Efficient Nonlinear Multiscale Prediction for Unseen Polycrystalline Textures via Self-Supervised Microstructure Pretraining

Abstract

Predicting the nonlinear mechanical response of polycrystalline materials across diverse crystallographic textures remains computationally prohibitive, and existing reduced-order surrogates are typically fit to a single microstructural realization, precluding reuse on unseen textures. We address both limitations through a self-supervised pretraining strategy. A three-dimensional masked autoencoder is pretrained on 100,000 voxelized synthetic face-centered cubic (FCC) microstructures whose textures systematically span the texture hull via hierarchical simplex sampling, yielding transferable, texture-aware latent representations. A differentiable homogenization operator then maps these representations to the parameters of an orientation-aware interaction-based deep material network (ODMN). Given a previously unseen microstructure, the pretrained encoder infers a standalone ODMN that, coupled with crystal plasticity, reproduces the nonlinear loading-unloading-reloading stress-strain response with mean relative error below 2%, at a 304x CPU-time speedup over full-field crystal-plasticity direct numerical simulation. The pretrained representation is also highly data-efficient: on a label-limited homogenized-stiffness regression task, pretraining raises the validation R^2 from below 0.1 (trained from scratch) to above 0.8. Together, these results demonstrate that self-supervised pretraining yields physically meaningful, transferable microstructural representations and provides a scalable framework for microstructure-property inference. The present scope (FCC systems with equiaxed grains) is a deliberate first step, with extensions to morphological texture and other crystal systems outlined.