(582f) A Data-Driven Approach Towards Assessing the Stability and Activity of Nanoparticles Containing on Order of Thousand Atoms: A Case Study Using the Oxygen Reduction Reaction

Designing nanoparticles (NPs) with density functional theory (DFT) is typically performed using periodic slab approximations, as DFT can handle around N=200 atoms with computational effort increases exponentially with N. Hence DFT-based approaches are too computationally demanding to screen NPs of arbitrary shape, size, and composition. We employ the alloy stability model (ASM), a surrogate model parametrized by DFT, to design stable and active NPs that catalyze the oxygen reduction reaction (ORR).

By screening across NPs of arbitrary structures, our surrogate model can identify those NPs with specific structures that meet operational constraints for ORR. We first determine the cohesive energy of the nanoparticle to evaluate its overall stability. We then calculate the stability of individual metal atoms on the nanoparticle, on-the-fly. These atomic site stabilities serve as descriptors for reactivity metrics like the free energy of adsorption of hydroxyl (Δ_adsG(OH*)). Scaling relations between the oxygenates O*, OH* and OOH* are then capitalized to compute the free energy of adsorption of other oxygenates. These free energies are, in turn, used to calculate the limiting potential for oxygen reduction and estimate the current density at each active site. The microscopic reactivity metrics are averaged over the nanoparticle, yielding macroscopic metrics like the mass-averaged current density.

Using this data-driven approach, we screen across > 5000 Pt-based nanoparticles with diameters between 2 (300 atoms) and 8 nm (10000 atom) and in diverse shapes (cubo-octahedrons, octahedrons, decahedrons, icosahedrons, and Wulff constructions). Considered compositions include random alloys, core-shell, and edge-decorated architectures. We identify thermodynamically stable architectures that yield mass specific activity higher than the Pt (111) benchmark. Our workflow provides an intrinsic link between the atomistic properties of each NP and macroscopic observables, thus enabling the high-throughput screening of an ensemble of NPs with near-DFT accuracy and significantly lower computational effort.