Generative Design of Polymer Formulations via Conditional Graph Diffusion with Feasibility Filters
DOI:
https://doi.org/10.71465/csb180Keywords:
Graph Neural Networks, Denoising Diffusion Probabilistic Models, Polymer Informatics, Inverse Material DesignAbstract
The inverse design of polymer formulations—identifying molecular structures and mixture ratios that satisfy specific property constraints—remains a grand challenge in materials informatics. Traditional combinatorial chemistry and high-throughput screening are often constrained by the vastness of the chemical space and the non-linear structure-property relationships inherent in macromolecular systems. In this work, we propose a novel generative framework based on Conditional Graph Diffusion with Feasibility Filters (CGD-FF). Unlike conventional variational autoencoders or generative adversarial networks, which often suffer from mode collapse or generate thermodynamically unstable structures, our approach models the generation process as a reverse diffusion process on molecular graphs. We introduce a conditional guidance mechanism that directs the denoising trajectory towards regions of the latent space associated with desired target properties, such as glass transition temperature and tensile strength. Furthermore, we integrate a physics-based feasibility filter that evaluates synthetic accessibility and thermodynamic miscibility in real-time during the generation phase. Our experimental results demonstrate that CGD-FF outperforms state-of-the-art baselines in generating valid, unique, and novel polymer structures while adhering to strict property targets. The integration of domain knowledge via feasibility filters significantly reduces the generation of chemically invalid candidates, bridging the gap between computational design and experimental realizability.
Downloads
Published
Issue
Section
License
Copyright (c) 2025 Elizabeth Moore

This work is licensed under a Creative Commons Attribution 4.0 International License.
All articles published in the Computer Science Bulletin are distributed under an open access license, allowing unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited. Authors retain copyright and grant the journal the right of first publication. The journal encourages knowledge sharing while maintaining academic integrity and proper attribution.
