Enhancing Robustness of Neural Network Interatomic Potentials through Sampling Methods and Uncertainty Quantification
Neural network interatomic potentials (NNIPs) are a significant advancement in computational materials science and chemistry for their ability to accurately approximate the potential energy surface (PES) of atomic systems with significantly reduced computational costs compared to quantum mechanical...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2024
|
| Online Access: | https://hdl.handle.net/1721.1/155357 https://orcid.org/0000-0001-6731-5531 |
| Summary: | Neural network interatomic potentials (NNIPs) are a significant advancement in computational materials science and chemistry for their ability to accurately approximate the potential energy surface (PES) of atomic systems with significantly reduced computational costs compared to quantum mechanical methods. Without relying on predefined interaction parameters, NNIPs offers greater flexibility and adaptability than classical force fields, and can be used for atomistic simulations of complex materials and biological systems. However, NNIPs face inherent limitations due to their dependence on diverse training data and limited extrapolative capabilities. This thesis proposes methodologies to address these challenges through analysis of uncertainty quantification (UQ) techniques, introduction of novel data sampling strategies, and development of structural similarity analysis algorithm to extract physical insights from diverse data sets. First, we examine the efficacy of UQ for single deterministic neural networks, demonstrating that a Gaussian mixture model-based approach can significantly reduce computational demands without sacrificing prediction accuracy and UQ reliability, although it does not significantly outperform the baseline ensemble method. Utilizing insights gained from the UQ analysis, we introduce a PES sampling technique based on adversarial attacks on predicted uncertainties, which samples atomic configurations with maximized uncertainties and mitigates the typical correlation issues associated with molecular dynamics sampling. Additionally, recognizing the limitations of the proposed adversarial sampling method, we introduce an enhanced sampling method using predicted uncertainty as collective variables (CVs) to enable more thorough exploration of under-sampled regions and to reduce confinement within local minima/maxima of energy and uncertainty landscapes. Finally, we propose a graph-based method to analyze structural variances in amorphous bulk systems that could be difficult to capture using conventional CVs, and yet can provide physical insights to explain the macroscopic properties of the materials. Overall, the methodologies proposed in this thesis improve the robustness and applicability of NNIPs in atomistic simulations and provide a groundwork for further advancements in this space. |
|---|