Application of Machine Learning in Chemical Space: Machine Learning Methods Applied to Autocorrelation Vectors Generated from the Vaska’s Space Dataset.
Year: 2023
Type: dissertation
Abstract: The field of chemistry is increasingly applying machine learning methods to aid in the exploration and understanding of the chemical space. This thesis explores a set of supervised machine learning (ML) methods consisting of multi-layer perceptron (MLP) models, Gaussian processes (GPs), multi-layer perceptrons models with a Gaussian process at the output layer (FEGP), and deep Gaussian processes (DGPs) and their performance against the predictions of the DFT-calculated activation barrier associated with the activation of dihydrogen and the H – H bond distance in the transition state to a constructed chemical space surrounding Vaska’s complex. From graph representations of the molecules in this space autocorrelation vectors were provided using generic and electronic property sets for the atoms and bonds in the graph representation. Gradient boosting regressors (GBR) were employed to assess the relevance of each descriptor in the autocorrelation vectors with great success, increasing the interpretability of, in addition to reducing the computational costs associated with, the ML methods. For the prediction of the activation barrier the GP model yielded the lowest mean absolute errors on test sets of 0.75 ± 0.03 kcal/mol, and for the prediction of the H – H distance of the transition state we found that the GBR method outperformed the ML methods, with a mean absolute error of 0.0161 Å.
Author: Betten, Jørn Eirik
Language: English
Subfield: Materials Chemistry
Field: Materials Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7026767270