Coordination-layer engineering of Pt single atoms on h-BN for propane dehydrogenation via a p-band descriptor validated by multiscale modeling and knowledge-informed machine learning

Abstract

Single-atom catalysts promise maximal metal and site efficiency, but predictable bidirectional electronic control remains elusive because most anchoring strategies enforce monotonic charge transfer. The rational design of single-atom catalysts (SACs) demands predictive electronic descriptors that quantitatively link the coordination environment to catalytic activity. Here, we establish the p-band center of the first coordination shell (ε_p) as a universal descriptor for electronically tuning Pt single atoms supported on hexagonal boron nitride (h-BN). Through density functional theory calculations, microkinetic modeling, a neural network, and SISSO symbolic regression, we uncover a striking bipolar regulation mechanism: carbon doping in the coordination sphere monotonically passivates cationic Pt sites at boron vacancies (Pt/V_B) while systematically activating anionic Pt sites at nitrogen vacancies (Pt/V_N). This seemingly paradoxical behavior originates from the context-dependent electronic duality of carbon, which acts as an electron donor relative to nitrogen but as an acceptor relative to boron. This dual behaviour drives ε_p in opposite directions across the two vacancy families. The ε_p descriptor governs the p–d hybridization strength, which in turn controls transition-state stabilization through metal–support covalent bonding, establishing a mechanistic chain: deep p-band, strong p–d hybridization, enhanced Pt–support bonding in the transition state, and a low activation barrier. Reactor-scale simulations translate this electronic optimization into a ∼1.8-fold difference in propylene yield. To rigorously validate the descriptor transferability under sparse, temperature-resolved data, we introduce PhysFormer, a physics-informed neural network that factorizes intrinsic activity and temperature response and SISSO symbolic regression, which identifies ε_p as an irreducible component of the structure–activity relationship in both vacancy families. This work establishes a descriptor-driven and physically constrained learning framework for programmable SAC design via coordination-layer engineering.