Melting of thin silicon films: a molecular dynamics study with two machine learning potentials

Yu. D. Fomin, E. N. Tsiok, V. N. Ryzhov · 2026

Read Melting of thin silicon films: a molecular dynamics study with two machine learning potentials by Yu. D. Fomin, E. N. Tsiok, V. N. Ryzhov (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-11722.txt.

Introduction

Fomin, Tsiok, and Ryzhov simulate how thin silicon films melt, using two machine learning potentials, SNAP and GAP. Films run from a single silicene layer up to 36 layers under SNAP. Silicene loses its structure at 500 K. The breakdown temperature climbs with thickness and levels off near 28 layers, at the model's bulk melting point of 1380 K. Films up to 8 layers show two-phase coexistence, while thicker ones melt from the surface inward. GAP, though sharper for bulk silicon, fails on the gas phase and shreds silicene into small clusters.

Catalog record

CREATORYu. D. Fomin, E. N. Tsiok, V. N. Ryzhov
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE15143
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.11722

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Melting of thin silicon films: a molecular dynamics study with two machine learning potentials Yu. D. Fomin, E. N. Tsiok, V. N. Ryzhov Abstract Thermal stability of silicene and thin silicon films is studied by molecular dynamics using two machine-learning potentials, SNAP and GAP. For SNAP potential, systems ranging from a single silicene layer to films of 36 layers are considered. Silicene is found to lose its structure at 500 K. The decomposition temperature increases with film thikness and reaches saturation at about 28 layers, corresponding to the bulk melting point of the SNAP model (1380 K). Thin films up to 8 layers exibit two-phase coexistence upon decomposition, while thicker films undergo surface melting followed by complete collapse into the liquid state. The GAP potential, although more accurate for bulk silicon, fails to describe the gas phase: silicene modelled with GAP decomposes into a set of small clusters. The results are compared with earlier data for the Stillinger-Weber potential. Keywords : silicene, melting, molecular simulation, machine learning potential pacs 61.20.Gy, 61.20.Ne, 64.60.Kw I Introduction Silicon is the base of modern electronics. It makes this material of great technological importance. In particular, silicon structures with low dimensionality, such as silicene, are interesting as promising two-dimensional materials. Silicene was predicted in a theoretical work si-1 in 1994 and synthesized in 2012 si-syn . Unlike graphene, silicene is not a perfectly flat material, but a corrugated layer, i.e., some of the atoms are above the symmetry plane and others are below. It is expected that silicene can be employed in different technological applications, including electronics, which is a traditional field for silicon si-electronics , but also promising technologies of green energy, such as hydrogen storage si-h , lithium-ion batteries si-li-ion , etc. Although the properties of silicene were widely studied both experimentally (see the review articles rev-1 ; rev-2 and references therein) and in simulations (see, for instance, dft-1 ; dft-2 ; dft-3 ; dft-4 ; dft-5 ; dft-6 for density functional theory studies of silicene), there are still a lot of opened questions. It is of particular interest to study the thermal stability of silicene. At the same time this is a difficult problem both for experiments and simulations. The melting of silicene and thin silicon film is challenging for DFT calculations since it requires rather large samples (thousands of atoms) at rather high temperature. For this reason it looks appropriate to employ the methods of molecular simulation with effective potentials, which allow to calculate even millions of atoms. However, these effective potential can be not accurate enough to reproduce the behavior of a real system. It results in a great spread of data for the thermal stability of silicene available in the literature. For instance, melting of silicene modelled with the Tersoff potenital tersoff was reported in Ref. melting-1 . Two sets of parameters of the potential were used: the original one tersoff and the one developed for the void-nucleated melting of silicon tersoff-a . The melting temperature was estimated to be T_{m}=3600 K in the former case and T_{m}=1750 K in the latter. This example clearly show that the melting point of silicene is extremely sensitive to the model used. In our recent papers we employed a well recognized Stillinger-Weber model sw to simulate the thermal stability of silicene and thin silicon films of different width we1 ; we2 . It was found that the temperature of thermal decomposition of a silicon film increases with increasing of the film width. However, when the number of layers is larger than 12 the decomposition temperature comes to a saturated value consistent with the melting temperature of bulk silicon: T_{=}1725 K (the experimental value is T_{m}^{exp}=1683 K tm ). As it was discussed above, the estimation of thermal stability of silicene evaluated withing the framework of empirical force fields strongly depends on the employed model. In this paper we continue the study of thermal decomposition of silicene and thin silicon films by employing two machine-learning potentials: SNAP snap-silicon and GAP gap-silicon models. It is expected that machine-learning models should better reproduce the interaction of the atoms which should lead to more accurate data. In our recent study melt-sn a comparison of melting lines of SNAP and GAP models was given (the melting line of GAP model was taken from Ref. gap-silicon ). Both models give correct qualitative behavior of the melting line of the diamond phase of silicon (melting line of silicon is a straight line with negative slope), but both lines are systematically lower than the experimental one. The melting temperature of silicon at ambient pressure obtained with SNAP model is T_{m}=1380 K which is substantially below the experimental value. Based on the fact that these models give correct qualitative behavior, we perform simulations of thermal decomposition of silicene and thin silicon films and compare the qualitative results with our previous calculations with empirical SW model. II System and Methods In the present work we investigate crystal structure collapse of a silicon sample of different thicknesses by means of molecular dynamics simulation. Two machine-learning potentials are used: SNAP snap-silicon and GAP gap-silicon . A film of a given thickness is surrounded by vacuum. Periodic boundary conditions are employed in all directions. The size of the system in z direction, where vacuum is added, is 543 \AA , which is approximately one hundred times the lattice constant of silicon. Such vacuum gap is sufficient to prevent interaction of the film with its periodic images. The unit cell parameters of silicene are a=b=3.861 \AA , and \alpha=\beta=90^{o} and \gamma=120^{o} , therefore, the box is triclinic. A rectangular box with equal dimensions in X and Y directions was employed for the thin films with diamond structure. The number of atoms used in simulations with SNAP model in all considered systems is given in the Table I. System Number of particles silicene 3200 4 2048 8 layers 4096 12 layers 6144 16 layers 8192 20 layers 10240 24 layers 12288 28 layers 14336 32 layers 18432 36 layers 20480 Table 1: Number of atoms in systems with different number of silicon layers. In the case of GAP model only a system of silicene with 1800 atoms was used. The reasons for this will be described below. In all cases we start from ideal crystal structure and simulate the system in canonical ensemble (constant number of particles N, volume V and temperature T) at given temperature. The set of temperatures depends on the system. The goal of this work is to find out at which temperature the crystalline structure collapses. The time step is 1 fs and all simulations are conducted up to 100 ps. All simulations were performed with lammps simulation package lammps . III Results and Discussion III.1 SNAP model First of all we consider the case of silicene. Figure 1 shows snapshots of the system at two temperature values: T=475 and 500 K. One can see that at T=475 K the silicene layer is preserved, although it experiences strong corrugation due to thermal motion, but the layer is broken at T=500 K. At the same time some crystalline particles are still observed in the system. One can conclude that this point corresponds to a

Read the complete text