Ultra-Compact and Parsimonious Machine Learning Frameworks for High-Velocity Scientific Discovery in Materials Microscopy

Machine learning (ML) offers unparalleled opportunities for accelerating advancements in materials microscopy. Yet, practical implementation often stumbles due to the lack of models that are both machine-interpretable and conforming to underlying physical laws, as well as the absence of specialized computational infrastructure for automated data analytics. We address these bottlenecks by focusing on the codesign of microscopy technique, ultra-compact physics-aware ML models, connection to physics-inspired models, and bespoke hardware solutions.<br/>Our contribution is three-fold. First, we present a robust infrastructure for the automated aggregation and management of high-velocity synthesis data, including real-time in-situ diagnostics. We introduce a federated, searchable scientific data management system optimized for high-throughput analysis. This infrastructure allows robust linkages between experiment and theory. Second, we go beyond traditional ML approaches by introducing highly parsimonious, physics-conforming neural networks specifically tailored for real-time analysis of band-excitation piezoresponse force microscopy data. These models are meticulously codesigned to be executable on field-programmable gate arrays (FPGAs) for edge-real-time analytics.<br/>Lastly, we unveil a neural network architecture aimed at automating 4D–scanning transmission electron microscopy strain mapping. Leveraging a cycle-consistent spatial transforming autoencoder, we incorporate an affine transformation layer to parsimoniously capture the governing equations of geometric transformations. This is further optimized through advanced compression, regularization, and optimization strategies to facilitate the otherwise computationally intensive training process. Remarkably, our method outperforms traditional template-matching techniques, achieving a sub-pixel precision of 0.3, as benchmarked against py4DSTEM.<br/>In sum, this work offers a comprehensive blueprint for seamlessly incorporating ultra-compact and parsimonious AI systems into the analysis of large-scale, high-velocity, and noise-prone experimental datasets. This facilitates closer connections between experiments and theory thereby revolutionizing the practical utility of AI in scientific discovery.