In recent years, machine learning has quietly rewritten the underlying gameplay of concrete material development.
Random Forest and XGBoost automatically identify key factors that affect strength from dozens of matching data sets; SVR uses kernel functions to capture nonlinear interactions between components; MLP neural network provides end-to-end strength prediction directly from the mix proportion; Bayesian optimization no longer relies on manual trial and error, allowing the algorithm to search for the optimal hyperparameters on its own; SHAP interpretability analysis eliminates the “black box” of the model and clarifies “why this ratio is strong and that ratio is weak”.
At the forefront, physical information neural networks have begun to embed hydration reactions and strength development laws into the network, allowing for physical consistency even in small samples.
.