Designing efficient, interpretable, and generalizable machine learning interatomic potentials

Interatomic potentials (IPs) are invaluable tools in the fields of computational materials science and chemistry for their ability to accelerate atomic-scale simulations beyond the length- and time-scales that are accessible using first-principles techniques. In recent years, the application of machine learning (ML) and deep learning (DL) models and algorithms towards IP development has been a major area of interest, where machine learning interatomic potentials (MLIPs) are seen as being more flexible and accurate than their classical potential counterparts. The resounding success of MLIPs has led to a major shift away from the physical foundations characteristic of classical models, and towards more data-driven methods. While a data-centric fitting approach guided by error-based training metrics can be expected to correlate well with accuracies on higher-level property predictions, it has also resulted in a pattern of developing slower, more complex, and less general models. This work aims to address these issues by rectifying the false dichotomy between classical and ML IPs, and developing models and techniques that show how IP design can be improved with a focus on speed, interpretability, and generalizability. By first performing an in-depth comparison of the performance of a classical spline-based MEAM (s-MEAM) IP relative to a collection of MLIPs, I demonstrate the competitive nature of s-MEAM on a variety of common benchmarking tests. s-MEAM is shown to be capable of achieving errors comparable to those of the benchmarked MLIPs while maintaining its high speeds and interpretability, establishing its position on the accuracy-speed pareto front. These results demonstrate that high model complexities may not be strictly necessary in order to achieve near-DFT accuracy for certain benchmarking tasks and suggest an alternative route towards sampling the high accuracy, low complexity region of model space by starting with forms that promote simpler and more interpretable interatomic potentials I then build upon these results by leveraging the strengths of both s-MEAM and modern neural network (NN) architectures to propose a novel MLIP framework. The proposed framework, which I call the spline-based neural network potential (s-NNP), is a simplified version of the traditional NNP that can be used to describe complex datasets in a computationally efficient manner. I demonstrate how this framework can be used to probe the boundary between classical and ML IPs, highlighting the benefits of key architectural changes for improving model accuracy and interpretability. Finally, I present a metric using the entropy of the loss landscape, and show how it can be used to predict model performance on out-of-domain data and provide insights regarding model design and optimization. Using this metric, I demonstrate how architectural and optimization choices influence the generalization capacity of neural network (NN) IPs, revealing trends in molecular dynamics (MD) stability, data efficiency, and loss landscapes. With a large-scale study on two state-of-the-art MLIPs, and their optimizers, I show that the metric of loss entropy predicts out-of-distribution error and data efficiency despite being computed only on the training set.

Name thesis:degree_name

Ph.D.

Level thesis:degree_level

Dissertation

Discipline thesis:degree_discipline

Materials Science & Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2023

Author dc:creator

• Vita, Joshua Alexander

Contributors dc:contributor

• Trinkle, Dallas
• Bellon, Pascal
• Tadmor, Ellad
• Schleife, Andre

• Machine Learning
• Interatomic Potentials
• Materials Science

Statement dc:rights

• Copyright 2023 Joshua Vita

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/121918

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F121918