MELT-3D: Zero-shot cross-material prediction of three-dimensional full-field crystal-plasticity evolution with a latent transformer

Abstract Full-field crystal-plasticity surrogates require simulations from each target material. We demonstrate zero-shot prediction of the complete three-dimensional material state of an unseen material, including all stress, elastic-strain, and plastic-strain tensor components and crystallographic orientation. Prediction requires only the initial full-field state and constitutive descriptors, with no target-material trajectories for training, validation, fine-tuning, or normal-ization. MELT-3D (Microstructure Evolution via Latent Transformer) combines a variational autoencoder, a non-autoregressive conditioned transformer, and a hybrid mechanics–machine-learning model that predicts field statistics for normalization and reconstruction. Trained on Cu and CoCrFeMnSi and evaluated on CoCrFeMnNi, MELT-3D achieves accuracy comparable to held-out training-material microstructures. For example, its normalized voxelwise stress error is 0.196, compared with 0.184 and 0.192 for the held-out materials. Volume averaging the predicted axial-stress fields gives a 1.0% end-state macroscopic stress error. Error decomposition shows that encoding and decoding account for three-quarters of the stress-field error and most orientation error, while transformer-based latent-state prediction has smaller overall contribution.