Abstract: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 to 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 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 uncertainty quantification 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.

What problem does this paper attempt to address?

The problems that this paper attempts to solve are as follows: On metal surfaces (especially copper surfaces), the reaction kinetics of molecular hydrogen, such as dissociative adsorption and recombination desorption, have important practical application values in fields such as energy storage and fuel cells. However, theoretical research on these processes faces many challenges. On the one hand, due to the complex electronic structure of the metal surface and the high sensitivity of the dynamics to the reaction barrier, it is very difficult to perform kinetic simulation calculations on the surface; on the other hand, although ab initio molecular dynamics (AIMD) based on density functional theory can provide accurate simulation results, its computational cost is extremely high, and it is difficult to accurately predict the adsorption or desorption probability of the reaction, because this requires averaging thousands of initial conditions. To solve these problems, this paper proposes a method based on machine - learning - based interatomic potentials (MLIPs), especially using message - passing neural networks (MPNNs), such as SchNet and PaiNN, to simulate the reaction scattering of molecular hydrogen on different copper crystal planes. Through ensemble learning and iterative data generation strategies, the author can not only adaptively generate training data, but also comprehensively evaluate the model performance and quantify the uncertainty. This method focuses in particular on how to generate reliable training data and how to evaluate the impact of model uncertainty on the prediction of dynamic observables, thus providing a new way to solve the above - mentioned computational problems. Specifically, the goals of the paper are: 1. **Improve computational efficiency**: Reduce the computational cost by using machine - learning methods while maintaining high precision. 2. **Generate reliable data sets**: Generate data sets that can accurately describe the interaction between hydrogen gas and the copper surface through adaptive sampling and uncertainty quantification. 3. **Evaluate model performance**: Compare the performance of different machine - learning models (such as SchNet and PaiNN) in predicting reaction probabilities, especially when dealing with high - dimensional systems. 4. **Reveal the limitations of existing models**: Through experimental verification, discover the shortcomings of the SchNet model based on invariant distance feature representation in gas - surface dynamics. In summary, this paper aims to overcome the limitations of traditional computational methods in simulating the reaction kinetics of hydrogen on metal surfaces through advanced machine - learning techniques, and provide more efficient and accurate tools for research in related fields.

Machine learning interatomic potentials for reactive hydrogen dynamics at metal surfaces based on iterative refinement of reaction probabilities

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

Using machine learning to go beyond potential energy surface benchmarking for chemical reactivity

Beyond potential energy surface benchmarking: a complete application of machine learning to chemical reactivity

Searching Configurations in Uncertainty Space: Active Learning of High-Dimensional Neural Network Reactive Potentials

Training Transferable Interatomic Neural Network Potentials for Reactive Chemistry: Improved Chemical Space Sampling

Reactive Dynamics and Spectroscopy of Hydrogen Transfer from Neural Network-Based Reactive Potential Energy Surfaces

Active learning of reactive Bayesian force fields: Application to heterogeneous hydrogen-platinum catalysis dynamics

Unified and transferable description of dynamics of H 2 dissociative adsorption on multiple copper surfaces via machine learning.

Training Neural Nets To Learn Reactive Potential Energy Surfaces Using Interactive Quantum Chemistry in Virtual Reality

Unified and transferable description of dynamics of H<sub>2</sub>dissociative adsorption on multiple copper surfaces<i>via</i>machine learning

Exploring the frontiers of condensed-phase chemistry with a general reactive machine learning potential

Machine learning hydrogen adsorption on nanoclusters through structural descriptors

Data Generation for Machine Learning Interatomic Potentials and Beyond

Long-range dispersion-inclusive machine learning potentials for structure search and optimization of hybrid organic-inorganic interfaces

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

Harnessing Physics-inspired Machine Learning to Design Nanocluster Catalysts for Dehydrogenating Liquid Organic Hydrogen Carriers

Including physics-informed atomization constraints in neural networks for reactive chemistry

Data-efficient modeling of catalytic reactions via enhanced sampling and on-the-fly learning of machine learning potentials

Hydrogen diffusion on Ni(100): A Combined Machine-Learning, Ring Polymer Molecular Dynamics, and Kinetic Monte Carlo Study

Micron-scale heterogeneous catalysis with Bayesian force fields from first principles and active learning