Physics-Guided Machine Learning Framework for Real-Time Multi-Scale Materials Characterization at Light Sources

The real-time characterization of material microstructures during in situ experiments is fundamental to advancing our understanding of material behavior and properties. Traditional methods for analyzing three-dimensional multiscale data from advanced light sources face several critical challenges during dynamic loading conditions. While conventional 3D measurements require collecting multiple projection data through sample rotation or rocking curves, dynamic experiments preclude such mechanical movements, necessitating alternative approaches such as multi-detector configurations and advanced algorithm developments.



Advanced light sources offer sophisticated techniques for probing materials across multiple length scales, yet integrating these multi-modal techniques within a single dynamic experiment remains challenging. Computational bottlenecks in data analysis prevent real-time identification of critical phenomena, limiting our ability to adaptively guide measurements and efficiently utilize multi-modal characterization techniques.



In this work, we present machine learning approaches that provide real-time feedback for 3D multi-scale microstructure characterization during in situ experiments. We developed a generative deep learning framework for reconstructing crystal orientations from diffraction data, combining conditional generative networks with physics-informed constraints. The framework processes multiple diffraction patterns simultaneously through an optimized convolutional neural network design for real-time predictions while maintaining high accuracy.



We also developed physics-informed deep learning surrogate models for predicting crystal orientation and strain field evolution during deformation. Our hybrid architecture integrates fundamental principles of crystal plasticity and conservation laws directly into the neural network through custom loss functions.



The primary impact of this work will be realized through the integration of adaptive ML-accelerated diffraction analysis with physics-informed surrogate models, which together can provide continuous feedback for guiding multi-scale characterization during in situ experiments.