Unveiling Process-Structure Mapping with a Deep Variational Autoencoder

Avanish Mishra, Brenden William Hamilton, Mashroor Shafat Nitol, Nithin Mathew, Kipton Marcos Barros, Timothy Clark Germann, Saryu Jindal Fensin; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 3652-3661

A comprehensive understanding of the process-structure-property relationship is crucial for designing materials with superior performance. While the structure-property relationship is well developed, the process-structure mapping remains less explored. Herein, we present a pre-training framework for establishing process-structure mappings via deep variational autoencoders (VAEs). We generate a large dataset of simulated 3D microstructures with varied grain sizes, distributions, and orientations using molecular dynamics. We produce virtual inverse pole figures (IPFs) for microstructures, generating images that mimic experimental electron backscatter diffraction (EBSD) micrographs. We then train VAEs to learn a compact, semantically meaningful latent space of microstructures, to enable efficient and smooth interpolation across simulated data. We validate our approach by latent-space interpolation on synthetic microstructures and by reconstructing large-area experimental micrographs using a patch-based "context-zooming" strategy. These results demonstrate an initial exploration of process-structure mapping via generative modeling. Establishing explicit and quantitative correlation between latent variables and processing conditions remains future work, as it will require a large, controlled experimental dataset.