Abstract

Soil constitutive models are essential for predicting geomaterial behavior under various loading conditions, serving as a foundation for reliable and efficient construction and infrastructure design. Advanced plasticity models can enhance predictive accuracy in geotechnical applications involving complex environmental and mechanical loads. However, the development, calibration, and numerical implementation of these models are often challenging, time‑intensive, and susceptible to subjectivity. This study presents a novel framework called ‘generative computational artificial intelligence,’ designed to automate the formulation, calibration, and numerical implementation of constitutive models. Focusing specifically on SANISAND plasticity, the framework features a collection of cooperative agents that autonomously handle experimental data as inputs, develop constitutive models, calibrate parameters, and generate implementation code, all with minimal human intervention. The methodology involves several core stages: (1) feature extraction to identify critical behaviors, such as elasticity and critical state, from data; (2) parameter identification to optimize parameters aligned with these behaviors; (3) model ranking to assess the performance of candidate models; (4) construction and implementation of the final constitutive model based on ranked candidates; and (5) final optimization using particle swarm optimization.

Animation guide: The left sidebar lists stages. The canvas panel shows (1) Feature Probing extracting Elasticity/Critical State/Plasticity, (2) Parameter Identification filling parameter chips per candidate, (3) Model Ranking computing RMSE and selecting winners, (4) Formulation placing SANISAND blocks from winners, and (5) PSO converging to a global best.

Publications & Repositories