![Center for Computational Sciences](../../assets/md5images/84cd06ec17d4a85a36bb45fb475ccdf3.png)
Research Experiences for Undergraduates in Computational Methods with Applications in Materials Science
Major
Computer Science
College
James Worth Bagley College of Engineering
Research Mentor
Doyl Dickel
Research Mentor's Department
Department of Mechanical Engineering
Research Center
Center for Computational Sciences
Abstract
Poster created as part of the Center for Computational Sciences' Research Experiences for Undergraduates in Computational Methods with Applications in Materials Science and presented at the 2024 Undergraduate Research Showcase.
The development of interatomic potentials via machine learning on first principles results is driven by the need to accurately and quickly analyze atomic environments. When encoding atomic environments for neural networks, a series of different two and three-body functions are used to transcribe the local environments to produce a vector of inputs, making it imperative that the local environments are encoded optimally. Given a set of atomic environments, we generate a series of different fingerprints and by analysis of the data select the most optimal ones which demonstrate a reduced validation error and computation time in our model. To select the best fingerprints for the network, we first construct a large dense matrix A, where each row is an atomic environment and each column is a fingerprint with a different set of parameters. We look to create the most optimal rank k approximation of A, by factoring it as A = CUR where C is made up of k columns of A, R is made up of k rows of A and U is square and of rank no greater than k. We employ the singular value decomposition followed by two rounds of sampling the column space of our matrix, where we utilize weights based on spectral properties and then adaptively compute leverage scores to select fingerprints that were not initially chosen and that require a low computational cost with high probability. By computing the matrices C and R we find the optimal representation of our fingerprint and training data respectively. Our optimal fingerprints achieve on average a 39% reduction in validation error with comparable training errors to naive fingerprint selections for unary and binary systems, while minimizing the computational cost for fingerprint generation.
Presentation Date
Summer 8-2-2024
Keywords
machine learning, interatomic potentials
Recommended Citation
Evans-Cole, Spencer; Barrett, Kip; and Dickel, Doyl, "Validation and Runtime Improvements in Neural Networks for Interatomic Potentials via Automatic
Fingerprint Selection" (2024). Research Experiences for Undergraduates in Computational Methods with Applications in Materials Science. 7.
https://scholarsjunction.msstate.edu/ccs-reu/7