A Generator of Synthetic Liquefaction Safety Factor Profiles for Training Machine Learning Models

• Introducing an innovative method to generate large synthetic soil profiles. • Addressing uncertainties in defining Ishihara’s critical layers H and H. • Validating synthetic data through a deep learning model. • Potential of combining synthetic data and machine learning in Geotechnical Engineering. This paper presents an innovative approach to soil liquefaction assessment through the development of a data generator for training machine learning (ML) models. The research focuses on the procedure proposed by Ishihara in 1985 discussing the uncertainties in identifying critical soil layers for liquefaction, specifically of the non-liquefiable crust and the potential liquefiable layer, which are essential parameters for predicting liquefaction susceptibility. The current manual identification process presents two significant challenges. It is time-consuming, when analyzing large cone penetration test (CPT) datasets such as microzonation studies, and identification varies considerably based on expert judgment and site-specific characteristics. Moreover, the Ishihara’s work not only lacks clear guidelines for parameter identification but was also developed using a small dataset, which may not represent the variety of geological conditions worldwide. To address these constraints, a methodology is proposed for generating synthetic liquefaction Factor of Safety (FS) profiles to train machine learning models for automated H 1 and H 2 identification. This synthetic data generation approach overcomes the scarcity and high cost of real CPT data by providing sufficient training examples for robust machine learning models, trained and tested to simulate expert judgment. The approach is designed to be adaptable to both professional subjectivity and site-specific conditions. The validity of this approach is demonstrated through the training and testing of a deep learning model, achieving coefficient of determination ( R 2 ) values of 98.77% for H 1 and 82.83% for H 2 when tested against real-world CPT data from 41 sites. This approach not only significantly reduces the time required for liquefaction assessment in large-scale studies but also represents a transformative and region-adaptable tool in geotechnical engineering practice that can be calibrated to diverse geological contexts through systematic recalibration of the synthetic data generator.