Integrating Multi-Task Surrogate Modelling with Reinforcement Learning for Autonomous CO2 Reduction Catalyst Discovery: A Proof-of-Concept Framework

<jats:p>Computational design of heterogeneous catalysts for CO2 reduction remains constrained by the prohibitive cost of density functional theory (DFT) calculations and the vast compositional search space of transition-metal alloys. We present a proof-of-concept framework that integrates three AI components: (i) a multi-task neural network surrogate trained on physics-informed synthetic data that simultaneously predicts activation energy (R2 = 0.984), turnover frequency (R2 = 0.984), selectivity (R2 = 0.974), and stability (R2 = 0.559); (ii) a proximal policy optimisation (PPO) agent that autonomously navigates an 18-dimensional action space of metal–facet–alloy compo-sitions, identifying Fe(111) as the top-ranked catalyst within the synthetic model (TOF = 103.93 s−1, selectivity = 0.72); and (iii) an explainability pipeline (SHAP, learned attention, counterfactuals) that recovers the dband centre as the dominant descriptor—providing an internal-consistency check on the physics encoded in the data generator. All synthetic targets are grounded in established computational catalysis models: Sabatierprinciple volcano curves, Brønsted–Evans–Polanyi rela-tions, and a dual-volcano competitive-kinetics selectivity model for CO2RR versus HER. We discuss the frame-work’s limitations—including the discrepancy between our synthetic prediction and experimental Faradaic-efficiency trends—with full transparency and propose a path toward DFT-validated deployment.</jats:p>