Green hydrogen is one of the most promising candidates for replacing current carbon-based energy sources. From 2021 to 2024, H2 production from renewable sources grew by 50%. The introduction of H2 into the existing gas distribution and storage network is uncertain because the structural properties of materials degrade considerably when exposed to hydrogen atmosphere. H2 embrittlement is a complex metallurgical phenomenon that combines physical and chemical processes, where several different mechanisms act simultaneously to cause the material to collapse. Machine-learning algorithms are being used to predict the severity of fatigue caused by metal contact with H2. Reliable databases are one of the most critical elements for training a machine-learning model. The Technical Database for Hydrogen Compatibility of Materials maintained by Sandia National Laboratories is a very useful data source. It is a repository of test results conducted in hydrogen environments. Several descriptors are available, such as hydrogen purity, pressure, temperature, test method, stress ratio, and frequency, material chemical composition, microstructure, processing technique, yield and tensile strengths, elongation, and reduction of area. An analysis of the database showed a high frequency of incomplete data in mechanical tests as well as a lack of standardization of these tests. In this study, we use numerical methods of molecular simulation to obtain stress-strain curves with different H2 concentrations and to simultaneously measure hydrogen diffusivity. Our goal is to provide first insights into how molecular dynamics methods can be used to fill gaps in existing databases and to propose new, uncatalogued descriptors, such as diffusivity. We use classical and reactive force fields (ReaxFF) to measure diffusivity and stress-strain curves at different H2 concentrations in Fe (alpha), Fe₃C, and Ni lattices. The results showed that the diffusivity of alpha iron showed a difference of one order of magnitude between the two types of force field, while the mechanical strength values were comparable. All materials showed a reduction in mechanical strength as the percentages of H2 in the lattice increased from 0 to 5%. Tests introducing the descriptors obtained from the molecular simulation calculations into a neural network model showed an increase in the R2 value and a reduction in the number of descriptors. The feasibility of using the descriptors obtained by molecular simulation allows for considerable savings in time and resources, since the experiments depend on extremely expensive equipment that demands a high workload.
