Reinforcing transformer-based polymer electrolyte discovery through high-fidelity fine-tuning and application-oriented screening

The discovery of high-performance polymer electrolytes has been hindered by inefficient trial-and-error approaches. Herein, we present an integrated AI-driven framework combining GPT-2 for de novo polymer generation, RoBERTa for high-accuracy property prediction, and a multi-stage pipeline for rigorous validation. A high-fidelity dataset of ionic conductivities obtained from 200 ns molecular dynamics simulations ensured reliable fine-tuning and downstream screening within our framework. As a result, we identified polymer electrolytes with properties surpassing known benchmarks. Notably, three polymers exhibited high ionic conductivities (−3.11, −3.32, and −3.30 log₁₀ S cm⁻¹), all exceeding the best-performing material in the training set. Further oxidative stability analysis of top candidates led to the rediscovery of polycaprolactone (PCL) with an electrochemical stability window of 4.86 V. Additionally, two novel structures, a branched PEO-PPO copolymer and a fluorinated polyether, demonstrated oxidative stability of 4.58 V and 4.50 V, respectively, which aligns with established modification strategies such as branching and fluorination. The results underscore the model's ability to encode structure-property relationships and generate novel, chemistry-informed molecular variants. Overall, this work delivers a portfolio of high-performance polymers ready for experimental evaluation and provides a reliable and feasible roadmap for advancing AI-driven discovery of functional energy materials.