<br/>High-dimensional chemistry and physics problems are generally modelled by neural networks (NNs). However, NNs require large data and are generally problem-specific. An alternative approach is probabilistic modelling based on Gaussian processes (GPs), which can be made problem-agnostic, can be fully automated and can be used for Bayesian optimization. However, GPs suffer from the numerical complexity and have been limited to low-dimensional problems for applications that require high accuracy. In this talk I will illustrate that it is possible to build accurate GP models of a 51-dimensional physical system based on 5000 inputs, using fragmentation of the input space and automated model complexity enhancement. I will illustrate the algorithms by GP models of the global potential energy surfaces (PES) for molecular systems with up to 19 atoms. I will show that GP models thus constructed have generalization power, allowing one to extrapolate PESs from low energies to high energies. I will argue that this work opens the prospect for building machine-learning models that can be used to search for new physics, such as phase transitions, in high-dimensional physics and chemistry problems with unknown property landscapes by extrapolation.<br/>

Accurate Gaussian process models for high-dimensional molecular systems

ACS Fall 2020 Virtual Meeting

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.