Fast and interpretable prediction of seismic kinematics of flexible retaining walls in sand through explainable artificial intelligence

Assessing the response of earth-retaining walls to seismic loading is key for safety. Nonlinear dynamic analyses using the finite element method (FEM) calculate the permanent displacements (PDs) the soil–wall system accumulates during an earthquake, but are computationally intensive. Simplified dynamic displacement-based methods that obtain PDs based on a failure mechanism (FM) and critical acceleration are faster. Still, they require geotechnical expertise, e.g., in strength, failure criteria, and stability under pseudo-static conditions. This paper presents a new method using eXplainable Artificial Intelligence (XAI) to predict the FM and kc (critical acceleration as a fraction of gravity) of flexible retaining walls in dry, coarse-grained soil, based on geometry and soil/wall properties. We created a large dataset (∼1000 samples) via FEM simulations based on plasticity limit theorems, with 7 types of FMs affecting 3 types of walls: propped, cantilever, and anchored. Each sample contained geometrical parameters, soil/wall properties, and FEM-derived FM and kc. Trained on these data, our system has a weighted F1-score of 97.28% and an R2 above 0.987 for FM and kc predictions, respectively. Shapley explanations (SHAP) and a new coherence model help experts interpret the results and check explanation consistency. Its high accuracy and millisecond-level inference make our method ideal when assessing the seismic response of numerous structures within short timeframes. Also, its interpretability promotes AI-expert collaboration, enhancing decision-making.