Exploring the transferability of machine-learning models for analyzing XRD data of shocked microstructures: from single crystal to polycrystals

Abstract

This study explores the transferability of machine-learning models to analyze X-ray diffraction (XRD) profiles of shock-loaded single-crystal and polycrystalline data. Transferability in this context refers to the ability of these models to accurately predict microstructural descriptors for crystal orientations and structures not included in its training data. Supervised machine-learning models were trained on XRD profiles and microstructural descriptors from atomistic simulations to extract properties like pressure, temperature, phase fractions, and dislocation density. We assessed two aspects of transferability: (1) the ability of models trained on specific single crystal orientations to predict microstructural descriptors for other orientations, and (2) the capacity of models trained on single crystal data to analyze polycrystalline structures. Results show promising accuracy in predicting certain descriptors within the same orientation and improved transferability to new orientations and polycrystalline systems when trained on multiple orientations. However, the accuracy of these predictions depends on the microstructural descriptor being targeted and the specific crystal orientations included in the training dataset. This work highlights the potential and limitations of machine learning for analyzing XRD data of shock-loaded materials and emphasizes the need for diverse training data to enhance model transferability and robustness.