Using ML Tools to Enable High-throughput Studies of Amorphous Material Surfaces, and Its Application to Plasma Etching

Several high-impact energy applications, including photovoltaics, energy storage, transistors, phase change memory devices, and catalysis, employ functional, amorphous materials. For several of these technologies, amorphous surfaces and interfacial chemistry and reactivity provides a key piece of information to understanding the performance of the material. One such complex surface process, of importance to the silicon semiconductor industry, is plasma etching. Plasma etching is commonly used in patterning processes where a proper etch rate is critical to transistor manufacturing and performance. Dry etching is ubiquitous in the semiconductor manufacturing process due to its highly anisotropic and selective etching mode.<br/>Today's accelerated materials design paradigm is increasingly leveraging modeling and in-silico screening of materials properties, however, it has been less utilized for amorphous materials due to the increased structural complexity and simulation time needed to properly capture amorphous properties. Even more challenging is the modeling reactions on amorphous surfaces due to the large chemical and structural diversity of surface sites. In principle, an a priori intractable number of local surface environments with potentially different reaction kinetics needs consideration.<br/>In this work, we present a method to perform site reduction on amorphous material surfaces and benchmark the site reduction technique by utilizing it to model a plasma etching reaction. Amorphous material surfaces are generated in high-throughput by combining Parametrized Method 6 (PM6), with Perdew-Burke-Ernzerhof (PBE) Density Functional Theory (DFT). We perform liquid-quench molecular dynamics simulations using PM6 and follow with DFT calculations, generating six potential surfaces per material. We perform site featurization by using the Smooth Overlap of Atomic Placement (SOAP) and make use of a Bayesian Gaussian Mixture model to cluster similar sites. For further analysis, we then perform calculations at sites closest to the cluster center and assume all sites within that cluster have similar values using a weighted average.<br/>To benchmark the site reduction, we created an automated method for generating Nudged Elastic Band (NEB) inputs for modeling the kinetics of etching reactions. The SOAP and Gaussian Mixture model used in this work allowed us to get a representative value of the reaction barrier for etching of a-Si forming SiCl₄ with a 70meV error from the true value with 6 sites sampled out of 17 potential sites, an over 20-fold reduction of potential error in comparison to not utilizing this model, in addition to vastly increasing the turnaround time to study processes on these surfaces. Additionally, the model was generalized to other combinations of etchants and amorphous materials, including a-Si etched by SiF₄, and a-C etched by both CH₄ and CCl₄. The model captured proper trends with a systemic error across all combinations tested. This material agnostic approach has the potential to dramatically speed up the scientific study of amorphous material surfaces.