A Molecular Generation–Validation Workflow Combining Reversible Junction Tree Reinforcement Learning and Semi-empirical Quantum Chemistry: A Case Study on Molecular Solar Thermal Fuels

Abstract

Molecular generative models can rapidly explore chemical space, but conventional metrics alone do not establish whether proposed structures possess application-relevant physical properties. To address this limitation, we developed an integrated generationvalidation framework coupling Reversible Junction Tree Reinforcement Learning (RJT- RL) with automated density-functional tight-binding (DFTB) evaluation. Azobenzenebased molecular solar thermal fuels served as a representative case study. RJT-RL was guided by maximum Tanimoto similarity to a reference dataset, while reference and generated molecules were evaluated under identical DFTB calculation conditions to obtain the energy difference between the cis and trans isomers (∆ E iso ) and the S 0 → S 1 excitation energy (∆ E ex ). As training progressed, the ∆ E iso distribution extended toward higher values, while ∆ E ex values were more concentrated within the visible-light window, yielding candidates within the selected joint screening region, although neither descriptor was included in the reward. These results show that automated semi-empirical post-generation validation can identify candidates with targeted property combinations and reveal information beyond conventional molecular generation metrics.