Developing Magnetic and Nonmagnetic Machine-Learned Interatomic Potentials for Gold-Promoted Nickel Catalysts

As a result of fossil fuels being burned to produce energy, the emission of greenhouse gases has increased and are the leading cause for climate change. An alternative to this practice that shows promise in the transportation section is the use of hydrogen alkaline fuel cells, a clean and renewable energy source. Despite this, the widespread use of these fuel cells is hindered by its reliance on scarce and expensive noble metal catalysts (e.g. Pt). Previous work indicates that sustainable and cheaper Au-promoted Ni-based catalysts show promising catalytic activity for the hydrogen fuel cell. As catalytic reactions occur at the nano-scale, experiments cannot fully resolve key information about the working catalyst, such as atomic-level surface reconstruction and active site formation. To tackle this issue requires the use of atomistic molecular dynamics to uncover the mechanisms of this evolving catalytic activity. However, two challenges exist: (1) current computational modeling techniques are unable to reach the necessary length- and time-scales and (2) due to nickel displaying ferromagnetic properties, there is an open question on how magnetism affects model accuracy. To address these challenges and be able to capture the multiscale performance and structure of NiAu catalysts operating within hydrogen fuel cells, we developed two machine learned interatomic potentials (ML-IAPs) for NiAu catalysts. These two potentials differ in that one is magnetic while the other is nonmagnetic. This was done to assess the impact magnetism has when predicting accurate catalytic structures of NiAu. The methodology for building an accurate machine learned interatomic potential first involved the development of training data consisting of both domain and beyond domain knowledge structures. The acquired training data was then converted into descriptors using the Fit Spectral Neighborhood Analysis Potential (FitSNAP) software. The hyperparameters and group weights were optimized using a genetic algorithm called the design analysis kit for optimization and tera-scale applications (Dakota). The goal of the Dakota optimization was to minimize prediction errors for the energies, forces, and a set of known objective functions (e.g. vacancy energy). Once the potential was deemed optimized it was then applied to various surface and nanoparticle NiAu structures to run molecular dynamics simulations at varied temperatures using Large-scale Atomic/Molecular Massively Parallel Simulator (LAMMPS). The current magnetic and nonmagnetic potentials demonstrated low errors and successfully modeled the structures mentioned earlier. Both potentials were applied to surfaces of pure Ni and Au at the (100), (110), and (111) facets, displaying stability and no anomalies at a temperature of 300 K. Additionally, the magnetic potential was also applied to a NiAu nanoparticle and was tested at two separate temperatures of 300 K and 1000 K. During these tests on the nanoparticle, the structure maintained its overall shape and did not display irregularities. These results show that both potentials were able to accurately predict the provided configurations and can further be tested on different structures and reaction conditions. Overall, this study signifies a promising step towards building accurate ML-IAPs for NiAu catalysts and advancing towards cost effective hydrogen fuel cells.