Lisanne Knijff: Dipole and Charge Prediction for Electrochemical Systems from Atomistic Machine Learning

Abstract

Due to the increasing demand for energy, sustainable energy generation and storage are becoming more and more important in society and research. Electrochemical energy storage devices such as electrochemical double-layer capacitors (EDLCs) play an important role in fulfilling this need. To understand, control and design EDLCs at atomic precision, physical insights from atomistic simulation are clearly needed. However, atomistic simulation of EDLCs faces challenges such as the large system size and the complex chemistry involved at electrified solid-liquid interfaces. To address these challenges, machine learning models for charge prediction have been developed to aid the atomistic simulations of EDLCs in this thesis. Here, a divide-and-conquer approach was taken, and the electrolyte and electrode component were investigated separately.

Initially, a neural network approach called PiNet-dipole was developed to model the supercell polarization in liquid water using two constraints. First, the displacement of the atomic charges is proportional to the itinerant polarization. Second, each water molecule has a net charge of zero. In doing so, a molecular dipole moment distribution can be inferred for liquid water that is surprisingly similar to that computed from Wannier centers. More importantly, PiNet-dipole provides a way to predict atomic charge without resorting to any predefined charge partition schemes. This is followed by using a class of machine learning models called PiNet-chi to predict the response charge as the result of an applied electric field for both organic electrolyte molecules and graphene analogues. Both of these models were then upgraded through the addition of equivariant features in PiNet2. This opened up new ways of predicting dipole moment for both small molecules and condensed phase systems, allowing expansion to the PiNet2-dipole family and enabling a more expressive atomic charge prediction model.

Finally, by combining the PiNet(2)-dipole and the PiNet(2)-chi models and integrating them with the semi-classical molecular dynamics code MetalWalls, the PiNNwall interface was developed to model polarizable and heterogeneous electrodes. PiNNwall was then used to study chemically doped graphene and graphene oxide under different electrical boundary conditions, as well as to investigate the influence of the proton charge on aqueous EDLCs.