Enabling Real-Time Human/AI Collaboration During Data Intensive Synchrotron Light Source Studies with Constrained Matrix Factorization

With the National Synchrotron Light Source II (NSLS-II) coming online in 2015 as the brightest source in the world, the imminent upgrades at the Advanced Photon Source (APS-U), Advanced Light Source (ALS-U), and Linear Coherent Light Source (LCLS-II), and advances in detector technology, the data generation rates at x-ray light sources are skyrocketing¹. While such advances open the door to new high throughput and <i>in situ</i> studies, these data intensive studies make prompt analysis of the data difficult, leading to researchers often ‘flying blind’ while at the beamline, opening the door to mistakes or missed opportunities that are not revealed until weeks-to-months after the experiment completes². To fully leverage the capabilities offered by advanced synchrotron light sources, new methods of analysis must be developed that can keep pace with such data intensive experiments.<br/><br/>We have developed a method of Constrained Matrix Factorization (CMF) which is both efficient and highly scalable for real-time analysis of beamline data³. Beyond the positivity constraint found in standard Non-Negative Matrix Factorization (NMF), our algorithm allows users to apply additional constraints to both the weights and components used to fit the data. In this way, researcher provided intuition or prior knowledge can be injected into the fitting procedure, producing more physically relevant and interpretable results. This process can be done dynamically and interactively during an experiment, providing model-free insights into the progress of a study and guiding scientists towards their research objectives. We will present the details of the method, as well as several examples of its use during recent <i>in situ </i>studies including temperature dependent studies of molten salts⁴ and decomposition studies of Metal Organic Frameworks (MOFs) during gas flow reactions⁵. We will also present how this and other recently develop AI-driven tools^6,7 can be readily integrated with Bluesky⁸, the open source python data acquisition system initially developed at NSLS-II.<br/><br/><b>References</b><br/>[1] Ratner, Daniel, et al. (2019). <i>BES roundtable on producing and managing large scientific data with artificial intelligence and machine learning</i>. DOE SC Office of Basic Energy Sciences.<br/>[2] Olds, D. (2020). Synchrotron X-ray Diffraction for Energy and Environmental Materials: The Current Role and Future Directions of Total Scattering Beamlines in the Functional Material Scientific Ecosystem. <i>Synchrotron Radiat. News</i>, <b>33</b>(5), 4-10.<br/>[3] Maffettone, P.M., Daly, A.C., Olds, D., <i>Appl. Phys. Rev </i>(In Press)<i>.</i><br/>[4] Li, Q. J., Sprouster, D., Zheng, G., Neuefeind, J. C., Braatz, A. D., Mcfarlane, J., Olds, D.,... & Khaykovich, B. (2021). Complex Structure of Molten NaCl–CrCl3 Salt: Cr–Cl Octahedral Network and Intermediate-Range Order. <i>ACS Appl. Energy Mater</i>, <b>4</b>(4), 3044-3056.<br/>[5] Ganesan, A., Purdy, S. C., Yu, Z., Bhattacharyya, S., Page, K., Sholl, D. S., & Nair, S. (2021). Controlled Demolition and Reconstruction of Imidazolate and Carboxylate Metal–Organic Frameworks by Acid Gas Exposure and Linker Treatment. <i>Ind. Eng. Chem. Res</i>.<br/>[6] Maffettone, P. M., Banko, L., Cui, P., Lysogorskiy, Y., Little, M. A., Olds, D., ... & Cooper, A. I. (2021). Crystallography companion agent for high-throughput materials discovery. <i>Nat. Comput. Sci.</i>, <b>1</b>(4), 290-297.<br/>[7] Banko, L., Maffettone, P. M., Naujoks, D., Olds, D., & Ludwig, A. (2021). Deep learning for visualization and novelty detection in large X-ray diffraction datasets. <i>npj Comput Mater</i> <b>7, </b>104.<br/>[8] Campbell, S. I., Allan, D. B., Barbour, A. M., Olds, D., Rakitin, M. S., Smith, R., & Wilkins, S. B. (2021). Outlook for artificial intelligence and machine learning at the NSLS-II. <i>Mach. learn.: sci. technol</i>, <b>2</b>(1), 013001.