AI Learns to Predict Six Concrete Properties at Once from Plastic and Rice Husk Waste Blends

Concrete made partly from shredded plastic bottles and burnt rice husks could become far easier to design thanks to a new machine-learning framework that predicts six key performance properties of the material simultaneously — and explains which ingredients matter most. The study, published in the journal Discover Concrete and Cement, combines a Kolmogorov-Arnold Network (KAN) with an optimisation algorithm called the Marine Predators Algorithm (MPA) to estimate compressive strength, split tensile strength, flexural strength, water absorption, modulus of elasticity and rapid chloride permeability in a single model.

The motivation is environmental as much as computational. Cement production is a major industrial source of greenhouse gases, because clinker manufacture involves both limestone calcination and heavy thermal energy consumption. Recycled polyethylene terephthalate (PET) plastic offers a route for diverting persistent post-consumer waste from landfill while reducing demand for virgin aggregates, and rice husk ash (RHA), a silica-rich agricultural by-product, can partially replace cement and take part in pozzolanic reactions that refine the concrete matrix. Blending the two waste streams lets PET substitute for aggregate while RHA substitutes for binder, but the resulting material behaviour is notoriously difficult to anticipate because the effects of binder content, plastic dosage, ash replacement, aggregate proportions and water are strongly coupled — and not always in simple, monotonic ways.

The experimental foundation came from the authors’ own laboratory programme. Ten concrete mixtures, labelled M0 through M9, were produced with a constant water-to-binder ratio of 0.40. PET was introduced as a partial fine-aggregate replacement at levels of 0, 20, 30 and 40 kilograms per cubic metre, while RHA replaced part of the cement at 0, 60, 75 and 90 kilograms per cubic metre. A control mixture contained neither waste material. All six performance indicators were measured after 90 days of curing, an extended period chosen to allow the pozzolanic reaction of the ash to develop fully. Compressive strength followed IS 516:1959, split tensile strength ASTM C496, flexural strength ASTM C78, water absorption ASTM C642, elastic modulus ASTM C469 and chloride permeability ASTM C1202.

What sets the new framework apart from most prior concrete machine-learning studies is its multi-target design and its built-in transparency. Instead of training separate models for each property, the KAN predicts all six outputs at once through a shared representation. Unlike conventional neural networks that rely on fixed activation functions such as ReLU, a KAN assigns learnable B-spline functions to the connections, or edges, between nodes. These spline functions can be visualised after training, giving engineers a direct view of how individual mix-design variables influence the predicted response — an intrinsic interpretability that post-hoc explanation tools cannot replicate.

The Marine Predators Algorithm handled the hyperparameter tuning that would otherwise be left to manual trial and error. Inspired by predator-prey foraging behaviour, the population-based optimiser searched for the spline order, grid resolution, regularisation coefficient and learning rate that minimised validation error, using 50 candidate solutions over a maximum of 150 iterations. The optimised configuration settled on a spline order of 5 and a grid size of 8, a compromise the authors describe as balancing expressive capacity against overfitting risk. Because the search space was defined in advance and the objective function was the validation root-mean-square error, the process reduced dependence on subjective parameter choices.

Data scarcity posed a serious challenge: laboratory testing of six properties across a factorial design is expensive, and the entire database consisted of just ten physically produced mixtures. To enlarge the effective training set, the researchers applied MixUp augmentation, a technique that generates synthetic observations by convex interpolation between pairs of existing samples. Crucially, the data were partitioned into training, validation and testing subsets before any augmentation, and only the training subset was enriched — with an interpolation coefficient sampled from a Beta distribution with the shape parameter set to 0.4, favouring synthetic points close to real parent mixtures. Using an eightfold augmentation multiplier, eighty synthetic samples were generated, producing a training set of ninety while the validation and testing observations remained untouched originals. All synthetic points were constrained to remain within the convex hull of the experimentally observed design space, so no generated mixture exceeded the practical replacement limits actually investigated.

