Machine learning-driven reverse prediction of concrete mix design: a comparative study of twelve algorithms

Concrete mix design is usually treated as a forward, iterative task: proportions are adjusted until the required compressive strength is met. Here, we work in the reverse direction. From a target 28-day compressive strength and curing age, the model estimates seven proportions directly: cement, blast furnace slag, fly ash, water, superplasticizer, coarse aggregate, and fine aggregate. The workflow combines machine learning with a deterministic physics-informed correction layer. Twelve algorithms from six model families (linear, instance-based, tree-based, ensemble, kernel, and deep learning) were trained on the UCI Concrete Compressive Strength dataset 1 ( n = 1,030). For a fair comparison, each model used the same multi-output regression wrapper, hyperparameters were tuned with 30-iteration RandomizedSearchCV and 5-fold internal cross-validation, robustness was assessed by 10-fold cross-validation, and pairwise differences were tested with paired t-tests and Wilcoxon signed-rank tests. The study contributes a tuned twelve-algorithm benchmark with statistical comparisons, a per-component analysis linked to concrete-material mechanisms, and a post-processing layer that enforces 0.30 ≤ W/C ≤ 0.80, cement ≥ 200 kg/m³, supplementary cementitious replacement ≤ 50% of binder, and non-negative superplasticizer dosage. After tuning, Extra Trees gave the highest test-set R² (0.341), while Gradient Boosting gave the highest 10-fold CV R² (0.262 ± 0.041). The leading four models were not significantly different (paired t-test, p > 0.05). Per-component R² ranged from about 0.52 for superplasticizer and 0.46 for cement to below 0.16 for fine aggregate and slag. This pattern is consistent with concrete chemistry: hydration-controlled components can be partly inferred from strength, whereas aggregate proportions are set largely by grading and workability. For a 15 MPa target, the unconstrained model returned a W/C of about 1.0; the physics layer corrected it to 0.80 with only a modest test-set R² reduction (0.341 to 0.302). The full pipeline was then applied to an independent ultra-high-performance concrete dataset (863 mixtures, 17 variables). The algorithm ranking was preserved and the best inverse R² increased to 0.608, indicating that input dimensionality strongly affects inverse predictability. A conditional-variance analysis shows the tuned models operate near the upper bound of the two-input formulation: mixtures with the same strength form several distinct design families, so the prediction is best read as a screening anchor within a family of valid designs. The proposed framework should be used as a screening tool, not as a replacement for ACI 211.1 design or laboratory verification.