Toward autonomous discovery of governing deformation mechanisms through interpretable spatial intelligence

Jean-Charles Stinville, University of Illinois Urbana-Champaign, USA

Advanced characterization techniques generate increasingly rich datasets to describe material microstructure and local deformation fields at high spatial resolution and over large fields of view. However, identifying the features that govern macroscopic material behavior within these fields remains a significant challenge. This presentation introduces Material Spatial Intelligence (MSI), an interpretable machine learning framework that learns latent representations of deformation and microstructure directly from multimodal microscopy data, including electron backscatter diffraction and high-resolution digital image correlation measurements. MSI establishes quantitative relationships between microstructural and deformation fields and macroscopic properties, including monotonic and cyclic mechanical properties, across a wide range of metallic materials. The framework goes beyond prediction by incorporating interpretable artificial intelligence approaches based on variational autoencoders (VAEs) or joint embedding predictive architectures (JEPAs) to identify the microstructural and deformation features, as well as their spatial organization, that most strongly influence material properties. The results demonstrate MSI’s ability to reveal physically meaningful deformation mechanisms and provide a foundation for autonomous discovery from advanced characterization multimodal datasets to accelerate the design and optimization of structural materials.

M. Calvat, H. Park, D. Anjaria, J.C. Stinville*
Materials Science and Engineering Department
University of Illinois Urbana-Champaign
Urbana, IL, USA
* Presenter