Predicting Catalyst Surface Stability Under Reaction Conditions Using Deep Reinforcement Learning and Machine Learning Potentials
Catalysts are critical for most large-scale energy intensive chemical transformation processes, such as energy storage, liquid fuel production, and the formation of chemical building blocks. The catalyst composition, structure, and morphology impact the performance under reaction conditions and influences properties like activity and selectivity. Many industrial catalysts offer imperfect activity/selectivity, or contain expensive metals. Further, catalysts can deactivate over time as the morphology changes or harsh reactive environments alter the surface structure. Methods to automatically model and predict the kinetics of how catalyst surfaces will restructure would enable engineers to design around these challenges, improve performance, increase longevity. This project investigated a specific type of machine learning model, deep reinforcement learning, machine learning models to act as surrogates for the physical system, and compared the results of these approaches with a specialized high-throughput experimental synthesis and measurement platform. This project leveraged a rapidly growing class of machine learning models, deep reinforcement learning models, to systematically explore and identify kinetic pathways for catalyst surfaces. Specifically, these models were allowed to interact directly with an atomistic simulation by moving atoms, running short simulations, and observing the impact of these changes on the energy of the system. By creatively exploring many of these pathways, the system was autonomously able to identify kinetic pathways that were similar to those that would be identified by a trained expert in the area. A key component of this approach is the availability of accurate and fast simulation techniques as the reinforcement learning approach employed typically ran thousands of simulations to identify reasonable pathways. A specific type of machine learning surrogate model based on atom-centered symmetry functions was employed in this project to accurately represent the catalyst surface and behavior. The approaches developed in this project were applied to metal alloy catalyst surfaces with three components but the approach could be used by other groups in the future for other types of surfaces and simulations. High-throughput experimental techniques were used to measure and confirm predictions for near-surface segregation in ternary alloy materials. A specialized high-throughput deposition system was used to prepare surfaces with a wide range of possible compositions. These surfaces were then exposed to varying static and dynamic temperatures which induced segregation to varying degrees. The surfaces were then characterized to measure the changes from before the experiments. These experiments qualitatively agreed with the predictions from the ML modeling approach above, and provided insight into how the models could be further optimized. Together, the modeling, reinforcement learning, machine learning, and high-throughput experiments address the challenges of predicting how catalyst surfaces will change under reaction conditions. This project also supported a tech to market strategy that included consultations with potential users from the chemical industry, connections with national labs and other potential users through the ARPAE SUMMIT, and release of software and documentation as open source software. These advances could be leveraged both by catalyst manufacturers, and by the broader atomistic modeling community for other possible materials that require discovery of kinetic pathways.
- Research Organization:
- Carnegie Mellon Univ., Pittsburgh, PA (United States)
- Sponsoring Organization:
- USDOE Advanced Research Projects Agency - Energy (ARPA-E)
- DOE Contract Number:
- AR0001221
- OSTI ID:
- 2324766
- Report Number(s):
- DOE-CMU-0001221-1
- Country of Publication:
- United States
- Language:
- English
Similar Records
Annual Report: Fuels (30 September 2012)
Improving surrogate model accuracy for the LCLS-II injector frontend using convolutional neural networks and transfer learning