Machine learning transferable physics-based force fields using graph convolutional neural networks
Name
1227031771-MIT.pdf
Size
1.4 MB
Format
Adobe PDF
Checksum (MD5)
3c59fa8c0cf3ac210d6bb2b4202a3e41
Author(s)
Harris, William H.(William Hunt)
Advisor(s)
Rafael Gomez-Bombarelli.
Date Issued
2020
Publisher
Massachusetts Institute of Technology
Abstract
Molecular dynamics and Monte Carlo methods allow the properties of a system to be determined from its potential energy surface (PES). In the domain of crystalline materials, the PES is needed for electronic structure calculations, critical for modeling semiconductors, optical, and energy-storage materials. While first principles techniques can be used to obtain the PES to high accuracy, their computational complexity limits applications to small systems and short timescales. In practice, the PES must be approximated using a computationally cheaper functional form. Classical force field (CFF) approaches simply define the PES as a sum over independent energy contributions. Commonly included terms include bonded (pair, angle, dihedral, etc.) and non bonded (van der Waals, Coulomb, etc.) interactions, while more recent CFFs model polarizability, reactivity, and other higher-order interactions.
Simple, physically-justified functional forms are often implemented for each energy type, but this choice - and the choice of which energy terms to include in the first place - is arbitrary and often hand-tuned on a per-system basis, severely limiting PES transferability. This flexibility has complicated the quest for a universal CFF. The simplest usable CFFs are tailored to specific classes of molecules and have few parameters, so that they can be optimally parameterized using a small amount of data; however, they suffer low transferability. Highly-parameterized neural network potentials can yield predictions that are extremely accurate for the entire training set; however, they suffer over-fitting and cannot interpolate.
We develop a tool, called AuTopology, to explore the trade-offs between complexity and generalizability in fitting CFFs; focus on simple, computationally fast functions that enforce physics-based regularization and transferability; use message-passing neural networks to featurized molecular graphs and interpolate CFF parameters across chemical space; and utilize high performance computing resources to improve the efficiency of model training and usage. A universal, fast CFF would open the door to high-throughput virtual materials screening in the pursuit of novel materials with tailored properties.
Description
Thesis: S.M., Massachusetts Institute of Technology, Department of Materials Science and Engineering, 2020
Cataloged from student-submitted PDF of thesis.
Includes bibliographical references (pages 22-24).
Subjects
Materials Science and Engineering.
MIT Department
Massachusetts Institute of Technology. Department of Materials Science and Engineering
Terms of Use
MIT theses may be protected by copyright. Please reuse MIT thesis content according to the MIT Libraries Permissions Policy, which is available through the URL provided.
Persistent DSpace Link
