Accelerating and Improving the Accuracy of Parameter Calibration in a Phenomenological Crystal Plasticity Model Through High-Volume Machine Learning Simulations

Phenomenological crystal plasticity (CP) models are widely used in Integrated Computational Materials Engineering (ICME) to link microstructural features with engineering-scale mechanical behaviour. Their practical use, however, is limited by the high computational cost of physics-based simulations and the labour-intensive nature of parameter calibration, challenges that are amplified in additively manufactured materials with location-dependent properties. To address these obstacles, we first developed deep neural network (DNN) surrogate models of physics simulations to predict the stress–strain response of an additively manufactured AlSi10Mg alloy. Twenty-five experimentally derived scenarios (five microstructures × five sets of grain orientations) were used for training 25 separate DNNs, with datasets for validated material behaviour generated using the Düsseldorf Advanced Material Simulation Kit (DAMASK) platform and a Fast Fourier Transform (FFT)-based solver. Once trained, the DNNs produced stress–strain curves almost instantaneously, enabling an exhaustive grid-search exploration of a vast parameter space. Our approach yielded significant efficiency gains, which were comprehensively quantified. The best-fit CP parameters obtained through this approach are expected to be more accurate than those derived from conventional trial-and-error calibration, which is restricted to a limited number of candidate values. In addition, the minimum number of CP-FFT simulations required to train the DNNs with sufficient accuracy was identified, reducing the need for costly physics simulations in future studies. The proposed framework enhances the practical utility of CP models for simulation-informed materials engineering and optimisation and is broadly applicable to parameter identification in phenomenological models of other domains.