ReSolved-Flow: Synthesis-Informed GFlowNets for Solvent-Conditioned Redox Molecular Design

Property-guided discovery of redox-active molecular materials for energy storage must identify candidates that satisfy a target property while remaining structurally diverse and practically synthesisable. We present ReSolved-Flow, the AI-guided design system, which uses a fragment-based Generative Flow Network (GFlowNet) combining a frozen solvent-conditioned redox-potential predictor, a learned synthesisability proxy, and chemically constrained BRICS assembly. The proxy predicts both retrosynthetic success and route length and is trained on 60,000 generated molecules labelled using AiZynthFinder. The GFlowNet is trained to approximate a distribution weighted by a multi-objective reward for redox alignment, predicted synthesisability, structural diversity, and molecular size. By considering route feasibility during generation, the framework is intended to reduce the number of computationally promising but practically unusable candidates. We apply it to three redox targets relevant to organic batteries in acetonitrile and water and evaluate the top-100 reward-ranked candidates for each target. These selected candidates include multiple high-reward structural families.