Machine learning and data-driven methods in computational surface and interface science

Introduction

Many modern technological challenges crucially depend on the properties of surfaces and interfaces. This includes the control of charge and energy transfer across electrode/electrolyte interfaces in batteries¹ and fuel cells², the characterization of structure and dynamics of wear and lubrication at tribological interfaces³, the optimization of chemical transformations at metal surfaces in heterogeneous catalysis⁴, corrosion science⁵ and surface functionalization⁶. Surface/thin film growth processes such as atomic layer deposition, chemical vapor deposition, or molecular beam epitaxy are highly industrially relevant. Increasing demand for functional thin-films and surface nanostructures also increases their structural complexity. As most modern materials involve multiple components, understanding the structure and properties of thin films, composites, buried interfaces, and exposed surfaces is now more important than ever. The atomic-scale characterization and design of functional interfaces require understanding and manipulation at the nanoscale, which often cannot be delivered by experimentation alone.

Computational simulation of surface and interface processes has become central to modern surface science. Few fields rely as strongly on the synergy between atomistic simulation and experimental study. This synergy is achieved by minimizing the gap between experimental complexity and simulation models, often through the use of model surfaces and ultra-high vacuum conditions, which, for example, enable atomic resolution to be achieved in scanning probe microscopy. However, surfaces in real-world applications are often more complex, featuring defects and partial disorder. Additionally, ambient pressure and interacting molecules play crucial roles in many applications, such as catalysis. By advancing the study of complex surface systems and dynamic processes at large length- and time scales with high throughput, machine learning (ML) and data-driven approaches have the potential to bring atomistic simulation and experiment even closer, offering improved mechanistic understanding of surface dynamics, reaction pathways, growth processes, and mechanical and electronic properties.

Common ML methods used in surface science encompass neural networks (NN), Bayesian regression methods, decision trees, support vector machines, and genetic algorithms. Citations for specific uses are given in later sections. These methods can learn expressions for formation energies, potential energy surfaces (PES), and other properties, provide frameworks to efficiently explore the configuration space of the material, or facilitate the optimization of a target property. MLIPs, in particular, are highly impactful and have revolutionized...