Machine Learning Based Constitutive Models for Predicting Stress Strain of Sands

Document Type

Conference Proceeding

Publication Date

1-1-2026

Abstract

Predicting soil behavior is one of the most important topics in geotechnical engineering for design and analysis purposes. With advances in computational capabilities, data-driven and machine learning (ML) models have gained attention in predictive studies, including constitutive modeling of material behavior. In this study, four ML models, Recurrent Neural Network (RNN), Long Short-Term Memory (LSTM), Gradient Boosting (XGBoost), and Radial Basis Function Networks (RBFN), were used to develop a data-driven constitutive model and predict the stress-strain behavior of three sandy soils during direct shear tests. The stress-strain data for Ottawa, Glacial, and Dune sands were obtained under 10 normal loads ranging from 222 N to 1,779 N. The dataset was preprocessed to normalize features and structured for each ML approach. Strain, applied normal load, and soil type were used as input features, with shear stress as the output. A universal model was also developed for all soils. In addition, the performance of each of these models was discussed based on the metrics, the soil type, and the comparisons with actual soil behavior. Model evaluation showed that LSTM achieved the highest performance with overall R = 0.77, capturing nonlinear soil behavior. RNN achieved an overall R = 0.50, while RBFN and XGBoost performed poorly with overall R values of 0.32 and -0.41, respectively. This study demonstrates that memory-based sequential models, particularly LSTM, provide superior predictive capability in modeling nonlinear and evolving soil stress-strain behavior. By comparing different approaches across soil types and loading conditions, this research provides valuable insights into the suitability of ML techniques for developing data-driven constitutive models in geotechnical engineering.

Publication Title

Geo Congress 2026 Soil Properties Modeling and Computational Geomechanics Selected Papers from Geo Congress 2026

ISBN

[9780784486719]

Link to Full Text

Share

COinS