A Novel Genetic Optimized LSTM Algorithm for Stress-Strain Modeling of High Performance Cementitious Materials in Rock and Geotechnical Applications

Predicting the nonlinear stress-strain response of high-performance cementitious materials is essential for the safe and efficient design of underground, geotechnical, and structural systems. Owing to the het erogeneous nature and complex damage evolution of these materials, conventional constitutive models of ten fail to capture their full mechanical behavior under varying loading conditions. This study introduces a Genetic Algorithm Optimized Long Short-Term Memory (GA-LSTM) framework that combines evo lutionary optimization and deep learning to model the complete stress-strain curve of high-performance concrete. Experimental data from uniaxial compression tests were systematically preprocessed using cleaning, normalization, interpolation, and Savitzky-Golay filtering to preserve physical integrity. The genetic algorithm optimized key hyperparameters including network depth, hidden units, dropout rate, and sequence length under a multi-objective, physics-informed strategy. The proposed model achieved superior predictive accuracy (R2 ≈ 0.995, RMSE < 3 MPa) with minimal deviation in peak stress and post-peak toughness. Comparative statistical analyses (ANOVA, Tukey’s HSD, Pearson correlation) ver ified the model’s reliability and robustness. The developed GA-LSTM framework offers a scalable and interpretable tool for mechanical characterization, design optimization, and geomechanical simulation of rock-concrete composite systems in civil, mining, and geotechnical engineering.