Prediction and Sensitivity Analysis of Foam Concrete Compressive Strength Based on Machine Learning Techniques with Hyperparameter Optimization

Foam concrete is widely used in engineering due to its lightweight and high porosity. Its compressive strength, a key performance indicator, is influenced by multiple factors, showing nonlinear variation. As compressive strength tests for foam concrete take a long time, a fast and accurate prediction method is needed. In recent years, machine learning has become a powerful tool for predicting the compressive strength of cement-based materials. However, existing studies often use a limited number of input parameters, and the prediction accuracy of machine learning models under the influence of multiple parameters and nonlinearity remains unclear. This study selects foam concrete density, water-to-cement ratio (W/C), supplementary cementitious material replacement rate (SCM), fine aggregate to binder ratio (FA/Binder), superplasticizer content (SP), and age of the concrete (Age) as input parameters, with compressive strength as the output. Five different machine learning models were compared, and sensitivity analysis, based on Shapley Additive Explanations (SHAP), was used to assess the contribution of each input parameter. The results show that Gaussian Process Regression (GPR) outperforms the other models, with R², RMSE, MAE, and MAPE values of 0.95, 1.6, 0.81, and 0.2, respectively. It is because GPR, optimized through Bayesian methods, better fits complex nonlinear relationships, especially considering a large number of input parameters. Sensitivity analysis indicates that the influence of input parameters on compressive strength decreases in the following order: foam concrete density, W/C, Age, FA/Binder, SP, and SCM.