Advanced neural network methods for atomistic materials chemistry

Thesis

Machine learned force fields (MLFFs) are functions trained to map the positions of atoms within a material or molecule to the potential energy and atomic forces generated by expensive, quantum-mechanical calculations. Their high accuracy and low cost are enabling faithful simulations of matter at the atomic scale over unprecedented time- and length-scales, revolutionising the field of computational atomistic modelling. Despite their great success, however, key problems inhibit the potential and widespread adoption of MLFFs. In this thesis, I present new methodologies to address several of these in turn.

To address the high upfront costs of labelling MLFF training data, I present work that uses cheap and approximate methods to generate and label large datasets of atomic structures. I demonstrate that “pre-training” MLFFs on this “synthetic data” can simultaneously enhance model accuracy, reduce the overall cost of training data generation, and improve the robustness of models’ predictions during dynamic simulations.

To address the sub-optimal inference speeds of large MLFF models, and in particular of recently developed atomistic foundation models, I present a new approach to MLFF “distillation”. I develop an automated pipeline to explore chemical space using an expensive “teacher” model, before training a cheaper “student” model to mimic the teacher’s predictions on this dataset.

To address the lack of interpretability of modern MLFFs, which are typically black-box neural network models, I present work that learns symbolic force field models directly from data, and in an end-to-end manner. These simple models recover the speed and interpretability of existing empirical models while (greatly) improving upon their accuracy.

In all of the above, I use gradient descent to train neural-network-based MLFF models on large datasets of labelled atomic structures. I therefore begin by exploring and motivating this approach, starting from a simple toy system, and building up to a real world use case. As part of this demonstration, I present graphpes, a software package I have developed over the course of my DPhil for training, evaluating and applying MLFF models.

Authors: Gardner, J

• DOI: 10.5287/ora-5jmxb8ovv
• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English
• Keywords: machine learning; synthetic data; neural networks; distillation; chemistry; MLFF
• Subjects: Science; Chemistry, Physical and theoretical; Machine learning; Supervised learning (Machine learning)