Purdue research team develops machine learning framework to describe complex catalyst surfaces using Gilbreth, Halstead clusters
Purdue chemical engineers have used Research Computing community clusters to develop a machine learning framework that creates various structural models of how a catalyst, aiding a chemical reaction, might be transformed under the reaction environment. The machine learning framework helps develop an atomic description of the complex surface, enabling research to build computational simulations which are closer to real-life materials.
The work was conducted by recent graduates Pushkar Ghanekar and Siddharth Deshpande and forms the basis of Ghanekar’s thesis. Ghanekar, who completed his doctorate under the supervision of Jeffrey Greeley, professor of chemical engineering, now works as a research scientist for Eli Lilly and Company.
“The problem that we usually face in doing this research is two-fold, the first being the efficient exploration of the enormous chemical space, and the second being the lack of tools to systematically generate configurations in said space,” explains Ghanekar.
That’s where the machine learning tool comes in, which quickly and cheaply ranks tens of thousands of possibilities and allows a researcher to save time and money by only running further calculations on a subset of models that look the most promising.
The machine learning framework, known as the Adsorbate Chemical Environment-based Graph Convolutional Neural Network (ACE-GCN), was developed on the Gilbreth cluster, which is optimized for GPU-intensive applications such as machine learning. The team also used the Halstead cluster to run electronic structure optimization calculations to feed into the machine learning tool.
Ghanekar worked with Research Computing staff experts Xiao Zhu and Amiya Maji to set up and maintain the electronic structure optimization code he used.
In addition to using the clusters for computation, the team used Data Depot to store their calculations and Research Computing’s GitHub service to host the machine learning models.
Ghanekar presented this work virtually at the Fall 2021 meeting of the American Chemical Society.