EAGER: ADAPT: Hypotheses Generation in Heterogeneous Catalysis using Causal Inference and Machine Learning

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

With support from the Chemical Catalysis program in the Division of Chemistry (CHE), the Catalysis program from the Division of Chemical, Bioengineering, Environmental and Transport Systems (CBET), and the Office of Multidisciplinary Activities (OMA), Bryan R. Goldsmith, Yixin Wang, and Suljo Linic of the University of Michigan, Ann Arbor will work to advance knowledge generation in heterogeneous catalysis using machine learning. They will develop the methods of interpretable machine learning and causal inference to generate hypotheses and extract insight of catalytic materials that are expected to lead to more predictive models of heterogeneous catalysts. This team’s research seeks to advance machine learning methods to find descriptors of catalytic performance (e.g., activity and selectivity) and, in this way, identify structure-property relationships that have the potential guide catalyst discovery efforts for important reactions pertaining to sustainability and energy applications. Their research addresses the National Science Foundation focus area of “AI for Concept Discovery”, and will benefit many areas beyond catalysis, such as enabling researchers to apply state-of-the-art machine learning algorithms to generate hypotheses and find new electrolytes or systems for renewable energy storage applications. This team will provide interdisciplinary training at the nexus of machine learning, statistics, and catalysis, which will help train an AI-aware workforce. They also will use a summer research internship program as a mechanism to broaden participation in AI-related STEM fields. Physically transparent models that can accurately quantify chemical and physical interactions between a surface of a material and adsorbate molecules (i.e., chemisorption) are crucial in many fields of chemistry and materials science. It has been known for a long time, going all the way back to the early 1900s, that chemisorption energies of adsorbates at gas/solid and liquid/solid interfaces are predictive descriptors of catalytic performance. There is a need to develop predictive theories of chemisorption that give insight into the underlying physical principles that govern chemical interaction at catalytic interfaces. Physically transparent and simple models that can accurately relate electronic and geometric features of a surface to its chemical properties and catalytic activity can allow us to rapidly predict or intuit which materials have specific chemical and catalytic features required for a particular application. This team will develop interpretable machine learning (i.e., models that can give researchers meaningful physical insights) and causal inference tools to generate hypotheses and extract insight that could lead to more predictive chemisorption models of heterogeneous catalysts. The team will focus on two state-of-the-art approaches; namely, generalized additive models and causal representation learning, and will advance these methods to understand adsorption of molecules on dilute alloy surfaces. A major goal is to identify causal links between electronic-structure, geometry, and chemisorption for dilute alloy catalysts. Although the focus here is on chemisorption and chemical catalysis on alloys, the developed methods are expected to be seamlessly integrated for use in other fields. 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.