(394i) Property-Guided Design of Topologically Complex Polymers Using Generative Modeling

The topology of a polymer chain can substantially influence their properties and those of derivative materials [1-2]. Establishing quantitative relationships between polymer topology and material properties remains challenging. Both experimental and computational investigations have enhanced the understanding of how polymer topology influences properties, yet such efforts are labor-intensive or computationally expensive, restricting studies to a small set of systems.

In recent years, efforts to design materials with target structural and functional properties have been increasingly enhanced by data-driven machine learning (ML) techniques [3-4]. However, successful applications of ML approaches for the design of polymers, or other soft materials, are just now emerging due to data scarcity, property complexity, and other factors. Nearly all demonstrations of ML over polymeric materials are restricted to simple linear architectures, yet the fact that polymers can possess complex architectures that dictate emergent properties is one of the motivating factors for their consideration as design constructs.

To address these limitations, we create a multi-task variational autoencoder (VAE) to generate polymers with specified topology and desired characteristics [4]. This model is developed with a dataset containing coarse-grained molecular dynamics (MD) data for 1,342 polymers. These polymers feature a variety of technologically relevant architectures such as star, comb, branch, linear, cyclic, and dendrimer structures, and cover a wide range of molecular weights. Input and encoding strategies are evaluated by training several models that aim to reconstruct the polymer topology and also perform auxiliary tasks of estimating the characteristic size (i.e., average squared radius of gyration) of the polymer and classifying its topology. Specifically, graph-derived topological descriptors (e.g., graph diameter, algebraic connectivity) and graph-explicit features are encoded and integrated to create a probabilistic latent space with physical interpretability. We leverage the generative modeling framework to produce sets of topologically diverse polymers that exhibit the same characteristic size in dilute solution. The rheological behaviors of these generated polymers, such as shear viscosity and viscoelastic moduli, are subsequently analyzed at finite concentrations. Overall, this study presents a scalable generative model for designing topologically diverse polymers with desired properties as an alternative to experiments or simulations.

[1] Khabaz, F. and Khare, R. Effect of chain architecture on the size, shape, and intrinsic viscosity of chains in polymer solutions: a molecular simulation study. The Journal of chemical physics 141 21, 214904 (2014).

[2] Wijesinghe, S., Perahia, D. and Grest, G. S. Polymer topology effects on dynamics of comb polymer melts. Macromolecules 51, 7621–7628 (2018).

[3] Gómez-Bombarelli, R. et al. Automatic chemical design using a data-driven continuous representation of molecules. ACS central science 4, 268–276 (2018).

[4] Kosuri, S. et al. Machine-assisted discovery of chondroitinase abc complexes toward sustained neural regeneration. Advanced Healthcare Materials 11, 2102101 (2022).

[5] Jiang, S., Dieng, A. B. and Webb, M. A. Property-guided generation of complex polymer topologies using variational autoencoders. ChemRxiv (2024).