A Hybrid FCM–EANN–XGBoost Framework for Predictive Modeling of Arching in Zoned Earth Dams Using Field and FEM Data

Arching in zoned earth dams is a complex stress-transfer phenomenon resulting from the differential compressibility between the central clay core and the adjacent granular shells. This study presents a monitoring-driven hybrid framework that integrates field instrumentation data, advanced nonlinear finite element method (FEM) simulations, and artificial intelligence (AI) techniques to develop a robust predictive model for arching behavior. The Eyvashan earth dam in Iran was used as a comprehensive validation case, supported by more than 850 quality-controlled records from 29 total pressure cells and 75 piezometers. Two-dimensional FEM analyses considered staged construction and hydro-mechanical processes during reservoir filling. The proposed framework incorporates FEM-derived engineered features within a multi-stage AI pipeline, including Fuzzy C-Means (FCM) clustering to identify behavioral regimes, an Evolutionary Artificial Neural Network (EANN) for nonlinear feature extraction, and XGBoost for final arching coefficient prediction. The hybrid FCM–EANN–XGBoost model achieved high accuracy with \(\:{R}^{2}=0.974\), \(\:NRMSE=0.074\), and \(\:MAPE=5.6\text{\%}\), representing more than 40% improvement over standalone FEM models and conventional machine learning approaches. SHAP analysis confirmed the dominant role of FEM-derived features in predictions, demonstrating the effectiveness of physics-informed learning. The framework successfully captures the spatiotemporal evolution of arching, including its asymmetry between the upstream and downstream faces of the core and its progressive intensification during reservoir filling, providing a foundation for proactive dam safety management.