(158f) Leveraging DFT with Machine Learning: Applications in Catalysis
Density functional theory has been a standard computational tool in catalysis for many years. With it, electronic structure/activity correlations have been discovered, linear scaling relationships, and new ideas about catalytic sites and mechanisms. The dramatic growth in computing power over the last decade has made it possible to routinely run thousands of DFT calculations today. There are many scenarios, however, where even thousands of calculations is not enough, but where the accuracy of DFT is desirable. For example, in Monte Carlo or molecular dynamic simulations, hundreds of thousands of calculations might be required to get reasonable statistical sampling or simulation times. Machine learning (ML) has emerged as a way to leverage DFT calculations. In this approach, a flexible model is trained from DFT calculations and then after it has been validated, the model is much more efficient than DFT with the same accuracy. We will show some examples of this from our work in modeling diffusion and adsorption on palladium surfaces. We will show how the use of ML enables us to use DFT to compute new physical properties that aren't easily obtained with DFT, and atomistic properties that are not easily obtained by experiments. If time allows, we will discuss some implications of modern machine learning tools for other applications in catalysis.