A new parallelized computational approach for parameter calibration of elastoplastic constitutive models

抄録

In geomechanical parameter searches for elastoplastic constitutive models conducted through element tests, state-of-the-art techniques are predominantly based on stochastic optimization algorithms to calibrate parameters against experimental stress-strain curves. However, these methods are prone to falling into local optima instead of identifying the global optimum and require extensive computation time. To overcome these challenges, we propose a novel parallelized computational method, implemented as an open-source program. By leveraging vectorization and parallelization on GPUs and TPUs, the method achieves over 1000× computational speedup, making global exhaustive search a practical alternative to stochastic optimization for complex parameter calibrations. To demonstrate its effectiveness, the method was applied to calibrate the geomechanical parameters of silica sand in consolidated-drained triaxial compression tests under different confining pressures. The results showed that exhaustive searches using GPUs completed the calibration in just 11 minutes, whereas Genetic Algorithms (GA) and Particle Swarm Optimization (PSO) required 15–21 hours. Additionally, the proposed method outperformed stochastic optimization in terms of Mean Absolute Error (MAE) and Mean Squared Error (MSE), achieving superior accuracy in matching experimental data. This advantage significantly enhances the accuracy and the efficiency of computational geotechnics, enabling more precise and scalable implementations of sophisticated elastoplastic models.