Neural Network Representations for Studying Gas‐Surface Reaction Dynamics: Beyond the Born‐Oppenheimer Static Surface Approximation^†

Abstract

In the past a few years, there has been significant progress in theoretical characterizations of gas‐surface reaction dynamics at the atomic level. One of the major breakthroughs is the machine learning representations of the potential energy surfaces and related properties for molecules on metal surfaces from first‐principles, particularly neural networks based methods. In this review, we focus on recent advances of the development and applications of high‐dimensional symmetry‐preserving neural network representations in gas‐surface systems, which have enabled efficient Born‐Oppenheimer molecular dynamics simulations with inclusion of all molecular and surface degrees of freedom, as well as some nonadiabatic molecular dynamics simulations with effective treatment of hot electrons, at the density function theory level. Despite these advances, further challenges remain. More accurate electronic structure theories and more efficient machine learning (and active learning) algorithms are needed towards a more quantitative description of more complex gas‐surface reactions involving multiple surfaces and adsorbates or multiple electronic states.

First-Principles Insights into Adiabatic and Nonadiabatic Vibrational Energy-Transfer Dynamics during Molecular Scattering from Metal Surfaces: The Importance of Surface Reactivity

“…Because of the huge number of electrons in the metal, the electronic structure of gas–surface systems is typically characterized by density functional theory (DFT). On the basis of molecule–surface PESs calculated on the fly , or fitted to DFT points, , the nuclear motion for both the molecular and surface DOFs can be further determined. This adiabatic treatment allows the intramolecular vibrational energy redistribution (IVR) and/or energy exchange between molecules and surface phonons.…”

Dynamics of CO2 Dissociative Chemisorption on W(110)

“…40 The key advantage of this EANN method is that the density-like descriptors given in eq 2 scale linearly with respect to the number of neighboring atoms. 42 The final CO 2 + W(110) PES was fitted to 9418 points with both energies and forces. These data points were divided into training and test sets with the ratio of 90:10.…”

“…The EANN PES is invariant with respect to translation, rotation, and permutation . The key advantage of this EANN method is that the density-like descriptors given in eq scale linearly with respect to the number of neighboring atoms …”

Investigating the Eley–Rideal recombination of hydrogen atoms on Cu (111)viaa high-dimensional neural network potential energy surface

“…Recent advances in atomistic neural network (AtNN)-based approaches have shown great promise in constructing highdimensional PESs with surface DOFs for studying the gassurface dynamics from first principles at affordable costs. 43,44 Such PESs, once well-trained, will allow efficient calculation of a larger number of trajectories at the same level of accuracy as AIMD with merely about ten to hundred thousandth of cost. There have been successful applications of AtNN to molecular adsorption and desorption on metal surfaces.…”