Machine Learning Approaches for Modeling Metal and Alloy Surfaces

We use models in science and engineering extensively. We use them to make predictions about the behavior of systems, to optimize designs, and to understand why systems behave the way they do. Most of our models are built from physical principles, and the parameters in the models are usually determined from measured data. That data is often expensive to gather, but the model is then cheap to evaluate. The accuracy of these models depends both on the depth of understanding we have, and the quality of the data. When we hit the limit of our understanding it is difficult to make better models. Machine learning offers a path forward to build models that are not necessarily based on physics, but which more accurately predict outputs.

We are interested in building models that allow us to perform molecular simulations that require many (hundreds of thousands) of calculations. These are not practical with quantum chemical calculations, which are too expensive to run at this scale. Existing molecular force fields are efficient enough for this, however, they lack the accuracy required to obtain meaningful results. I will present how we are using machine learning in conjunction with quantum chemical calculations to develop efficient models that can be used to simulate effects such as segregation, diffusion, etc., which can only be probed using simulation methods such as Monte Carlo and molecular dynamics. This approach is not fool-proof though, and we will show examples that worked well, and what we have learned from examples that did not work as well.

Machine learning has more to offer science and engineering than just model development. I will also discuss some aspects of how machine learning works, particularly the role that automatic differentiation has in machine learning. This has implications for many types of scientific programming, and may enable new ways to think about science and engineering problem solving.