Machine Learning-Accelerated Computing of Molecular and Catalytic Properties of Metallic Alloys

Metallic alloys are important materials in engineering for their versatile properties. With the development of computational power, molecular simulation plays an increasingly important role to study the properties of metallic alloys. Molecular simulation is used to calculate the properties that are hard or expensive to be measured experimentally. Traditional molecular simulation relies on density functional theory (DFT) or empirical potentials to calculate the energies and forces during the simulations. DFT provides higher accuracy, but it is more time-consuming. Empirical potentials are much faster but they are not as accurate as DFT. Recently, machine learning (ML) potentials have become attractive because they are potentially both accurate and fast. When trained with sufficient data, machine learning potentials could be as accurate as DFT. Also, because machine learned potentials are essentially regression models, their inferences could be as fast as empirical potentials. With the help of machine learning models, simulations for large molecular systems can be accurately finished with a reasonable computational cost. 

In this dissertation, we combined ML methods, DFT, and Monte Carlo (MC) simulations to study the surface segregation of the CuPdAu alloy under vacuum conditions. We built a neural network (NN) model to accurately approximate the DFT potential energies during the MC simulations which were used to estimate the surface compositions of the CuPdAu alloy with various bulk compositions. A range of factors that might contribute to surface segregation was investigated, such as surface relaxation, vibrational contribution, and orientation dependence. Next, we developed a NN ensemble-based active learning method to accelerate the geometry optimization process, which enabled us to obtain the ground-state structures in a faster way. We used a NN-ensemble approach to provide the uncertainty estimation of the NN prediction during the molecular simulation, such that we were able to replace the DFT calculation using the NN model appropriately during the geometry optimization process. After that, the surface segregation and aggregation phenomenon was investigated using semi-grand canonical Monte Carlo simulation with the help of ML surrogate models for the bulk, slab potential energies, and the pseudo-adsorption energies. Then, we illustrated using automatic differentiation to evaluate the degree of rate control (DRC). Automatic differentiation provided higher accuracy and faster computation of the DRC. Lastly, we developed an efficient method to search for similar molecular structures in a large database. The search method was based on the approximate nearest neighbor search and machine learning embedding.