Multi-Objective Generative Discovery of Liquid Organic Hydrogen Carriers

An effective liquid organic hydrogen carrier (LOHC) must combine high gravimetric hydrogen capacity, low dehydrogenation enthalpy, and a melting point below the operating range in both loading states. Improving one property typically degrades another, and no established carrier satisfies all three. We couple two complementary generators to Uni-Mol predictors fine-tuned on LOHC property data: SAGE-VAE, an in-house stack-augmented, RL-guided conditional VAE, and REINVENT4. Generation is conditioned on gravimetric capacity ($\mathrm{CapH_2}$), dehydrogenation enthalpy (DE), and the melting point. The two generators explore largely disjoint regions of chemical space. Multi-stage screening across structural, physicochemical, toxicity, and density-functional-theory (DFT) filters then yields 67 candidates, 18 of them Pareto-optimal. These improve on established carriers in all three objectives with DFT-confirmed energetics, and the ranking recovers independently-studied LOHCs among its top candidates. Coupling property-conditioned generation with reliability-aware prediction lets the search reach far beyond the small set of pre-characterized compounds that limits conventional screening.