Machine learned interatomic potentials for gas-metal interactions

Abstract Developing interatomic potentials for gas-metal systems is difficult due to the wide range of chemical compositions that the potential must be able to reproduce. There is a need for these types of potentials for studying plasma-material interactions in fusion reactors where gaseous plasma species will implant in metallic reactor components. The challenges presented by these material systems make them suitable candidates for treatment by a machine learning approach, such as that of the spectral neighbor analysis potential (SNAP). However, constraining the dynamics with these more flexible potentials is difficult. In this work, we have developed a SNAP potential for W-N and W-H in order to study the material degradation due to ion implantation in tungsten. We have developed a large set of density functional theory training data spanning multiple chemical environments including gas phase, surface, bulk, and gas-metal configurations. Additional methodologies for developing training data and optimizing the potential for accurately describing fast diffusing impurity species are detailed. The SNAP potential well-reproduces key material properties relevant for modeling plasma-material interactions including defect formation energies, surface adsorption energies, dimer binding energies, and tungsten nitride formation energies. In addition to testing on static energetic properties, the SNAP potential was also used to simulate thermal and dynamic gas-metal interactions, including bulk diffusion, molecular gas adsorption isotherms, and ion implantation. The SNAP potentials are demonstrated to well-reproduce behavior in the wide range of chemical environments investigated, demonstrating the suitability of these machine learned interatomic potentials for future studies of plasma material interactions.