Inverse design of conjugated polymers

Designing polymers with targeted properties is hindered by complex, nonlinear structure–property relationships and slow iteration between synthesis and validation. Here we describe a progression in capability, from efficient optimization within a known chemical space to discovery in a vastly larger space, to accelerate inverse design of conjugated polymers for electrochromic functionality. We first established an autonomous optimization framework to navigate the large but bounded composition space of statistical copolymers. By integrating data mining, DFT-derived electronic descriptors, a copolymer machine-learning model, and AI-guided robotic synthesis, the platform rapidly identifies formulations that achieve targeted electrochromic responses. Building on this foundation, we then moved beyond optimization to de novo structure discovery, where the search space expands by orders of magnitude and the challenge shifts from "which formulation is best?" to "what new building blocks should exist at all?" To address this, we developed a generative modeling framework: a Transformer-based conditional model trained on ~600,000 MD/DFT-simulated monomer structures and guided by reinforcement learning toward color accuracy and synthetic feasibility. Rather than selecting among known candidates, the model proposes chemically viable new structures that can access targeted electrochromic states. Together, these two stages form a unified inverse-design strategy: autonomous experimentation enables rapid, data-efficient optimization, while generative modeling enables scalable exploration and discovery in the much larger space of possible chemistries. We also built an interactive data interface to support electrochromic polymer discovery and accelerate closed-loop iteration across both stages.