Beyond Funnel-Style High-Throughput Screening: An Expert-Guided Multi-Scale Nexus for Polynary Perovskite Electrocatalyst Exploration

Abstract

Navigating the vast compositional landscape of polynary oxide electrocatalysts is traditionally limited by funnel-style screening paradigms that treat high-fidelity data as binary gates rather than informative training signals. Here, we present an expert-guided multiscale framework for Pr–Sr–Co-Fe perovskites. By integrating atomic descriptors, experimental structural features, and mechanism-informed electronic features, this framework addresses these limitations within a descriptor-accessible polynary perovskite space. By implementing a “Nearest Grid Interpolation” strategy, the computational cost of high-fidelity density functional theory calculations is reduced by nearly 2 orders of magnitude within the search space. Two rounds of active learning surge the discovery efficiency of high-performance candidates (η@10 mA cm–2 < 370 mV) from an initial 4.4% to 60.0%. Integrated with human-in-the-loop expertise, this workflow identifies a localized synergistic optimum at Pr0.01Sr0.99Co0.66Fe0.34O3−δ (PSC0.66F), which exhibits a low overpotential of 338 mV at 10 mA cm–2. Symbolic regression reveals that the intrinsic activity is closely associated with a nexus of O 2p-band center, its bandwidth, and Fe eg orbital filling within this local region. Experimental validation confirms that this optimized electronic configuration promotes highly efficient lattice oxygen mechanism (LOM) kinetics, supported by accelerated bulk oxygen diffusion and dynamic surface reconstruction. This multiscale and expert-guided strategy provides a practical and extensible framework for descriptor-accessible complex materials systems, transforming fragmented data tiers into a cohesive, learnable predictive landscape.