Prediction of liquefaction induced settlement using SPT dataset through hybrid CNN-BiLSTM-AM model

Soil liquefaction-induced settlement poses a major risk to infrastructure in earthquake-prone regions. This study introduces a hybrid deep learning model, CNN-BiLSTM-AM, that combines convolutional and bidirectional long short-term memory networks with an attention mechanism to improve the prediction of liquefaction-induced settlement using Standard Penetration Test (SPT) data. The model uses key input features such as depth (m), unit weight (kN/m3), corrected SPT-N (N1(60)) values, and cyclic stress ratio (CSR). These parameters reflect critical soil properties and seismic loading conditions. Actual vs. predicted graphs, and performance metrics including R2, MAE, RMSE, and MSE were utilized to evaluate the proposed model. Comparative analysis confirms the robustness of the CNN-BiLSTM-AM model showed the highest accuracy of 94.69%. Sensitivity analysis confirmed that N1(60) as the most crucial input feature, aligning with geotechnical understanding of soil resistance to seismic deformation. The model not only demonstrates high predictive accuracy but also reflects practical engineering relationships, offering a valuable tool for seismic design and risk mitigation. This work lays the groundwork for further investigation by emphasizing the potential and difficulties of utilizing SPT dataset to forecast soil liquefaction-induced settlement.