Parameters estimation for van Genuchten-soil freezing characteristic curve model by using synthetic data

• Introduced SMOTE and CTGAN to synthesize data to improve VG-SFCC model performance. • SMOTE-synthesized data significantly improves VG-SFCC model parameter prediction. • CTGAN-synthesized data introduces noise, failing to enhance model performance. • Data-driven framework enables efficient SFCC estimation in frozen soil engineering. Accurate estimation of the Soil Freezing Characteristic Curve (SFCC) is essential for understanding the thermal-hydro-mechanical behavior of frozen soils. The van Genuchten (VG) model has been widely adopted for characterizing the SFCC due to its flexibility and ability to fit various soil types. However, determining the parameters of the VG-SFCC model is often hindered by limited data and high experimental costs. The present study proposed a novel methodology to estimate the VG-SFCC model parameters based on synthetic datasets generated using the Synthetic Minority Oversampling Technique (SMOTE) and Conditional Tabular Generative Adversarial Network (CTGAN). A Multi-Output Neural Network (MONN) was constructed using soil physical properties as inputs and VG-SFCC model parameters as outputs. Results demonstrate that the limited size of the collected dataset did not yield satisfactory model performance. However, SMOTE-synthesized data significantly enhanced the model’s predictive performance, yielding reliable VG-SFCC parameter estimations for various soil textures. Critically, this study reveals that interpolation-based methods (SMOTE) preserve physical relationships better than generation-based methods (CTGAN) for VG-SFCC model parameters estimation. Discussions on the impact of synthetic sample size on model performance, as well as the differences between the two data generation methods are also present in detail. This study provides practical guidance for ML-based SFCC modeling under data scarcity, offering a validated framework that can amplify the value of limited experimental data in frozen soil engineering practice.