Abstract
Abstract Soil liquefaction refers to the loss of strength and bearing capacity of soil under dynamic loads. Structures built on liquefied soils experience settlement. In this study, the newly developed Kolmogorov–Arnold Networks (KAN) method, an innovative artificial neural network technique, was utilized to predict liquefaction-induced settlement. The KAN method, which extends beyond the traditional Multi-Layer Perceptron (MLP) approach, was compared with the Random Forest method for benchmarking purposes. Using data derived from laboratory and field studies, models were constructed using both methods, and their performances were analyzed. According to the results, the KAN model outperformed the RF model in terms of R2, MAE, MSE, and RMSE metrics. Additionally, the feature importance analysis on the KAN model identified cyclic stress ratio (csr) and corrected SPT blow count (n1_60) as the most significant variables. These results underscore the potential of the KAN method in enhancing predictive accuracy and reliability in geotechnical applications, paving the way for its broader acceptance and implementation in real-world scenarios.