Simulating Hydrogen Adsorption Trends on Early Transition Metals with Neural Network Potentials

The advent of machine learning techniques like neural network potentials, which serve to effectively model complex first-principles calculations like density functional theory, have shown tremendous potential for accelerating material science and catalytic research. These neural network potentials still require high-quality training data which can be computationally expensive, yet once trained, they allow one to estimate physical and chemical properties several orders of magnitude faster than first-principles methods, all within very reasonable error ranges per application. Moreover, these neural network potentials can be used to efficiently map their own state-space – that is, to interpolate between the training data points with reliable accuracy and precision. This latter benefit is of substantial interest for those applications seeking to explore otherwise intractable state-spaces – often referred to as high-throughput computational screening, wherein thousands if not many more data points are calculated efficiently in an effort to identify both unique local states and global patters within the whole state-space. Our investigation leans on neural network potentials and high-throughput computing to search the vast state space of molecular hydrogen adsorption sites on early transition metals of various sizes and compositions systematically. We perform a thorough statistical analysis, including unique sites, adsorption energy ranges and heights, nearest neighbor tendencies, and more, all within very pragmatic computational demands. We also study the specific role of so called high-entropy alloys, which combined with adsorption studies like ours, have shown promising application for fine-tuning, or otherwise optimizing energy storage and catalysis. Finally, we put our neural network potentials to the test within molecular dynamics simulations in order to determine how well these networks can capture the expected physics of adsorption, desorption, and dissociation. It is revealed that they hold up surprisingly well, offering hope that the computational fidelity and efficiency trade-off can indeed be optimized per application. These methods and results help to move computational material science and catalysis research forward by cautiously testing the utility of neural network potentials therein, especially with regard to optimizing the otherwise intractable state-space exploration of new materials for hydrogen energy storage.