8:15 AM - DS02.14.02
Synergistic Coupling in QM/MM Simulations of Dislocations via Machine Learning
Petr Grigorev1,James Kermode2,Mihai-Cosmin Marinica3,Thomas Swinburne1
CINaM Centre Interdisciplinaire de Nanoscience de Marseille1,School of Engineering, University of Warwick2,CEA, Service de Recherches de Métallurgie Physique3
Show Abstract
Ab initio modelling of dislocations is essential to gain quantitative insight into a vast range of important material phenomena, from elementary glide or climb mechanisms to predictive models of solute solution strengthening or post irradiation annealing. The simulation data can be used directly in phenomenological micro-structural models or in the construction of empirical interatomic potentials in a multi-scale framework. With some notable exceptions, ab initio size limitations mean that most dislocations cannot be contained in periodic supercells, requiring the use of flexible boundary techniques which either couple to an elastic Green's function or an empirical interatomic potential in QM/MM simulations. In the latter case, the potential should ideally have identical elastic properties as the ab initio medium. A key limitation of these flexible boundary methods is that whilst the ionic positions and forces can be extracted, the total energy becomes ill-defined.
In recent work [1,2], we showed how this information could be used to extract energetic barriers for glide or segregation through the principle of virtual work. A very good agreement was found with special cases treatable with periodic supercells, whilst in general our ab initio results deviate significantly from predictions of empirical potentials, particularly for prismatic (edge) dislocations. In this work, we show how linear-in-descriptor machine learning (ML) potentials [3] can be used to exactly match ab initio elastic properties, typically leading to efficiencies over "traditional" empirical potentials during QM/MM relaxations. We then go further, using the ionic positions and forces to train the ML potential, showing that the resultant "classical" migration barriers are in excellent agreement with that obtained from QM/MM. Our work shows that QM/MM methods can qualitatively expand the range of atomic structures in the training database of machine learning interatomic potentials, with significant impact on their transferability, or predictive power, for materials modelling.
References:
[1] Swinburne, T. D., & Kermode, J. R. (2017). Computing energy barriers for rare events from hybrid quantum/classical simulations through the virtual work principle. Phys. Rev. B, 96(14), 144102.
[2] Grigorev, P., Swinburne, T. D., & Kermode, J. R. (2020). Hybrid quantum/classical study of hydrogen-decorated screw dislocations in tungsten: Ultrafast pipe diffusion, core reconstruction, and effects on glide mechanism. Phys. Rev. Materials, 4(2), 23601.
[3] Dérès J, Goryaeva A. M., Lapointe C., Grigorev P., Swinburne, T. D., Kermode, J. R., Ventelon L., Baima J., and Marinica M.-C. Efficient and transferable machine learning potentials for the simulation of crystaldefects in bcc Fe and W. Submitted to Phys. Rev. Materials