Reactive force field potential for carbon deposition on silicon surfaces

Abstract

In this paper a new interatomic potential based on the Kieffer force field and designed to perform molecular dynamics (MD) simulations of carbon deposition on silicon surfaces is implemented. This potential is a third-order reactive force field that includes a dynamic charge transfer and allows for the formation and breaking of bonds. The parameters for Si-C and C-C interactions are optimized using a genetic algorithm. The quality of the potential is tested on its ability to model silicon carbide and diamond physical properties as well as the formation energies of point defects. Furthermore, MD simulations of carbon deposition on reconstructed (100) silicon surfaces are carried out and compared to similar simulations using a Tersoff-like bond order potential. Simulations with both potentials produce similar results showing the ability to extend the use of the Kieffer potential to deposition studies. The investigation reveals the presence of a channelling effect when depositing the carbon at 45° incidence angle. This effect is due to channels running in directions symmetrically equivalent to the (110) direction. The channelling is observed to a lesser extent for carbon atoms with 30° and 60° incidence angles relative to the surface normal. On a pristine silicon surface, sticking coefficients were found to vary between 100 and 73%, depending on deposition conditions.

Efficient global optimization of reactive force‐field parameters

“…Parameters in a specialized charge‐transfer force field were optimized with a GA . Pahari and Chaturvedi optimized ReaxFF parameters with a GA, but the focus of their paper was on determining a minimal set of parameters to vary in the GA based on prior sensitivity tests and cross‐correlations.…”

“…[20][21][22][23][24][25][26][27][28][29][30] The parameter optimization problem for reactive force fields is harder than that of traditional force fields, because there are far more parameters per atom, these parameters are more strongly coupled, a significantly larger reference data set is needed, and we have limited knowledge about the relationship between reference data items and force field parameters. GA methods have successfully been applied to this challenging task, [31][32][33] including a GA optimization study of ReaxFF parameters for SiOH [34] and azobenzene [35] by one of the present authors.…”

“…The techniques used in Ref. [31][32][33][34][35] are single-objective GA optimization techniques. Recently, a number of studies using multi-objective GA techniques to optimize ReaxFF parameters were published [36][37][38] (see Section Related Work).…”

Accurate meso-scale dynamics by kinetic Monte Carlo simulation via free energy multicanonical sampling: oxygen vacancy diffusion in BaTiO3

“…However, it is not straightforward to prepare such structure-related data, because it is difficult to perform exhaustive structure samplings. Recently, the genetic algorithm has been successfully applied to the parameter optimization for molecular dynamics [50][51][52][53][54]. In this work, we apply the MUCA-MC simulations not only to calculate free energy and kinetic parameters but also to prepare the training data set for the parameter optimization, because the MUCA-MC simulations can generate many structurerelated data without any intentional purpose.…”

Development of a new parameter optimization scheme for a reactive force field based on a machine learning approach

“…For fine refinement of the force field parameters, a new ML‐based parameter optimization process was developed. The distinct features of the proposed method are as follows: (1) obtaining several local minima with the k ‐nearest neighbor algorithm, (2) an efficient optimization process based on an ML approach, and (3) the ML‐based optimization can be combined with other optimization processes such as Monte Carlo and the genetic algorithm . We note that the third feature is particularly important, because the combination of an efficient algorithm with a flexible ML function model may provide a powerful tool for efficient and accurate MD simulation of many materials and contribute to the field of computational material science.…”