Machine Learning Interatomic Potentials for Reactive Hydrogen Dynamics at Metal Surfaces Based on Iterative Refinement of Reaction Probabilities

Abstract

The reactive chemistry of molecular hydrogen at surfaces, notably dissociative sticking and hydrogen evolution, plays a crucial role in energy storage and fuel cells. Theoretical studies can help to decipher underlying mechanisms and reaction design, but studying dynamics at surfaces is computationally challenging due to the complex electronic structure at interfaces and the high sensitivity of dynamics to reaction barriers. In addition, ab initio molecular dynamics, based on density functional theory, is too computationally demanding to accurately predict reactive sticking or desorption probabilities, as it requires averaging over tens of thousands of initial conditions. High-dimensional machine learning-based interatomic potentials are starting to be more commonly used in gas-surface dynamics, yet robust approaches to generate reliable training data and assess how model uncertainty affects the prediction of dynamic observables are not well established. Here, we employ ensemble learning to adaptively generate training data while assessing model performance with full uncertainty quantification (UQ) for reaction probabilities of hydrogen scattering on different copper facets. We use this approach to investigate the performance of two message-passing neural networks, SchNet and PaiNN. Ensemble-based UQ and iterative refinement allow us to expose the shortcomings of the invariant pairwise-distance-based feature representation in the SchNet model for gas-surface dynamics.

Benchmarking of machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces

“…This involved performing dynamics simulations, picking structures that are not well represented by our models trained on the current database, adding them, and retraining the models. More details about the database generation can be found in reference [40]. As the training database was generated with SchNet, none of the models used in this study (PaiNN, MACE, ACE, REANN) have been involved in the sampling of the dataset.…”

“…For DFT calculations we employed a specific reaction parameter (SRP) functional [44] containing 52% of PBE [45] and 48% of RPBE functional [46] (SRP48), which is known to correctly predict the dissociation barrier. [40,42,[47][48][49][50][51]. We employed a k grid of 12×12×1 and a 'tight' default basis set within the FHI-aims [52] all-electron electronic structure code.…”

“…Recently, Stark et al [40] have performed an adaptive sampling of training data based on the widely-used specific reaction parameter (SRP)-based functional [41,42] to construct MLIPs for H 2 chemistry on various facets of copper. Specifically, they have compared the difference in the performance of an equivariant (PaiNN) and a non-equivariant MPNN (SchNet) for the simulation of reactive scattering probabilities, with PaiNN MLIPs outperforming SchNet-based MLIPs in accuracy and data efficiency.…”

SpaiNN: Equivariant Message Passing for Excited-State Nonadiabatic Molecular Dynamics

“…These models have demonstrated improved accuracy and efficiency compared to traditional ML models by successfully capturing the underlying symmetries of molecules and exploiting them to achieve accurate predictions even with limited training data. [25][26][27][28][29] However, to the best of our knowledge, only one study has applied equivariant ML to describing excited state properties. Gómez-Bombarelli and co-workers developed a diabatic artificial neural network using the equivariant PaiNN 30 model, achieving a six-fold speedup in photodynamics simulations for azobenzene derivatives.…”

Virtual Special Issue on Machine Learning in Physical Chemistry Volume 2

“…In JPC C , the special issue papers describe ML and other data science techniques utilized in the scope of nanoparticles and nanostructures; surface and interface processes; electron, ion, and thermal transport; optic, electronic, and optoelectronic materials; and catalysts and catalysis; as well as energy conversion and storage materials and processes. As in the Parts A and B, a number of articles directly address either the development of new methods or the use of ML methods in new ways. − Several contributions relate to methods in the use or development of so-called machine learning potentials (MLP) or machine learning interaction potentials (MLIP), including a Perspective on improving these models authored by Maxson et al, as well as other contributions. − Interfaces and related phenomena are addressed in several articles. − Nanomaterials and their properties are also prominently featured. − Another large category in JPC C includes studies that address the calculation of properties of materials for wide-ranging applications. − …”