Physics-informed deep learning of dynamic 3D experiments

Successful applications of deep learning for data analysis such as classification and nonlinear regression are sometimes perceived as a ‘black-box’ magic. This is not a satisfying situation for experimental physics, where interpretation of observations through the framework of fundamental and extended physics is essential. Meanwhile, continuing advances in hardware and different imaging modalities such as visible light, X-ray and neutron, make image data increasingly accessible and as a result, machine learning appears to be unavoidable for automated information extraction from the oceans of data [1, 2]. Recent progress in data science indicates that physics and other knowledge such as geometry and material properties may be imbedded in physics-informed or physics-constrained deep neutral networks through a variety of means such as synthetic data generation, experimental data augmentation, automated uncertainty quantification, transfer learning, regularization by differential equations or scaling laws. Here we review and highlight recent progress in combining image data, deep learning, and physics models for dynamic experiments in a diverse set of 3D geometries: exploding wire experiments, ultrafast multi-phase processes in liquids, plasma-liquid interfaces, shock compression of materials, and laser/X-ray compression of materials. Our work aims at a holistic approach to a.) high-speed imaging, data science consistent with physics and human-level understanding and b.) experimental physics and material science involving dynamic 3D or 4D scenes. This research used resources of the Advanced Photon Source, a U.S. Department of Energy (DOE) Office of Science User Facility, operated for the DOE Office of Science by Argonne National Laboratory under Contract No. DE-AC02-06CH11357. The work is supported in part by LANL C2, LANL ICF, LANL LDRD, and NSF-MSGI programs.