Projection Pursuit Regression-Based Model for Concrete Stress–Strain Behavior Under Axial Compression
Abstract
The uniaxial compressive stress–strain curve of concrete is critical for structural nonlinear simulation and safety evaluation. Traditional constitutive models rely on empirical assumptions with poor generalization, while common machine learning methods such as ANN and LSTM act as black-box tools lacking mechanical interpretability and cannot precisely capture key damage features under multi-factor coupling. This study aims to develop an interpretable prediction model for the full-range compressive behavior of concrete. A projection pursuit regression (PPR) model is constructed and trained using experimental data, with water–binder ratio, fly ash content and mineral powder content adopted as input variables. The influences of mix proportions on elastic modulus, peak stress, peak strain, inflection point D and critical shear damage point E are quantified. Point D and E represent rapid crack propagation and the formation of through-cracks, respectively. The PPR model yields stable and accurate predictions. Within the tested range, lower water–binder ratios and a higher mineral powder content improve mechanical performance, whereas increased fly ash content reduces peak stress and enhances ductility. Without predefined functions, the PPR method effectively avoids overfitting. This work provides a reliable data-driven approach for concrete proportion design and structural damage assessment in engineering practice.