Mega-electronvolt (MeV) ultrafast electron diffraction (MUED) is a powerful structural measurement technique recently demonstrated to provide complementary capabilities for novel characterization of materials. MUED takes advantage of the strong interaction between electrons and matter and minimizes space charge problems seen in other low-energy electron diffraction systems.  An MUED instrument can resolve much finer structural details enabling us to see how atoms in molecules move and can make molecular movies of ultrafast chemical reactions by utilizing a pump-probe technique. While a few facilities have been constructed to provide these advanced capabilities to a growing user community, as a relatively young technology, progress is ongoing and much remains to be learned by improving these facilities, especially in maximizing throughput and ease of operation (e.g. rapid initialization and high stability). One class of advancement that can be immediately investigated at MUED facilities is the demonstration of realtime or near-realtime data processing enabled by data science/machine learning/artificial intelligence mechanisms in conjunctions with high-performance computing (HPC) to bring automated operation, data acquisition and processing to the facility control and experiment analysis, thereby enabling faster measurement throughout. For example, applying deep learning to MUED diffraction patterns generated through instrument probing of materials systems is an innovative step to turn-key, high throughput and high research output instruments. Such an approach has been demonstrated with success when applied to X-ray diffraction and many other experiments, and there is high confidence that significant gains are possible for MUED facilities.

The goals for this research project are to utilize the expertise of the PIs’ collaboration at DOE labs to identify the most promising machine learning (ML) and artificial intelligence (AI) techniques, as well as others used in the accelerator and facility user communities, apply and optimize them for improved operation of the Brookhaven National Laboratory Accelerator Test Facility MUED instrument in conjunction with enhanced computer interfacing and diagnostic capability of the machine, conduct online ML-control to provide the users the ideal beam with high stability for specific material types, provide experimenters with greatly improved usability and sample throughput, and consequently demonstrate increased scientific productivity of MUED facilities. The team will do this in seamless conjunction with a second user facility, the Argonne Leadership Computing Facility (ALCF). Scientific advancement will be enabled in a wide range of fields including accelerator science, materials science, ML control, and in AI for science with HPCs. Some of the materials science focuses the PIs are considering during the validation phase are on materials applied in diverse uses such as superconductors, battery and fuel cell components, high strength magnets, thermoelectrics, and quantum information storage and manipulation. MUED (and its cousin technique MEU Microscopy) is also emerging for interesting biomolecules. Of particular interest is the study of proteases that mediate diverse disease processes such as viral entry into cells (e.g. COVID-19) and cancer metastasis. This research will open the door to new collaborations between the UNM and LANL team members and an Office of Science User Facility at Brookhaven – the Accelerator Test Facility – that will enable high-throughput research with materials through the use of data science, ML and AI. Additionally, the UNM and LANL team members, as well as the BNL team members will have access to the ALCF. Finally, the Argonne team members will be interfacing their facility with the new genre of user facility (MUED) at BNL for new explorations in AI for science.