Liquefaction susceptibility using machine learning based on SPT data

Assessing the potential for liquefaction using traditional experimental or empirical analysis procedures is both time-consuming and arduous. Employing a machine learning model that can accurately predict liquefaction potential for a specific site can reduce the time, effort, and associated costs. This study proposes several empirical machine learning models, including deep neural network (DNN), convolutional neural network (CNN), recurrent neural network (RNN), long short-term memory (LSTM), and bi-directional long short-term memory (BILSTM), to assess the liquefaction potential of soil deposits based on SPT-based post liquefaction datasets. To train the proposed models, a dataset comprising 834 liquefied and non-liquefied cases was collected to perform the liquefaction analysis. A Pearson correlation matrix was also conducted to examine the correlation between soil and seismic parameters and the probability of liquefaction. Furthermore, a sensitivity analysis was adopted to assess the impact of soil and seismic parameters on the probability of liquefaction. The proposed model's prediction capability was assessed using several performance indices, including rank analysis, accuracy matrix, and AIC criteria. The comparative analysis of the proposed models' predictive ability to determine liquefaction probability revealed that the RNN model outperformed the others, displaying the highest accuracy and lowest error index values. Subsequently, the RNN model achieved the first rank with a total score value of 70, followed by the CNN (55), DNN (52), BILSTM (47), and LSTM (16) models. The parametric analysis, rank analysis, accuracy matrix, and AIC criteria collectively demonstrate the proposed models' ability to predict liquefaction probability. Furthermore, the robustness of these models was assessed through external validation and comparative analysis