When 1:00 PM - 3:00 PM May 26, 2023
Where 3158 HH Dow (Pod Room)
Add event to calendar vCal
iCal

Special seminar: "Practical deployment of machine learning for high-velocity science"


Joshua C. Agar
Drexel University

Machine learning has tremendous capabilities to advance scientific experiments. There are, however, significant challenges with the practical implementation of these tools. In particular, the absence of machine-interpretable physics-conforming models and computing infrastructure for automated analysis represent significant barriers. Here, we discuss our progress in codesign of scientific experiments, physics-informed machine learning models, and hardware implementations. We will discuss three tools. First, we will discuss our infrastructure for automated collection and collation of synthesis data including high-velocity in-situ diagnostics. This includes a federated and searchable scientific data management system. Secondly, we will show how physics-conforming neural networks can be developed to rapidly fit band-excitation piezoresponse force microscopy data. We show how to codesign this model for edge-real-time analysis on field-programable gate arrays. Finally, we develop a neural network architecture to automate 4D – scanning transmission electron microscopy strain mapping. This is achieved by building a cycle-consistent spatial transforming autoencoder where we embed an affine transformation to learn physics of geometric transformations parsimoniously. We discuss compression, regularization, and optimization methods required to automate this difficult training process. We achieve better performance (0.3 sub-pixel precision) compared to conventional template matching techniques as implemented in py4DSTEM. Ultimately, this work develops methodologies to seamlessly integrate AI systems with voluminous, high-velocity, and noisy experimental data.