Multi-fidelity Deep Learning And Active Learning For Molecular Optical Properties

Greenman, Kevin P.

Permalink

https://hdl.handle.net/2142/116660 Copy

Description

Title

Multi-fidelity Deep Learning And Active Learning For Molecular Optical Properties

Author(s)

Greenman, Kevin P.

Contributor(s)

• Gómez-Bombarelli, Rafael
• Green, William H.

Issue Date

2022-06-20

Keyword(s)

Mini-symposium: Machine Learning

Abstract

A variety of physics-based and statistical methods have been developed to guide molecular design based on optical properties. Each method has a trade-off between cost, accuracy, and generalizability. While methods such as time-dependent density functional theory (TD-DFT) are often generalizable across chemical space due to their foundations in physics, they are relatively slow and are less suitable for screening large libraries of molecules. Statistical or machine learning methods are fast, but their performance is highly dependent on the choice of training data and representation. This makes them useful for design within chemical families that already have large datasets available, but less useful for de novo design tasks that explore new parts of chemical space. We propose a new deep learning method that leverages a combination of low fidelity (TD-DFT) and high fidelity (experimental) data sets to predict molecular optical properties with improved accuracy and generalizability over existing statistical methods. We also illustrate the importance of non-random data splitting strategies to assess generalizability of predictions for spectra in condensed phase. Finally, we demonstrate the use of active learning for model improvement by gathering new experimental data in regions of high prediction uncertainty.

Publisher

International Symposium on Molecular Spectroscopy

Type of Resource

text

Language

eng

Handle URL

https://hdl.handle.net/2142/116660

DOI

https://doi.org/10.15278/isms.2022.MK03

Copyright and License Information

Copyright 2022 held by the authors

Owning Collections