Machine learning approach for evaluating soil liquefaction probability based on reliability method

Reliability analysis is necessary to address the many uncertainties, including both model and parametric uncertainties. This study systematically assesses the reliability index (β) and probability of occurrence of liquefaction (PL) using the first-order reliability method (FORM) approach on the cone penetration test (CPT) dataset, taking into account parametric uncertainties. Acknowledging the recent advancements in machine learning models and their ability to capture complex, non-linear relationships and interactions within the data, a deep learning model, namely a deep neural network (DNN), is developed and suggested based on its performance in predicting PL. We use eight statistical performance metrics to evaluate the DNN model's performance across three distinct dataset split ratios. Additional charts, such as regression plots like Taylor's diagrams, rank analysis, regression error characteristics curves, and loss and epoch curves, is provided to comprehensively assess the DNN model's performance. The current investigation demonstrates that the DNN model is promising for predicting PL on CPT datasets. Additionally, we conduct a reliability-sensitivity analysis to determine the contribution of each variable in evaluating PL. According to the sensitivity analysis, the most important parameter is the equivalent clean sand penetration resistance (qc1Ncs). It is followed by the magnitude scaling factor (MSF) and the stress reduction factor (rd). This study contributes valuable risk assessments for geotechnical engineering design and advocates for the broader integration of FORM-based ML models in liquefaction evaluation.