Real-time interpretable and cluster-stratified lightGBM framework for high-precision concrete strength prediction and instantaneous mixture optimization

This study presents a real-time, interpretable framework based on the light gradient boosting machine (LightGBM) algorithm for the accurate prediction and optimization of 28-day concrete compressive strength (Fc), validated using a dataset of 500 concrete mixtures. The proposed model was benchmarked against seven widely used regression algorithms, including linear regression (LR), ridge regression (RR), random forest (RF), K-nearest neighbors (KNN), support vector regression (SVR), decision tree (DT), and multivariate adaptive regression splines (MARS), to ensure a comprehensive comparative evaluation. The LightGBM model demonstrated superior predictive performance relative to the benchmark models, achieving an RMSE of 6.11 MPa and an R² of 0.951 during the initial evaluation. Model robustness and generalization capability were further verified using a 10 × 10 repeated k-fold cross-validation procedure, yielding stable results (R² = 0.940 ± 0.017; RMSE = 6.37 ± 0.49 MPa). To capture heterogeneity in mixture compositions, K-means clustering was applied to partition the dataset into four distinct mixture regimes, within which stratified LightGBM models further improved predictive accuracy, reducing RMSE to 3.7–5.1 MPa and achieving R² values exceeding 0.97. Model interpretability was enhanced through global and regime-specific SHAP (Shapley Additive Explanations) analyses, which provided transparent and physically consistent insights into feature contributions, consistently identifying cement as the dominant positive factor and water as the primary negative driver of CS. Furthermore, an interactive web-based prediction engine was developed to enable instantaneous strength prediction, real-time sensitivity analysis, 95% prediction interval estimation, and specification-driven mixture optimization with millisecond-level computational efficiency. Comprehensive diagnostic evaluations, including Taylor diagrams, residual control charts, calibration plots, and prediction-interval validation, confirmed the statistical reliability and practical applicability of the proposed framework. Overall, the developed LightGBM-based system provides an accurate, interpretable, and scalable decision-support tool for data-driven concrete mix design and performance optimization.