Synthetic Datasets and Machine Learning Algorithms for Enhanced Calibration of Plasticity Models for Liquefiable Geosystems

Machine Learning (ML) algorithms are increasingly adopted in engineering to extract insights from data and support decision-making. While they offer opportunities to enhance analysis accuracy and efficiency, their application is often hindered by challenges such as limited data availability, the need for curated datasets, and appropriate algorithm selection. One promising area for ML application is the calibration of constitutive models in nonlinear dynamic analyses (NDAs). Given the wide variability in soil properties encountered in practice, a well-trained ML model can significantly improve the robustness and efficiency of the calibration process. This study explores two strategies for dataset development and ML training aimed at calibrating the PM4Sand constitutive model, with one approach incorporating regional sand property data. Statistical analyses, including entropy metrics, are used to evaluate the influence of input features, and the models are validated using element-level simulations in FLAC. Findings demonstrate that the proposed methods improve calibration reliability and contribute to more accurate and efficient NDAs.