Machine-learning for force-fields in molecular simulation: Water, Metal Oxides and their Interfaces

Photoelectrochemical (PEC) water splitting cells, used to create hydrogen from solar energy, are crucial to the implementation of solar-to-fuel technology. Research in science and technology is focused on improving the functionality and efficiency of these devices while also ensuring that they will last for a long time and will be affordable. Machine-learning based computer-aided simulation is one of the methods exploited to design the next generation of photoelectrochemical cells. Such simulation techniques have filled the gap between conventional force fields, which are fast and inaccurate, and electronic structure simulations, which are slow and accurate. In this thesis, we developed and/or improved machine-learning interatomic potentials (MLPs) for all the chemical systems involved in the design of PEC cells. Having established the reliability of such potentials in describing the many-body share of energy for argon clusters, we then utilize neural networks as a powerful machine-learning method to construct potential energy surfaces for the other studied systems. High-dimensional neural-network (HDNN) model developed by Behler et.al (Behler, 2007) was used to obtain the coordinates of a chemical system as input and output its total energy. We used atom-centered symmetry functions (Behler, 2011) to encode the chemical environment into a fixed-length input vector. As part of this thesis, various sampling schemes are examined and improved to construct the training set needed for model development, as well as a new technique to manipulate publicly available data for model development is implemented using machine-learning. In addition, the network's weights and biases are adjusted using different optimization techniques. We also implemented a molecular dynamics code that includes long-range electrostatic energy via Ewald sums and delta-machine-learning, and we proposed and developed a technique based on checkpoint ensembles in machine-learning to improve the accuracy of a neural network model for total energy prediction.