greater mechanical performance and environmental benefits through the re-use of industrial waste materials as fine aggregate replacement, consistent with circular-economy principles. Formal life-cycle assessment or carbon quantification was not conducted in this study. When optimal proportions of marble dust and luffa fibers are used in concrete, it contributes to greater mechanical performance and environmental benefits through the re-use of industrial waste materials as fine aggregate replacement, consistent with circular-economy principles. Formal life-cycle assessment or carbon quantification was not conducted in this study. Using a proportion of 20% marble dust and 1% luffa fiber represented the optimum concrete mix, which resulted in compressive strengths (34.5 MPa), flexural strengths (6.2 MPa), and split tensile strengths (4.25 MPa). Using a proportion of 20% marble dust and 1% luffa fiber represented the optimum concrete mix, which resulted in compressive strengths (34.5 MPa), flexural strengths (6.2 MPa), and split tensile strengths (4.25 MPa). Marble dust contributed to strength improvements by both filler effects and better particle packing, thus providing a denser microstructure than the control (inferred from macroscopic strength data and literature on analogous systems; microstructural confirmation via SEM or MIP was not conducted in this study); luffa fibers improved crack resistance through effective crack bridging (inferred from flexural and split-tensile response). Marble dust contributed to strength improvements by both filler effects and better particle packing, thus providing a denser microstructure than the control (inferred from macroscopic strength data and literature on analogous systems; microstructural confirmation via SEM or MIP was not conducted in this study); luffa fibers improved crack resistance through effective crack bridging (inferred from flexural and split-tensile response). Increased quantities of luffa fibers (> 1%) decreased concrete workability and increased clustering of fibers resulting in reduced interfacial bonding and reduced overall strength. Increased quantities of luffa fibers (> 1%) decreased concrete workability and increased clustering of fibers resulting in reduced interfacial bonding and reduced overall strength. The single multi-output ANN developed (feedforward MLP with five inputs, two hidden layers of 64 and 32 neurons with ReLU activation, and a three-neuron linear output layer) demonstrated the ability to simultaneously capture non-linear relationships between the mix design input variables and all three mechanical strength properties of the concrete. The single multi-output ANN developed (feedforward MLP with five inputs, two hidden layers of 64 and 32 neurons with ReLU activation, and a three-neuron linear output layer) demonstrated the ability to simultaneously capture non-linear relationships between the mix design input variables and all three mechanical strength properties of the concrete. The ANN achieved high predictive accuracies (R² values of up to 0.96, low RMSE, bias and a20 index of 100%), indicating that the ANN is a reliable and efficient tool for predicting mechanical properties of luffa fiber–marble dust concrete within the tested M25 design space, and has potential utility in reducing the number of experimental trials needed during mix development. The ANN achieved high predictive accuracies (R² values of up to 0.96, low RMSE, bias and a20 index of 100%), indicating that the ANN is a reliable and efficient tool for predicting mechanical properties of luffa fiber–marble dust concrete within the tested M25 design space, and has potential utility in reducing the number of experimental trials needed during mix development. The limited size of the dataset may have affected the generalizing capability of the ANN method to a broader subject area. The limited size of the dataset may have affected the generalizing capability of the ANN method to a broader subject area. The present study demonstrates improved mechanical strength at the optimum mix proportion; formal life-cycle assessment (LCA), carbon footprint analysis, or multi-objective optimization were not performed and are not claimed. The mechanistic explanations offered for pore refinement, crack bridging, and interfacial bond improvement are consistent with the macroscopic experimental observations and with published literature on analogous systems; however, they remain inferential in the absence of dedicated microstructural characterization (SEM, MIP, XRD) at the mix level. Future studies shall incorporate such analyses to validate the proposed mechanisms. Only one concrete grade (M25) and particular concrete material sources were used, limiting the applicability of results relative to other grades and environmental conditions. Only one concrete grade (M25) and particular concrete material sources were used, limiting the applicability of results relative to other grades and environmental conditions. The long-term durability performance of the mix (e.g. settlement, shrinkage, creep, chemical resistance) was not adequately examined. Problems with fibre dispersion and workability due to not having enough fibres were not properly quantified. The long-term durability performance of the mix (e.g. settlement, shrinkage, creep, chemical resistance) was not adequately examined. Problems with fibre dispersion and workability due to not having enough fibres were not properly quantified. There needs to be more data to expand on the ANN method’s robustness by achieving a more varied mix of materials, curing structures, and environments to meet these goals. There needs to be more data to expand on the ANN method’s robustness by achieving a more varied mix of materials, curing structures, and environments to meet these goals. Author contributions S. Anandaraj: Conceptualization, Investigation, Writing – original draft. S. Shaniya: Investigation, Conceptualization, Formal analysis, writing-review and editing. S. Swaminathan: Validation, writing-review and editing. S. Suganya: Supervision, Writing – original draft. G. Sindhu Priya: Investigation, Validation, writing-review and editing. Balaji L: Formal analysis, writing-review and editing. Dhanalakshmi A: Validation, writing-review and editing. Dawit Tafesse Gebreyohannes: Methodology, project administration, writing-review and editing. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability All data is provided within the article. 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 S. Suganya, Email: suganya@src.sastra.edu Dawit Tafesse Gebreyohannes, Email: dawitt@hu.edu.et. References Associated Data Data Availability Statement All data is provided within the article.