Machine Learning-Enabled Exploration of the Electrochemical Stability of Real-Scale Metallic Nanoparticles

Abstract

Surface Pourbaix diagrams are critical to understanding the stability of nanomaterials in electrochemical environments. Their construction based on density functional theory (DFT) is, however, prohibitively expensive for real-scale systems, such as several nanometer-size nanoparticles (NPs) involving at least thousands of noble metal atoms, and this limitation calls for machine learning (ML)-driven approaches. Herein, with the aim of accelerating the accurate prediction of adsorption energies for a wide range of surface coverages on large-size NPs, we developed a bond-type embedded crystal graph convolutional neural network (BE-CGCNN) model in which four bonding types were treated differently. Owing to the much enhanced accuracy of the bond-type embedding approach compared to the original CGCNN, we demonstrate the construction of reliable Pourbaix diagrams for very large-size NPs involving up to 6,525 atoms (approximately 4.8 nm in diameter), which enables the exploration of electrochemical stability over various NP sizes and shapes. We reveal that ML-based Pourbaix diagrams well reproduce the experimental observations with increasing NP size, such as the increasing O- to OH-covered phase ratio and the decreasing Pt dissolution phase in the diagrams. This work suggests a new method for accelerated Pourbaix diagram construction for real-scale and arbitrarily shaped NPs, which would significantly open up an avenue for electrochemical stability studies.

Graph neural network with classification accelerating the discovery of Heusler catalysts for nitrogen reduction reaction

“…The SGCNN 27 predicts adsorption energy by bulk and slab graphs. Since the adsorbate and its neighboring play a crucial role in adsorption energy prediction 27 and are focused by many GNNs [28][29][30][31] , we abandoned the bulk graph and instead attempted to incorporate another graph representing adsorbate and adsorption site named as adsorbate-site graph (Fig. 3a).…”

“…5). Bonds are categorized into three types, i.e., covalent bond within adsorbate, interaction between substrate, and chemisorption between surface and molecule 29 . The distance to the adsorbate is defined as the maximum of the shortest distances to the adsorbate node from the two nodes corresponding to the edge.…”

“…This allows them to effectively capture structural features and outperform the conventional methods in cheminformatics 26 . Various models, such as SGCNN 27 , LS-CGCNN 28 , BE-CGCNN 29 , LEPool-DimNet++ 30 , ACE-GCN 31 , GemNetOC 32 , and MT-MD DimNet++ 26 , are proposed to accurately predict the adsorption energy of small molecules on metal surfaces. These efforts either aim to provide a more precise description of geometric structures, predict adsorption energy at different coverage, focus on the local structure around adsorbed molecules, or seek ways to enhance the predictive capabilities of adsorption energies by incorporating information from relaxed structures.…”

The influence of perovskite crystal structure on its stability

“…11,12 Graph neural networks (GNNs) are oen used to model chemical molecular data due to their unique adaptability to unstructured data. [13][14][15][16][17][18] However, in our study, we employed a model based on three-dimensional convolutional neural networks (3DCNN) to establish the predictive relationship between perovskite adsorption models and adsorption energies. Prior research has extensively explored the application of 3DCNN in the chemical domain, [19][20][21][22] and we further demonstrated the applicability of 3DCNN by enhancing its tting and generalization capabilities through the inclusion of coordinate attention 23 mechanism and the mixture-of-experts (MMOE) framework.…”

Machine-learning prediction of facet-dependent CO coverage on Cu electrocatalysts

“…Various methods have been proposed to address this, including cluster expansion, multi-order lateral interaction models, graph theory, and machine learning approaches. 5,[35][36][37][38][39][40][41][42][43][44][45][46][47][48][49][50][51] The combination of graph theory and machine learning has been regarded as one of the most effective means to analyze coveragedependent adsorption energies: graph theory tools for automating the enumeration of vast adsorption configurations, and machine learning models for predicting adsorption energies across the entire configuration space based on a limited DFT dataset. 44,[47][48]52 However, graphbased enumeration algorithms encounter significant computational bottlenecks at the stage of isomorphism comparison of configurations, as graph isomorphism comparison is an extremely time-consuming NP problem, exponentially growing with the number of atoms in adsorption configurations.…”