The selective oxidation of methane to methanol using homogeneous catalysts with earth-abundant metals remains a significant challenge, primarily due to the vast and complex chemical space that must be explored to identify promising catalysts. Trial-and-error approaches are inherently inefficient, while first-principles methods like density functional theory (DFT) become computationally prohibitive at scale. To address this, we employ data-driven strategies, leveraging active learning and efficient global optimization (EGO), to accelerate catalyst discovery while minimizing computational cost with machine learning. Inspired by the widespread use of macrocyclic ligands in catalysts studied for C–H activation, we construct a large space of Mn and Fe square pyramidal complexes comprising synthetically accessible tetradentate macrocyclic ligands in square planar and seesaw geometries, along with an aromatic monodentate ligand. We train neural network-based surrogate models on data from Fe and Mn complexes studied in prior work, incorporating epistemic uncertainty quantification via the Laplace approximation, which guides efficient DFT data acquisition from our expanded design space. We define optimal methane-to-methanol catalysts as those with high turnover numbers stemming from flatter reaction energy landscapes. Thus, we seek catalysts that readily activate methane’s strong C–H bonds while disfavoring the methyl radical rebound step which acts as a thermodynamic sink in the energy landscape. Our EGO-driven approach minimizes hydrogen atom transfer (HAT) energetics while maximizing rebound energetics, leading to the identification of optimal methane-to-methanol catalysts in previously unexplored regions of chemical spaces where established thermodynamic and kinetic scaling relations may not hold. Additionally, we uncover novel catalyst design rules for tuning C–H activation reactivity, offering promising avenues for catalyst development that could outperform current state-of-the-art catalysts.
