(169cl) Generative Artificial Intelligence for Property-Guided Design of Co-Polymers

Synthetic polymers are an important class of materials that one finds in many different consumer products, ranging from energy materials, over plastics, to medical applications. There is a constant demand for novel polymer materials with improved highly specialised functionalities. Traditional efforts in discovery and development of novel materials are often time-consuming, require much expert knowledge, and do not consider the entire possible design space. One promising step towards more efficient, while at the same time broader, exploration of the chemical polymer design space is the development of generative molecular machine learning models.

Generative artificial intelligence (AI) for the design of synthetic polymers still needs to overcome domain-specific challenges. One challenge is that unlike for small molecules, synthetic polymers are governed by multiple structural levels of information that go beyond the atomic structure of monomers. Polymers are governed by monomer stoichiometries, by chain architectures, linking structures, chain length and many more. This raises the question on how to best represent polymers for machine learning algorithms. Secondly, controlled design of novel materials necessitates much (property-) labelled data, which in the field of synthetic polymers is not yet easily accessible.

We present our currently developed approaches on molecular machine learning for co-polymer design. We build upon the representation of polymers as molecular graph ensembles [1] and work on the two challenges outlined above: learning with limited labelled data and learning beyond the atomic representation of monomer units. To this end, we extend our previous Graph-to-string variation autoencoder [2] to take partly labelled-data into account and to organise its latent space based on target property information. We additionally perform Bayesian optimisation in the models latent space to identify most relevant property regions.

We illustrate our approach on a case study for polymer photocatalyst design for the production of green hydrogen. Our results show that we can sample 100% valid co-polymers with high novelty (> 90%) and diversity (> 80%) from desired property regions in the latent space [3]. Notably our model extends previous generative AI works on polymers by designing for the first time molecular ensembles; therewith including stoichiometries and chain architectures. This presents an important step in the direction of generative AI for structurally diverse material classes.

References

[1]: Aldeghi, M., & Coley, C. W. (2022). A graph representation of molecular ensembles for polymer property prediction. Chemical Science, 13(35), 10486-10498.

[2]: Vogel, G., Sortino, P., & Weber, J. M. (2023). Graph-to-String Variational Autoencoder for Synthetic Polymer Design. In AI for Accelerated Materials Design-NeurIPS 2023 Workshop.

[3]: Vogel, G., Weber, J.M. (In preparation). Inverse design of copolymers with optimized properties.