This study investigates the effectiveness of various deep learning (DL) algorithms in predicting soil liquefaction susceptibility. We explore a spectrum of algorithms, including machine learning models such as Support Vector Machines (SVMs), K-Nearest Neighbors (KNN), and Logistic Regression (LR), alongside DL architectures like Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), Bidirectional LSTMs (BiLSTMs), and Gated Recurrent Units (GRUs). The performance of these algorithms is assessed using comprehensive metrics, including accuracy, precision, recall, F1-score, receiver operating characteristic (ROC) curve analysis, and area under the curve (AUC). Cross-entropy loss is employed as the loss function during model training to optimize the differentiation between liquefiable and non-liquefiable soil samples. Our findings reveal that the GRU model achieved the highest overall accuracy of 0.98, followed by the BiLSTM model with an accuracy of 0.95. Notably, the BiLSTM model excelled in precision for class 1, attaining a score of 0.96 on the test dataset. These results underscore the potential of both GRU and BiLSTM models in predicting soil liquefaction susceptibility, with the BiLSTM model’s simpler architecture proving particularly effective in certain metrics and datasets. The findings of this study could assist practitioners in seismic risk assessment by providing more accurate and reliable tools for evaluating soil liquefaction potential, thereby enhancing mitigation strategies and informing decision-making in earthquake-prone areas. This study contributes to developing robust tools for liquefaction hazard assessment, ultimately supporting improved seismic risk mitigation.