A Self-Improvable Generative AI Platform for the Discovery of Solid Polymer Electrolytes with High Conductivity

This study presents a novel generative AI-based approach for the efficient discovery and design of high-performance solid polymer electrolytes, crucial for next-generation battery technologies. The platform integrates three core components: a conditioned generative model, a validation module, and a feedback mechanism, creating a self-improving system for material innovation. To optimize the conditional generative model, we have compared different architectures including both the GPT-based and diffusion-based models, and employed pertaining and fine-tuning training strategies to achieve faster convergence and superior data efficiency. The implementation of validation from full-atom MD simulation and the feedback mechanism allows the platform to refine the output iteratively. We demonstrate that this platform facilitates the generation of hundreds of thousands of novel, diverse, valid, and synthesizable polymers, and more importantly, the effectiveness of the platform is underscored by the identification of 19 polymer repeating units, each displaying a computed ionic conductivity surpassing that of Polyethylene Oxide (PEO).