Early warning system for risk assessment in geotechnical engineering using Kolmogorov-Arnold networks

In this study, we used the Kolmogorov-Arnold networks (KAN) model based on the Kolmogorov-Arnold representation theorem for a comprehensive and fair evaluation. We compare its performance with four other powerful classification models across three datasets: a simple slope binary classification dataset, an imbalanced rockburst dataset, and a highly discrete liquefaction dataset. First, a thorough review of machine-learning algorithms for geohazard assessment was conducted. Subsequently, three datasets were collected from real engineering practices, and their data structures were visualized. Bayesian optimization was then used to adjust the parameters of all models across all datasets. To ensure model interpretability, a global sensitivity analysis based on Sobol indices was performed, establishing an interpretable visual analysis of the model’s decision-making process. For a fair evaluation, various metrics and repeated stratified 10-fold cross-validation were employed to comprehensively analyze the predictive results of the models. The results indicate that although the KAN model, based on the RBF kernel, achieves the expected performance on the binary classification dataset, it also performs well on imbalanced and highly discrete datasets, significantly surpassing other commonly used classification models. This demonstrated the broad application potential of the KAN model in geotechnical engineering.