Searching for an Optimal Multi‐Metallic Alloy Catalyst by Active Learning Combined with Experiments

Abstract

However, determining the optimum elemental components and composition is challenging because of the different physical behaviors and chemical activities of the metal elements in catalytic reactions. Furthermore, it is difficult to determine the metal combination that must be investigated as it is difficult to exactly determine which metal element will affect the catalytic performance within the alloy. Before the introduction of computational techniques, researchers have mainly investigated binary and ternary alloys, with the focus on searching for high-performance multi-metallic catalysts by experimental trial and error. [8][9][10] For example, noble and non-noble metals, for example, platinum (Pt), [11][12][13] ruthenium (Ru), [14,15] iridium (Ir), [16,17] palladium (Pd), [18,19] and nickel (Ni), [20] iron (Fe), [21] cobalt (Co), [22] molybdenum (Mo), [23] as well as their alloys, are typically used in the hydrogen evolution reaction (HER), rendering the dimensionality of the space of alloy composition candidates far beyond human intuition or brute-force search by trial and error. However, it was almost impossible to search for the optimal component and composition for high-performance catalysts as the determination of the catalyst performance is a timeconsuming and expensive process due to the enormous number of candidate combinations and compositions.Searching for an optimal component and composition of multi-metallic alloy catalysts, comprising two or more elements, is one of the key issues in catalysis research. Due to the exhaustive data requirement of conventional machine-learning (ML) models and the high cost of experimental trials, current approaches rely mainly on the combination of density functional theory and ML techniques. In this study, a significant step is taken toward overcoming limitations by the interplay of experiment and active learning to effectively search for an optimal component and composition of multi-metallic alloy catalysts. The active-learning model is iteratively updated using by examining electrocatalytic performance of fabricated solid-solution nanoparticles for the hydrogen evolution reaction (HER). An optimal metal precursor composition of Pt 0.65 Ru 0.30 Ni 0.05 exhibits an HER overpotential of 54.2 mV, which is superior to that of the pure Pt catalyst. This result indicates the successful construction of the model by only utilizing the precursor mixture composition as input data, thereby improving the overpotential by searching for an optimal catalyst. This method appears to be widely applicable since it is able to determine an optimal component and composition of electrocatalyst without obvious restriction to the types of catalysts to which it can be applied.

Exploring Optimal Water Splitting Bifunctional Alloy Catalyst by Pareto Active Learning

“…Details of the fabrication of CNF substrate could be seen in the previous work. [58] Preparation of the Precursor Solution: All chemicals were purchased from Sigma-Aldrich. The precursor chemicals were as follows: chloroplatinic(IV) acid hydrate, ruthenium(III) chloride hydrate, palladium(II) chloride, cobalt(II) chloride, nickel(II) chloride, copper(II) chloride, iron(III) chloride hexahydrate, and tin(II) chloride.…”

“…Even though 10 mA cm −2 is a conventional region to measure overpotential values, the overpotential of HER was measured at 20 mA cm −2 due to the usage of previous data done by this group. [58] After the measurements, overpotential values were utilized to develop a new model by Pareto active learning, which predicted the HER and OER overpotential values (Figure 1c). The predicted HER and OER overpotential values were designated as the x-and y-axes, respectively, and plotted in a 2D data space.…”

“…Previously, our group combined active learning and CTS experiments to search for a high-performance HER catalyst. [58] During the search of an optimal component and composition of a multimetallic alloy, 98 experimental data points were generated; the use of these data would be beneficial to the overall Pareto active learning process. By contrast, data for OER were not generated.…”

Active Machine Learning for Chemical Engineers: a Bright Future Lies Ahead!

“…Synthesizability is defined as the feasibility of the proposed queries, referring to whether the proposed catalysts or molecules can be synthesized. Often, a simple representation of a catalyst is a vector containing the catalyst composition [50,68]. This guarantees the synthesizability of the catalyst but limits the design space explored by the active machine learning algorithm as only the composition is varied but no structural or geometrical properties are considered.…”

Following Paths of Maximum Catalytic Activity in the Composition Space of High‐Entropy Alloys

“…[10][11][12][13][14][15] The idea of thinking of the catalytic activity as a continuous function of the HEA composition to be optimized in an experimental context where the number of experiments must be kept as low as possible has developed recently. [16][17][18][19] It was found with simulations that the function of ORR catalytic activity for the Ag-Ir-Pd-Pt-Ru HEA composition space has a relatively long length scale of around 0.3 with respect to molar composition. [16] This means that the correlation between molar compositions is still expected to be significant even when they are spaced 30 atomic percent (at%) apart in composition space.…”