A machine-learning study has combined prediction and optimisation to explore trade-offs in eco-concrete mixtures. Using 400 experimental records, the researchers trained models to estimate compressive strength, then searched for designs balancing strength against cement content and water-to-binder ratio. The work appeared in Scientific Reports on 23 October.[1]
The approach is intended to narrow a materials-testing problem. Engineers can vary several ingredients and curing conditions, but physically testing every combination is expensive. A model can rank candidate designs and expose trade-offs, helping determine which mixtures deserve the next laboratory experiment.
Predicting strength before searching for a mix
Seven inputs described mixture and curing characteristics. The best tested model, Bayesian-optimised XGBoost, achieved a held-out R² of 0.958, closely followed by random forests at 0.957 and support-vector regression at 0.950. The small separation between the first two is more informative than treating the winning model as categorically superior.[1]
R² describes how well predictions account for variation in measured strength under the study’s evaluation. It is not the probability that a proposed mixture will meet an engineering requirement. The result establishes performance on the dataset split, while a newly selected design requires a physical measurement.
The team also added SHAP, an interpretability method that estimates the contributions of input variables to a prediction. Curing age and porosity emerged as important factors, followed by cement dosage and the water-to-binder ratio. This is useful because a purely opaque prediction of strength would be of limited help to a materials researcher trying to understand where a formulation is sensitive.[1]
Choosing among competing objectives
Once the predictor was trained, the researchers embedded it in a multi-objective search algorithm called NSGA-II. The algorithm evaluated compromises among higher predicted 28-day strength, reduced cement content and a lower water-to-binder ratio. In optimisation language these compromises form a Pareto frontier: improving one target may worsen another, and there is no single best answer without defining the engineering preference.[1]
One selected compromise had a water-to-binder ratio of 0.250 and cement content of 351.47 kilograms per cubic metre, with predicted 28-day compressive strength of 10.87 megapascals. That is a prediction attached to a particular objective choice. Its suitability depends on the intended application and its requirements; the number alone does not make the design either a successful building material or a failed one.[1]
This distinction matters for optimisation. A search algorithm can identify a combination that performs well against its stated targets while leaving other important properties outside the calculation. Strength at a particular age is one measurement; durability, workability and the demands of a specific use may require additional constraints or experiments.
Lower cement content is also an environmental proxy rather than a complete carbon calculation. A meaningful comparison would examine the materials, energy, transport, curing and expected service life of the candidate and its alternative. The reported objective values do not by themselves quantify whole-life emissions savings.[1]
The next experiment follows from the model
A high R-squared on a split of 400 records says the model captured patterns in that set. Whether it generalises to a new quarry's aggregates, different weather, cement batches or unusual admixtures requires independent datasets and prospective mixes. Concrete is especially sensitive to testing procedures and the material's age. Small numerical forecast errors on the familiar data range can hide larger mistakes in formulations the model has never encountered.[1]
SHAP feature contributions can help researchers identify which model inputs deserve attention, but they explain the fitted predictor rather than certify a physical mechanism. If age and porosity strongly influence a prediction, targeted experiments can test that sensitivity and reveal whether it remains valid for a new mix family.[1]
A useful next step would cast and cure designs from several points on the proposed trade-off frontier, then compare measured strength and application-specific properties with the forecasts. That would test both the predictor and the choice of objectives. The study offers a structured way to choose those experiments, with the material’s performance still determined by the specimens.[1]
Sources & further reading
- Intelligent prediction of eco-concrete compressive strength based on SHAP-interpretable BO-XGBoost and NSGA-II multi-objective optimization Scientific Reports · 23 Oct 2026 · Accessed 26 Oct 2026