Identifying limitations in screening high-throughput photocatalytic bimetallic nanoparticles with machine-learned hydrogen adsorptions
Kirby Broderick, Eric Lopato, Brook Wander, Stefan Bernhard, John Kitchin and Zachary Ulissi, “Identifying limitations in screening high-throughput photocatalytic bimetallic nanoparticles with machine-learned hydrogen adsorptions”, Applied Catalysis B: Environmental 320, 121959 (2023). https://doi.org/10.1016/j.apcatb.2022.121959
This summary was written by a language model from the paper and has not been checked line by line. Rely on the published paper for details.
Question
Computational catalyst screening usually rests on the Sabatier principle: compute the binding energy of a key intermediate on a low-index slab and place the material on a volcano. Real catalysts, however, are often nanoparticles that form in situ with an unknown distribution of compositions, facets and sites. The authors asked how well hydrogen adsorption energies, predicted at scale by a machine-learned surrogate for DFT and aggregated over many sites and surfaces, describe a large experimental dataset of photocatalytic hydrogen evolution over bimetallic nanoparticles, and how much of the bimetallic search space such predictions can rule out (Broderick et al. 2023).
Approach
The experimental data came from the group’s parallel photoreactor, in which metal salts are reduced in situ to nanoparticles by an iridium photosensitizer, [Ir(Fmppy)2dtbbpy]PF6 (0.25 mM), with triethanolamine as sacrificial reductant in DMSO with a small amount of water, under blue LED illumination. Pairs of precursor salts were varied across rows and columns of a well plate, covering 25 metals (among them Ag, Au, Co, Fe, Ir, Ni, Os, Pd, Pt, Rh, Ru, Sn and several main-group and early transition metals), and hydrogen was tracked with colorimetric detection tape. For each well, cumulative H2 was smoothed with a Gaussian filter and the maximum rate taken; rates were averaged over replicate conditions, and the highest average across all compositions of a metal pair represented that system.
On the computational side, intermetallic bulk structures from the Materials Project (as enumerated for the Open Catalyst dataset, near or on the convex hull) were cut into slabs with Open Catalyst tools, and adsorption sites were placed with CatKit. Very large bulks (over 80 atoms) and slabs (over 100 atoms) were dropped. Hydrogen adsorption energies were predicted with GemNet, a message-passing graph neural network trained on the Open Catalyst data, after 90 relaxation steps with the same model; on validation data the predictions had a mean absolute error of about 0.1 eV. In all, about 16 million adsorption energies were compared with 5300 experiments across 84 monometallic and bimetallic systems. To build the volcano x-axis, structures were filtered by Pourbaix electrochemical stability (within 0.5 eV/atom at pH 8.5, -1.2 V and 0.6 mM metal ions), the strongest binding site on each surface was taken, and the surface whose value lay closest to -0.24 eV (roughly thermoneutral after the free energy correction of Nørskov and coworkers) represented the system.
Key findings
The most active systems all contained surfaces whose low-coverage binding energies lay at or slightly stronger than thermoneutral (between about -0.41 and -0.23 eV), and the upper envelope of activity versus binding energy formed a volcano shape. Many systems predicted to expose favorable surfaces were nonetheless inactive, so a favorable adsorption energy behaved as a necessary but not sufficient condition for activity. Systems containing at least one noble metal were on average about an order of magnitude more active than base-metal-only systems with comparable predicted binding energies; the authors suggest that active base-metal surfaces may form less often or be deactivated (oxides and sulfides have been seen on Fe and Ni particles in a similar system), or that compact 3d orbitals limit electron transfer. Au- and Ag-containing points sat above the trend, which the authors attribute to strain-enhanced binding in nanoparticles that slab calculations do not capture. Using predicted binding energies as a filter would remove roughly half of all metal combinations from an experimental screen, or about 30% of combinations where a bimetallic phase is expected to form. The authors also caution that rates were not normalized by surface area and that coverage effects, poisoning, nucleation and growth, and solvent (DMSO) effects on proton activity were not modeled.
Significance
The work is one of the first comparisons of a fully automated, multisite, machine-learned adsorption energy workflow against thousands of nanoparticle experiments. It shows that adsorption-energy screening remains useful as a coarse filter for complex in situ nanocatalysts, while making explicit that particle formation, surface expression, stability and deactivation often control observed activity. The authors argue that high-throughput, in-the-loop discovery will need additional computational tools for these phenomena, especially surface and materials stability in the reaction environment. Data and code are available from the Ulissi group’s DOEHER repository.