Classical interatomic potentials (CIPs) have a low-to-intermediate number of parameters and fixed functional forms that are based on the physical understanding of the chemical bonding [1]. Numerically lighter than machine learning potentials, they enable studies not otherwise accessible at the atomic scale, such as uncovering plasticity mechanisms, or validating mesoscale models of alloy mechanical properties [2]. Identifying the CIP parameters for a specific material system/physical problem remains a challenging step, knowing that the flexibility of CIPs is limited. A better understanding of the actual capabilities of a given potential type to reproduce some target combination of properties would thus be relevant when selecting, using and/or fitting such a potential. In addition, ways to improve CIPs when their transferability is unsatisfactory are not clearly established.
In this work, we select the case of a many-body central force potential and investigate its ability to reproduce bulk and defect properties in hcp zirconium: hcp is a crystallographic structure for which angular dependence matters, yet simple central force potentials (EAM, FS, etc.) are still widely used in the scientific community for non-fcc metals. We first combine efficient model screening tools to a variance-based sensitivity analysis to establish some of the intrinsic limits of this class of CIPs. We recover the known effect of the potential cutoff on the accessible range of c/a ratios and various stacking fault energies (important for both plasticity and irradiation). Interestingly, the influence of CIP parameters is different from one property to another; in particular, vacancy and self-interstitial energetics are mostly affected by two distinct parameters. We therefore propose to use the results of sensitivity analysis to re-optimize existing CIPs and improve their transferability; this process is illustrated by refitting a CIP for the study of small irradiation defects in hcp-Zr. Finally, we discuss the applicability of our approach to (i) more complex interaction models with up to tenth of parameters (MEAM, BOPs, etc.) and (ii) the development of entirely new potentials [3].
[1] Y. Michin, Acta Mater. 214 (2021)
[2] F. Maresca & W. A. Curtin, Acta Mater. 182 (2020), R. Gröger, MSMSE 30 (2022), D. Caillard et al. Nature 609 (2022)
[3] A. Del Masto, J. Baccou, G. Tréglia, F. Ribeiro & C. Varvenne, Comput. Mater. Sci. 231 (2024)
| Subject : | : | oral |
| Topics | : | 1-7 - Multiscale materials modelling from atoms to macroscale |
| PDF version | : |  PDF version |
