397401 Advanced catalyst design strategies: Integrating machine learning and probability theory with quantum mechanics and descriptor-based analysis

Sunday, November 16, 2014
Galleria Exhibit Hall (Hilton Atlanta)
Andrew Medford, SUNCAT Center for Interface Science and Catalysis, SLAC National Accelerator Laboratory, Menlo Park, CA

Chemical processes driven by heterogeneous catalysts are central to a broad range of existing and emerging renewable energy technologies. The ability to engineer novel catalysts which are more active, selective, or inexpensive for a given chemical transformation has the potential to enable or radically improve processes for sustainable energy conversion. A particularly exciting development in the search for novel catalysts is the ability to design catalysts in silico using computational methods. A number of promising strategies for computer-aided catalyst design have been developed, including descriptor-based optimizations, genetic algorithms, and high-throughput computational screening. The previous success of in silico catalyst design demonstrates the potential power of catalyst prediction through fundamental understanding.

The problem of catalyst design can be considered in terms of two essential challenges: establishing a connection between atomic-scale models and observed reaction rates, and systematically predicting novel catalysts and processes based on these models. My work will utilize machine learning and probability theory to more efficiently utilize existing theoretical and experimental results to provide fundamental understanding and robust predictions. 

Understanding the connection between atomic-scale models and observed rates is particularly difficult due to the ubiquitous nature of uncertainty in heterogeneous catalysis. The potential importance of transient species, rare events, and sparse active sites makes experimental investigations difficult; moreover, the complex electronic structure of surfaces and vast possibilities of active site geometries presents a challenge for even the most advanced computational methods. Using probability theory as an extended logic, my research will seek to account for uncertainty in order to develop quantitative connections between atomic-scale models and catalytic activity.

The development of predictive methods for catalyst discovery relies on access to large amounts of data and innovative approaches to data analysis. The current state of computational catalysis research places the field on the brink of a data-rich era, where research is limited by data analysis rather than data generation. The impact of available information can be maximized by the use of improved methods for data analysis and catalyst optimization. Descriptor-based analyses provide one route towards insight-based catalyst discovery. My research will build on descriptor methods by using more advanced statistical methods and generalizing the design space beyond surface properties.

Extended Abstract: File Not Uploaded
See more of this Session: Poster Session: Meet the Faculty Candidate
See more of this Group/Topical: Education Division