(45d) Trends in Catalyst Stability and Reactivity: Extracting Physical Insights Using Simple Data-Driven Approaches

The catalytic turnovers of a typical reaction pathway are determined by the energies of between 10 to 100 reaction intermediates.^[1] Discovering optimally performing catalysts entails maximizing catalytic turnovers across vast spaces of energy and materials. One way to simplify this formidable optimization problem is to parametrize the reaction mechanism using energy correlations.^[2] These energy correlations, often linear functions, are mappings between binding energies of reaction intermediates (e.g. C_xH_yO_z*) and one to two catalytic descriptors (e.g. binding energies of CH* and O*). Such energy correlations have enabled the high-throughput screening of metals, oxides, and zeolites. The slopes of these linear correlations (e.g. 0.5 for correlations between OH* and O*) across metals, oxides, and zeolites are determined by the ratios of bond orders of the corresponding species. Recent advances notwithstanding, these correlations are limited by three key features. First, since these correlations do not include catalyst stability metrics, they are unable to describe catalyst dynamics under reaction conditions. Second, within the existing formalism it is difficult to reverse-engineer active sites through property (e.g. rate) → structure (e.g. chemical composition) “inverse design” approaches. Third, their linearity, together with rigid constraints on the slope, imposes severe limitations in the accessible regions of materials space. There has been a concerted effort in the computational catalysis community to bridge these three knowledge gaps using data-driven statistical learning models.^[3]-[5] Herein, we illustrate our contributions towards reverse-engineering realistic active ensembles sites that possess targeted catalytic turnovers.^[6]-[9] We place a special emphasis on extracting physical insights from regressed parameters within our simple data-driven approach.^{[6], [10]}

In the first part, we present accelerated schemes for determining the energies of metal atoms that, in turn, influence catalyst stability.^[6] Predicting energies of metal atoms with atomic site specificity enables expeditious evaluations of surface energies, cohesive energies of nanoparticles having generic shapes, and binding energies of metal adsorption sites. We calculate the energies of metal atoms by melding quadratic and linear interpolations across the space of coordination numbers, composition, and external strain.^[7] These new interpolations shrink the training set size for 12 fcc p- and d-block metals from > 204 to as few as 24 Density Functional Theory (DFT) calculations, without sacrificing accuracy on the test set.^[6] We will specifically highlight physical insights regarding metal-metal bonding that are obtained from the regressed parameters. As a proof of concept, we employ this method to reverse engineer thermodynamically stable active site motifs of nanoparticles that catalyze C₁ conversion.

In the second part, we investigate the existence of energy correlations on supported nanoparticles using Au nanorods on doped MgO supports as a prototypical model.^[10] Using > 1700 DFT calculated binding energies, we show that although DFT-derived energy correlations between 31 reaction intermediates of CO oxidation, water gas shift, and methanol synthesis do exist, their slopes depart from bond order conservation constraints. This departure from bond order conservation at bifunctional metal/oxide interfaces enhances the flexibility in materials space during catalyst optimization. We build a conceptual framework using the multipole expansion in electrostatics to semi-quantitatively explain trends within the slopes. Through these trends, we elucidate the electrostatic origin of the departure in slopes from bond order conservation.

Taken together, these two studies reveal new functional forms for energy correlations that determine metal nanoparticle stabilities, and reactivity trends at bifunctional interfaces. More crucially, we show that parameters of simple data-driven models contain a wealth of intriguing physical insights that enhance model interpretability.

References

1. Norskov et al., Proc. Natal. Acad. Sci. USA, (2011), 108, 937.
2. Abild-Pedersen et al., Phys. Rev. Lett. (2007), 99, 016105.
3. Tran et al., Nature Catal. (2018), 1, 696.
4. O’Connor et al., Nature Catal. (2018), 1, 531.
5. Goldsmith et al., AICHE J. (2018), 64, 2311.
6. Choksi, Streibel, and Abild-Pedersen., J. Chem. Phys. (2020), 152, 094702.
7. Streibel, Choksi and Abild-Pedersen. J. Chem. Phys. (2020), 152, 094701.
8. Choksi et al. J. Phys. Chem. Lett. (2019), 10, 1852.
9. Roling, Choksi, and Abild-Pedersen. Nanoscale (2019), 11, 4438.
10. Choksi et al. Angew. Chem. Int. Ed. (2018), 57, 15410.