A Novel Approach to Predicting Liquefaction-Induced Settlements Using Kolmogorov-Arnold Networks (KANs)

This study investigates the applicability and effectiveness of Kolmogorov-Arnold Networks (KAN) in predicting settlements due to soil liquefaction, a critical issue in geotechnical engineering. Soil liquefaction, resulting from increased pore water pressure, diminishes soil bearing capacity and can lead to significant structural damage. Utilizing a comprehensive dataset derived from field and laboratory studies, the data was divided into training (70%), validation (15%), and testing (15%) sets and processed as torch tensors for the KAN model. The model, consisting of three layers with grid and k parameters set to 3 and 11, respectively, was trained using the LBFGS optimizer and MSE Loss function over 125 steps. The KAN model demonstrated superior performance with an R² value of 0.935 and an MAE of 0.14 on the training set, and an R² of 0.908 and an MAE of 0.176 on the test set. Comparative analysis with other studies showed that KAN outperformed traditional neural network models. Feature importance analysis revealed “unit_weight” as the most significant feature, aligning with previous studies. These results underscore the potential of KAN in enhancing predictive accuracy and reliability in geotechnical applications, paving the way for its broader acceptance and implementation in real-world scenarios.