Artificial Intelligence and Deep Learning Models for Bearing Capacity Prediction of Foundation Systems: A State-of-the-Art Review

The evaluation of the ultimate bearing capacity (UBC) of foundation systems remains a fundamental challenge in geotechnical engineering because of the complex interactions among soil properties, foundation geometry, loading conditions, embedment depth, and soil–foundation behavior. In recent years, artificial intelligence (AI) and deep learning (DL) techniques have emerged as powerful data-driven tools for modeling nonlinear geotechnical systems and improving bearing-capacity prediction. This study presents a comprehensive state-of-the-art review of AI- and DL-based approaches for foundation systems, including shallow foundations, deep foundations, pile foundations, and other geotechnical applications. Major models, including Artificial Neural Networks (ANNs), Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, Transformer models, Graph Neural Networks (GNNs), hybrid AI frameworks, and physics-informed deep learning approaches, are critically reviewed and compared. Particular attention is given to the integration of AI models with numerical methods, including the finite element method (FEM) and finite element limit analysis (FELA). The reviewed studies frequently report lower prediction errors than conventional empirical, numerical, and machine-learning approaches within the evaluated datasets. However, many of the highest reported accuracies are based on laboratory-scale experiments, simulation-generated data, or random train–test partitions of a single database. Consequently, these results may demonstrate effective interpolation within controlled data distributions rather than reliable performance under independent field conditions. Model performance is strongly influenced by dataset origin and diversity, feature selection, validation strategy, overfitting control, and generalization capability. Hybrid datasets combining field, laboratory, and numerical data offer a promising route toward more reliable prediction, but genuine external validation using independent sites, projects, or institutions remains uncommon. Moreover, architectural suitability should reflect the physical structure of the problem: CNNs are appropriate for spatial heterogeneity, LSTMs for time-dependent behavior, Transformers for long-range interactions, and GNNs for mechanically connected systems. Limited field-scale datasets, weak external validation, limited model interpretability, inadequate uncertainty quantification, and persistent data scarcity continue to restrict widespread engineering implementation. Future research should prioritize explainable AI, physics-informed learning, transfer learning, hybrid data frameworks, open benchmark datasets, and multi-site field validation to improve the robustness, transparency, and practical applicability of intelligent bearing-capacity prediction for diverse foundation systems.