Background: Geotechnical engineering relies on inherently variable, nonlinear subsurface data that challenges conventional analytical methods. Machine learning (ML) offers a data-driven pathway to more accurate and efficient prediction of soil behaviour and infrastructure performance.
Objective: To systematically evaluate ML algorithms for core geotechnical tasks — soil classification, slope stability, foundation bearing capacity, and settlement — and to identify optimal model selection strategies for practical deployment.
Methods: A comparative methodology reviewed 94 peer-reviewed studies (2014–2024). Six ML algorithms were benchmarked using prediction accuracy, RMSE, F1 score, and computational efficiency across standardised geotechnical datasets.
Results: Random Forest and LSTM networks achieved the highest mean accuracy (93.2% and 94.1%, respectively). The integrated ML workflow reduced analysis time by up to 62% compared to conventional numerical methods, with an average cross-validation R² of 0.924.
Conclusion: ML techniques substantially improve the precision and efficiency of geotechnical analysis. Future work should focus on uncertainty quantification, transfer learning for sparse datasets, and integration with Building Information Modelling (BIM) platforms.