The reported performance figures are striking. Across all six targets, the KAN-MPA framework achieved an overall coefficient of determination of 0.9995, with a mean absolute error of 0.0005, a root-mean-square error of 0.0016, a mean absolute percentage error of 0.004 percent, an explained variance score of 0.9948, an RMSE-to-standard-deviation ratio of 0.000144 and a composite performance index of 0.9974. Regression correlations between predicted and measured values reached 0.99993 on the training data, 0.99977 on validation, 0.99981 on testing and 0.99989 overall. Residual analysis showed errors concentrated around zero with no systematic drift, funnel-shaped dispersion or evidence of heteroscedasticity, and repeated validation runs with confidence interval assessment indicated the metrics did not depend on one favourable data split.

The explainability layer added practical insight. Integrated Gradients, an attribution method that integrates model gradients along a path from a baseline input to each observation, quantified how much each mix-design variable contributed to every predicted property. Cement content, PET content and RHA replacement level consistently emerged as the most influential variables. For compressive strength, cement and PET dominated, reflecting binder availability and aggregate-substitution effects; for water absorption and chloride permeability, the influence of RHA grew markedly, consistent with its role in pozzolanic pore refinement. The chloride-permeability attribution was particularly sensitive to PET content, which the authors note is physically plausible given PET’s known effects on the interfacial transition zone. They are careful to stress, however, that Integrated Gradients quantifies model sensitivity rather than direct causality, so the attribution patterns indicate relative influence within the trained model rather than definitive experimental proof of governing mechanisms.

The analysis also flagged a promising design window. Mixtures containing roughly 20 to 30 kilograms per cubic metre of PET and 60 to 75 kilograms per cubic metre of RHA showed a favourable balance of mechanical and durability performance, including improved compressive strength, reduced water absorption and enhanced chloride resistance. Within that region, the authors argue, appropriately selected PET-RHA combinations can support structural performance and sustainability goals simultaneously, cutting clinker demand, diverting plastic and agricultural waste, and potentially extending service life through better durability. The learned spline functions themselves identified steep-sensitivity regions where small dosage changes produce large property shifts, and flatter regions suitable for robust mix optimisation.

The authors are unusually candid about the limits of their results. The database comprised only ten mixtures from a single experimental programme, so the findings should be interpreted strictly within the investigated design space, and broader generalisation requires larger, multi-source datasets and independent external validation. No formal statistical significance testing was performed, so the performance comparisons against the benchmark models — Support Vector Regression, Random Forest and Multilayer Perceptron — are descriptive rather than statistically verified superiority, and widely used algorithms such as XGBoost, LightGBM, CatBoost and Extra Trees were not evaluated. Uncertainty quantification beyond confidence interval analysis, including bootstrap resampling and prediction intervals, was beyond the study’s scope. The team also acknowledges that simpler statistical models might perform competitively on such a small dataset, and frames the work as an initial demonstration of KAN-based multi-target modelling rather than evidence of superiority over all alternatives. Future directions include transfer learning, physics-informed constraints, life-cycle assessment integration, and exploiting the symbolic regression capabilities of KANs to derive closed-form predictive equations that could carry explainable waste-concrete design from the laboratory into engineering specification.

Subject of Research: Explainable multi-target machine learning prediction of mechanical and durability properties of concrete containing recycled PET plastic and rice husk ash

Article Title: Kolmogorov-Arnold network integrated with marine predators algorithm for explainable multi-target prediction of mechanical and durability performance of PET- and RHA-blended sustainable concrete

Article References: Kumar, P. P., Reddy, B. R., Premkumar, N., Reddy, A. R., Balanagaiah, K., & Durga, M. (2026). Kolmogorov-Arnold network integrated with marine predators algorithm for explainable multi-target prediction of mechanical and durability performance of PET- and RHA-blended sustainable concrete. Discover Concrete and Cement, 2(1), Article 74. https://doi.org/10.1007/s44416-026-00114-z

Image Credits: AI Generated

DOI: 10.1007/s44416-026-00114-z

Keywords: concrete, machine learning, Kolmogorov-Arnold network, recycled PET, rice husk ash, sustainable materials, compressive strength, durability, Integrated Gradients, data augmentation, cement replacement, Marine Predators Algorithm