Insights on the robustness, sensitivity and expressivity of simple many-body potentials: application to a-Zr

International audienceComputational material science, including irradiation physics, has been relying on classical in teratomic potentials for decades: they efficiently allow to simulate complex phenomena at lengthand time scales that outrun the capacity of the most accurate ab initio approaches. Two differentclasses of interatomic potentials are currently used. The so-called machine learning potentials[1],on one hand, have a high number of parameters and thus a high flexibility to reproduce a numberof physical quantities, but a high computational cost. On the other hand, many-body interatomicpotentials, like Embedded Atom Method (EAM) and tight binding Second Moment Approximation(SMA) potentials, have a more limited accuracy due to a reduced number of parameters. However,they are based on physically-inspired functional forms which ensures a reasonable transferability.Their low computational cost thus allow studies that require multiple evaluations of system en ergy and forces[2]. The identification of the potential parameters for a given system is of crucialimportance, but challenging.In this work, we apply efficient tools of model screening and sensitivity analysis to get insightson the capabilities of two SMA potentials [3, 4], in the context of irradiated α-zirconium. We focuson properties that are relevant for irradiation, i.e. elastic constants, point defect properties andstacking faults energies. The local screening of the potential parameters appears to be an efficientmethod to check the robstness of the potential, i.e. to identify the possible artefacts related to theroughness of the potential. After correcting for these artefacts, the most influencial parameterson the computed properties are identified. In particular, one parameter affects most of the self interstitial properties, while having a limited effect on bulk and vacancy properties. This suggestsa sensitivity indices-oriented technique to refit existing potentials. Following this approach, werefine a potential for irradiation defects in α-Zr, thanks to a multi-objective optimisation package