Collaborative Research: DMREF: Machine Learning-aided Discovery of Synthesizable, Active and Stable Heterogeneous Catalysts

Catalytic materials have long been used to improve the efficiency and product selectivity of many processes of vital importance to chemical manufacturing, petroleum refining, and pollution control. Given the complexity of catalytic reactions, and the need for the catalyst to operate under harsh conditions in many cases, considerable development effort – particularly from industry - has gone into the design of catalyst materials that can be readily synthesized, and that maintain stable performance for long time-on-stream. Academic research efforts, in contrast, have largely focused on theoretical, computational, and experimental identification of more active and/or lower-cost catalytic materials, but with little attention to synthesizability and stability. The project creates a new catalytic materials research framework that combines the search for more active materials with screening for synthesizability and stability under reaction conditions. The added complexity is addressed through the addition of powerful machine learning (ML) approaches that augment theoretical and computational tools to yield a more complete set of properties, or “descriptors,” associated with synthesizable, highly active, and stable catalytic materials. Ultimately, the goal is to package the various discovery tools in the form of an intuitive approach that delivers optimal results for catalysis practitioners. The project builds on the widely practiced descriptor approach to catalysis research, where a descriptor of catalytic activity (e.g., adsorption energy of an adsorbate) is computed using quantum chemical Density Functional Theory (DFT) calculations on various catalyst surfaces. Research efforts extend the current approaches by developing synthesizability, stability, and activity descriptors, using ML tools to rapidly screen through these descriptors, and collaborating with experimentalists in an iterative feedback loop to examine the accuracy of the predictions and to ensure the “catalysis practitioner-friendliness” of the combined methods. The approach will be developed in two case studies focusing on bimetallic catalysts for low temperature preferential CO oxidation in the presence of H2 (CO PROX) and partial oxidation of ethylene to ethylene oxide. The project will create a computer-aided workflow and open-source tools for predicting the synthesizability, activity, and stability of catalysts. By combining ML and DFT modeling with operando experimental characterization and testing, new structure-function relations will be identified for both reactions. In doing so, ML methods will advance beyond the prediction of activity for highly idealized systems to more realistic catalytic systems under operating conditions. Predicted materials structures and compositions will be validated against open-source high-fidelity experimental datasets in a feedback discovery loop that accelerates catalyst discovery. Beyond the technical component, the project will include outreach efforts focused on student professional development, broadened science participation, and informal science communication to help create a world-class scientific workforce. Cross-disciplinary training activities at the University of Michigan (U-M) and Wayne State University (WSU) will provide graduate and undergraduate students with a foundation to continue making scientific advances throughout their careers. A Data Science for Catalysis Training Program will enable undergraduates from WSU to visit U-M during the summer to learn the basics of data science and catalysis. Underrepresented students from Detroit schools, and their parents, will engage in science outreach events hosted by team members. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.