(88e) Simulating Closed-Loop Catalyst Discovery Processes Using an Experimental Band Gap Surrogate Model
AIChE Annual Meeting
2022
2022 Annual Meeting
Topical Conference: Applications of Data Science to Molecules and Materials
Applications of Data Science to High Throughput Experimentation
Monday, November 14, 2022 - 9:00am to 9:15am
Heterogeneous catalyst discovery faces many challenges, some of which are common to materials discovery efforts: a curse of dimensionality arises from combinatorial design spaces; a lack of open-source datasets and plethora of experimental techniques make it difficult to predict experimental synthesizability or activity; and high-throughput synthesis techniques outpace structural characterization, meaning that the structures or properties models rely on often need to be inferred. Other factors are unique to catalyst discovery: because heterogeneous catalysis is a surface phenomenon, activity and stability must be considered at the level of the surface, and modifications made to increase surface area or stability often have impacts on activity that are hard to model computationally. Despite these difficulties, computational tools have proven useful in predicting catalyst synthesizability and activity in various settings, predicting effective catalysts as obscure as MoS2 and Pt3Sc on the basis of purely computational descriptors. However, closed-loop optimization processes that develop models of catalyst activity and iteratively acquire new experiments based on their expected utility have proven challenging to develop.
To better understand closed-loop catalyst discovery processes, I developed an optimization process to acquire experimental band gaps from a database on the Materials Project website. Crystal structures corresponding to experimental compositions were taken from the Materials Project database, and prototype structures were created in cases where there were no structures, and a graph model was used to predict various materials properties from these structures. An acquisition model using these structural features as well as composition features was then used to iteratively discover the experimental data set. Methods and knowledge developed by this effort are then applied to a simulated catalyst discovery problem with a similar structure.