Productive automation of calibration processes for crystal plasticity model parameters via reinforcement learning

Abstract

Crystal Plasticity Finite Element (CPFE), which merges crystal plasticity principles with finite element analysis, can simulate the anisotropic grain-level mechanical behaviour of polycrystalline materials. Due to the benefit of CPFE, it has been widely utilised to analyse processes such as manufacturing, damage, and deformation where the microstructure plays a prominent role. However, this method is computationally expensive and requires the robust calibration of its parameters, which can be many. In this work, we propose a framework to address difficulties in calibrating multi-parameter CPFE. The Deep Deterministic Policy Gradient (DDPG) algorithm, a Deep Reinforcement Learning (DRL) approach, is utilised to optimise the CPFE parameters. Additionally, a Python-based environment is developed to fully automate the calibration process. To allow comparison with the conventional optimisation method, the Particle Swarm Optimisation (PSO) algorithm is also used, which shows the DDPG framework yields more accurate calibration. The generalisation performance of the proposed framework is also demonstrated by calibrating each parameter set of two different CPFE models for monotonic loading of the stainless steel type, 316L. Moreover, the effectiveness of the framework in the more complex condition is also demonstrated by calibrating the CPFE parameters for a two-cycle cyclic behaviour of a 316H stainless steel material. The reliability of these calibrated parameters is also validated in the cyclic simulation after two cycles.