Slab Graph Convolutional Neural Network for Discovery of N2 Electroreduction Catalysts

TL;DR

This paper introduces SGCNN, a novel machine learning model based on slab graph convolutional neural networks, designed to predict surface adsorption energies for N2 electroreduction catalysts using only elemental properties, thus accelerating catalyst discovery.

Contribution

The development of SGCNN, a flexible ML model that predicts adsorption energies without ab initio inputs, enabling faster and more accessible catalyst screening for NRR.

Findings

SGCNN achieves a mean absolute error of 0.23 eV in predicting adsorption energies.

Binary intermetallics with specific d-electron occupations may enhance NRR performance.

The model leverages elemental properties, avoiding complex ab initio calculations.

Abstract

The catalyst development for N2 electroreduction reaction (NRR) with low onset potential and high Faradaic efficiency is highly desired, but remains challenging. Machine learning (ML) recently emerged as a complementary tool to accelerate material discovery; however a ML model for NRR has yet to be developed. Here, we develop and report slab-graph convolutional neural network (SGCNN), an accurate and flexible ML model that is applicable to catalytic surface reactions. With the self-accumulated database of 2,699 surface calculations, SGCNN predict binding energies, ranging over 8 eV, of five key adsorbates (*H, *N2, *N2H, *NH, *NH2) related to NRR performance with the mean-absolute-error of only 0.23eV. Unlike previously available models, SGCNN avoids using ab initio level inputs, instead is solely based on elemental properties that are all readily available in Periodic-Table-of-Elements; true accelerations can be realized. t-distributed stochastic neighbor embedding (t-SNE) analysis reveals that binary intermetallics of averaged d-electron occupation between 4 and 5 could potentially lower the onset potential in N2 electroreduction.