Ab Initio to Activity: Machine Learning-Assisted Optimization of High-Entropy Alloy Catalytic Activity

Abstract

High-entropy alloys are slowly making their debut as a platform for catalyst discovery, but conventional methods, theoretical as well as experimental, may fall short of screening the vast composition space inhabited by this class of materials. New theoretical approaches are needed to gauge the catalytic activity of high-entropy alloys and optimize the alloy composition within a feasible time frame as a prerequisite for further experimental studies. Herein, we establish a workflow for simulations of catalysis on high-entropy alloy surfaces. For each step of the modeling we present our choice of method, however, we also acknowledge that alternative options are available. We apply the developed methodology to predict the net catalytic activity of any alloy composition, within the composition space spanned by Ag-Ir-Pd-Pt-Ru, for the oxygen reduction reaction. Based on first-principle calculations, a graph convolution neural network is used to predict adsorption energies of *OH and *O. Subsequently, taking competitive co-adsorption of reaction intermediates into account, we couple the net adsorption energy distribution of a high-entropy alloy surface to the expected current density. Lastly, this procedure is used in conjunction with a Bayesian optimization scheme to search for optimal alloy compositions, which yields several promising compositions. This result shows that an unbiased in silico pre-screening and discovery of catalyst candidates is viable and will help scale the otherwise insurmountable challenge of searching for high-entropy alloy catalysts. It is our hope that our computational framework, which is freely available on GitHub, will aid other research groups to efficiently identify promising high-entropy alloy catalysts.

A Flexible Theory for Catalysis: Learning Alkaline Oxygen Reduction on Complex Solid Solutions within the Ag-Pd-Pt-Ru Composition Space

“…To address this, we employed an adsorption energy inference model in the form of an earlier published graph neural network (GNN) with a lean architecture [31] . This inference model was trained using multiple datasets of DFT calculations of *OH and *O on an fcc(111) surface with broadly sampled compositions within the Ag-Ir-Pd-Pt-Ru composition space, as this was readily available to us from the earlier work [31] . The composite dataset also included several suballoys, e.g., binary, ternary, and quaternary combinations of Ag-Ir-Pd-Pt-Ru.…”

“…This constituted the gross adsorption energy distribution for that particular alloy composition. Subsequently, we simulated the occupation of binding sites according to the lowest adsorption energy and mimic inter-adsorbate blocking, which have also been described in detail elsewhere [31] . From this simulation and the resulting adsorbate coverage, we obtain the net adsorption energy distribution of *OH and *O.…”

“…In previous work [28,29,31] , we have applied a theory-derived model to the adsorption energies of *OH and *O based on the Butler-Volmer and Koutecký-Levich equations: applies. is an adjustable parameter to fit the scale of the experiments and thus have a unit of mA times number 𝑎 of sites per cm 2 .…”

Demystifying the Chemical Ordering of Multimetallic Nanoparticles

“…These multimetallic NPs have been applied to a plethora of catalytic applications including ammonia decomposition where they show high stability and an increased reaction rate, almost 4-fold, in comparison to ruthenium (Ru) catalyst (at 500 °C) . Toward understanding the stability of multimetallic systems, Clausen et al developed a computational methodology for screening different alloy compositions to find optimal catalysts through the use of first-principles-based graph neural networks and Bayesian optimization . This method was applied to the ORR with an alloy composition space over the HEA Ag–Ir–Pd–Pt–Ru.…”

“… 33 Toward understanding the stability of multimetallic systems, Clausen et al developed a computational methodology for screening different alloy compositions to find optimal catalysts through the use of first-principles-based graph neural networks and Bayesian optimization. 34 This method was applied to the ORR with an alloy composition space over the HEA Ag–Ir–Pd–Pt–Ru.…”