Multi-Constraint Molecular Generation using Sparsely Labelled Training Data for Localized High-Concentration Electrolyte Diluent Screening

TL;DR

This paper introduces ConGen, a semi-supervised molecular generative model that effectively handles sparsely labeled training data to generate molecules with multiple property constraints, demonstrated in electrolyte diluent screening.

Contribution

The work extends the semi-supervised variational auto-encoder to handle sparsely labeled data, enabling multi-constraint molecule generation from limited property labels.

Findings

ConGen outperforms existing models in multi-constraint molecule generation.

It successfully generates candidate molecules for Lithium-ion battery electrolytes.

The approach reduces the need for fully labeled datasets in molecular design.

Abstract

Recently, machine learning methods have been used to propose molecules with desired properties, which is especially useful for exploring large chemical spaces efficiently. However, these methods rely on fully labelled training data, and are not practical in situations where molecules with multiple property constraints are required. There is often insufficient training data for all those properties from publicly available databases, especially when ab-initio simulation or experimental property data is also desired for training the conditional molecular generative model. In this work, we show how to modify a semi-supervised variational auto-encoder (SSVAE) model which only works with fully labelled and fully unlabelled molecular property training data into the ConGen model, which also works on training data that have sparsely populated labels. We evaluate ConGen's performance in generating molecules with multiple constraints when trained on a dataset combined from multiple publicly available molecule property databases, and demonstrate an example application of building the virtual chemical space for potential Lithium-ion battery localized high-concentration electrolyte (LHCE) diluents.