A data-driven approach for soil parameter determination using supervised machine learning

Soil constitutive models have significantly advanced over the years, often with an increase in parameters. Accurate determination of these parameters is critical, as inaccuracies can lead to unreliable numerical simulations. Conventional calibration practicetypically relies on laboratory testing, which is often impractical for applications, particularly at early project stages. An ongoing research projectofthe Computational Geotechnics Group at Graz University of Technology focuses on developing an Automated Parameter Determination frameworkthat integrates a graph-based approach to derive constitutive model parameters from in-situ tests. This framework uses established correlations to identify parameters based on in-situ tests. However, the multiplicity of correlations for a given parameter introduces inherent challenge in selecting the recommended value. An alternative methodinvolves using advanced regression algorithms to enhance the robustness of parameter determination through data-driven techniques.This approach can improve the quality of numerical simulations and minimizes the uncertainty in the parameter calibration process. In this paper, supervised machine learning regression models are employed to predict soil parameters, including saturated unit weight, undrained shear strength, and small-strain shear modulus (via shear wave velocity) using cone penetration test data as input. The performance of these models is benchmarked using data from two Norwegian GeoTest Site locations and are validated based on laboratory-derivedvalues, in-situ measurements, and traditional correlation-based methods. The findings demonstrate the potential of advanced regressionmodels to generalize soil parameter predictions across test sites, significantly improving reliabilitywhile reducing site-specific biases.