Machine Learning for Metallurgy: A Neural-Network Potential for Zirconium
For an accurate description of metals and their alloys it is mandatory to control their performance and predict their long-term stability. This includes knowledge about defect structures as well as fracture behavior of the material. Since the large simulation cells required by the accurate description of complex defects are beyond the accessibility of ab initio calculations as based on density functional theory, machine-learned potentials (MLPs) have emerged as a powerful approach in this field.
In a recent collaboration, Manura Liyanage and William Curtin of #EPFL and David Reith and Volker Eyert of Materials Design have developed a neural-network potential (NNP) for zirconium, which has proven successful in describing a variety of important properties of this metal. Zr and its alloy are extensively used in the nuclear industry due to its low neutron absorption cross section, good mechanical properties, and corrosion resistance. Generation of the neural-network potential was preceded by ab initio calculations for a large training set including hydrostatically and uniaxially strained structures of the different crystalline phases of Zr, vacancy structures, surface slabs, snapshots from ab initio MD simulations as well as structures with stacking faults and self-interstitials.
The resulting NNP was able to successfully describe dislocation structures and their relative energies and fracture behavior, along with bulk, surface, and point-defect properties and structures, and significantly outperforms the best available traditional potentials. This allows to perform large scale simulations of complex processes of zirconium.