Multimodal Diffusion to Mutually Enhance Polarized Light and Low Resolution EBSD Data

A team of researchers led by Harry Dong from Carnegie Mellon University and Air Force Research Laboratory has introduced a novel multimodal diffusion model that tackles a critical bottleneck in materials science: the time-consuming process of 3-D electron back-scattered diffraction (EBSD) microscopy. Published on arXiv (2604.22212), the model leverages an unconditional diffusion approach to mutually enhance two imaging modalities—polarized light (PL) data and low-resolution EBSD data. The key innovation is that the model is trained exclusively on synthetic data yet demonstrates strong generalization to real-world datasets, which are often low-resolution, noisy, corrupted, and misregistered. This eliminates the need for expensive, high-quality training data.

In practice, the model achieves performance nearly indistinguishable from full-resolution EBSD using just 25% of the original EBSD data (i.e., 1/4 the resolution) combined with corrupted PL data. This represents a potential 4x speedup in data collection for serial-sectioning microscopy. The model also excels at multiple downstream tasks including grain boundary prediction, super-resolution, and denoising—all critical for materials characterization. By using inference-time scaling, the team demonstrated consistent performance gains across these objectives. This work bridges the gap between computer vision and materials science, offering a practical solution for accelerating 3-D microstructure analysis without sacrificing accuracy.