PolymerGPT: Multi-property Optimization with a Decoder-Based GPT Model for Generative Polymer Design

We introduce PolymerGPT, a decoder-only GPT-style transformer trained from scratch on polymer SMILES (pSMILES) representations across 1M, 10M, and 100M-compound subsets of the PolyOne100M dataset, using multi-GPU distributed training on Google Cloud (Vertex AI). The model achieves Validity, Uniqueness, and Novelty (VUN) scores above 95% across all data subsets, surpassing existing literature benchmarks, while aligning generated polymer properties closely with real-world measurements as validated through closure-set experiments. A TransPolymer property-prediction model, optimized via Ray Tune with Bayesian optimization and Hyperband scheduling, guides multi-property conditional generation. C. Pyle and A. Gadari contributed equally to this work; C. Zhu is the corresponding author.

In arXiv preprint, 2026. arXiv:2608.01431.