Prediction of Concrete Compressive Strength for Ready-Mix Concrete: A Comparative Evaluation of Algorithms with Explainability Analysis

Abstract

Ready-mix concrete is accepted or rejected on cube strengths that arrive weeks after placement, so an estimate available at the batching plant would be worth having. This paper benchmarks twelve machine learning algorithms for that purpose on the public concrete compressive strength database, under one preprocessing and evaluation protocol. Twenty-five exact duplicates were removed before any split was drawn, leaving 1005 records divided 804 for training and 201 for testing. A stacked ensemble of XGBoost, LightGBM, CatBoost and extremely randomized trees was the most accurate model, reaching R² = 0.942, RMSE = 4.15 MPa and MAE = 2.65 MPa, and leading in 15 of 20 repeated random splits. LightGBM and XGBoost followed at 0.938 and CatBoost at 0.932, while the multilayer perceptron and support vector regression reached about 0.87 and the linear baselines only 0.580. The more useful finding concerns how such results should be read. Tested with the corrected resampled t-test over the repeated splits, the stacked ensemble cannot be separated from LightGBM (p = 0.207), although an ordinary paired test on a single split declares a comparable difference highly significant. Permutation importance, partial dependence and SHapley Additive exPlanations all rank age, cement and water first, with mean absolute SHAP values of 8.25, 6.93 and 3.74 MPa, and adding five domain-driven ratio features produced no detectable gain. With an RMSE comparable to reported plant standard deviations of 1.92 to 6.57 MPa, a model of this accuracy suits pre-dispatch screening within the IS 10262 and IS 4926 framework rather than acceptance testing.