Amorphous silicon structures generated using a moment tensor potential and the activation relaxation technique nouveau
Karim Zongo 1,∗, Hao Sun 2, Claudiane Ouellet-Plamondon 1, Normand Mousseau 3, Laurent Karim Béland 2,∗ · 2025
Read Amorphous silicon structures generated using a moment tensor potential and the activation relaxation technique nouveau by Karim Zongo 1,∗, Hao Sun 2, Claudiane Ouellet-Plamondon 1, Normand Mousseau 3, Laurent Karim Béland 2,∗ (2025) 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-2501-10549.txt.
Introduction
Karim Zongo and co-authors grow glass-like silicon forms with a learned force field in a 2025 paper. True-to-life forms of this glass have taxed the field for decades of hard work. Their route pairs a hop-and-relax hunt with a force field trained on the fly for the task.
Seven forms span from a few hundred atoms to a few thousand atoms in size. Each is checked on near and mid shape traits, spare energy, and strength marks in depth. The forms match lab data and top past forms well, with low flaw counts in the small ones.
Catalog record
| CREATOR | Karim Zongo 1,∗, Hao Sun 2, Claudiane Ouellet-Plamondon 1, Normand Mousseau 3, Laurent Karim Béland 2,∗ |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 56 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-10549
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Amorphous silicon structures generated using a moment tensor potential and the activation relaxation technique nouveau Karim Zongo 1,∗, Hao Sun 2, Claudiane Ouellet-Plamondon 1, Normand Mousseau 3, Laurent Karim Béland 2,∗ Abstract Abstract Preparing realistic atom-scale models of amorphous silicon (a-Si) is a decades-old condensed matter physics challenge. Herein, we combine the Activation Relaxation Technique nouveau (ARTn) to a Moment Tensor Potential (MTP) to generate seven a-Si models containing between 216 and 4096 atoms. A thorough analysis of their short-range and medium-range structural properties is performed, alongside assessments of excess energy and mechanical properties. The seven ARTn-MTP models are compared with available experimental data and other high quality a-Si models present in the literature. The seven ARTn-MTP a-Si models are in excellent agreement with available experimental data. Notably, several of our models, including the 216-atom, 512-atom, and 1000-atom a-Si models, exhibit low coordination defects without any traces of crystalline grains. Historically overlooked in previous research, our study underlines the need to assess the validity of the continuous random-network hypothesis for the description of perfect amorphous model by characterizing local crystalline environment and to explore the crystallisation process of a-Si through modelling. I Introduction Amorphous silicon (a-Si) is used in applications that include photovoltaics, thin-film transistors, battery electrodes, and liquid-crystal displays street1999technology , carlson1976amorphous , powell1989physics , cui2009crystalline , key2009real . Recent advances include its use as coating materials for next-generation gravitational wave detectors birney2018amorphous . Additionally, a-Si serves as the canonical disordered solid prototype for the development of modeling methods. a-Si lewis2022fifty , madanchi2024future is idealized as a continuous random network (CRN) zachariasen1932atomic , characterized by a local atomic environment akin to crystalline silicon wooten1985computer , vink2001device . It maintains a local tetrahedral crystalline environment without long-range order. Real a-Si, however, is believed to contain defects associated with its disordered nature at a level that depends on its preparation roorda1991structural , laaziri1999high , laaziri1999high_xray , levesque2022internal , birney2018amorphous . These include local coordination defects, medium-range ring defects and two-level systems–the latter contribute to noise in sensitive gravitational wave detectors and cause decoherence in quantum computers levesque2022internal , birney2018amorphous . Variations in tetrahedral geometry and bond lengths also significantly affect a-Si’s performance. Controlling structural anomalies during preparation poses challenges from both an experimental and a theoretical perspective. Indeed, the structure of a-Si is highly influenced by preparation methods holmstrom2016dependence ; for instance, chemical vapor deposition kail2011configurational and ion implantation roorda1989calorimetric , van1992evidence , laaziri1999high can introduce voids and coordination defects. Thermal annealing roorda1989calorimetric , fortner1989radial , laaziri1999high , haberl2009structural often refines these imperfections, enhancing material quality. Additionally, a recent study rosset2025signatures has raised an important issue that questions the definition of the true model of a-Si. This study explored the footprints of paracrystallinity in amorphous silicon using molecular dynamics simulations with a machine learning potential rosset2025signatures , morrow2022indirect similar to the one employed herein. As reported by the authors, while optimal models of a-Si are traditionally based on the continuous random network model, some groups have proposed that real a-Si could be better described as ’paracrystalline’. However, since both optimal types of structure conflict with experimental observations treacy1998paracrystallites , voyles2001structure , treacy2012local , roorda2012comment , the authors have further investigated the paracrystalline model, concluding that realistic experimental models of a-Si may contain some degree of crystalline grain traces. The observed degree of crystallinity is likely influenced by the preparation method—i.e. MD. Notably, while the degree of crystallinity was examined, its relationship to coordination defects was not. In other words, the central question remains how to balance crystallinity and coordination defects when preparing a-Si. For instance, a high cooling rate in melt-quench simulations will likely lead to a higher number of coordination defects and lower crystallinity, while a low cooling rate will most likely yield fewer coordination defects and higher crystalline-like environments. Extensive theoretical research has focused on developing realistic structural models for a-Si wooten1985computer , barkema1996event , barkema2000high , deringer2018realistic , justo1998interatomic , ishimaru1997generation , wright2013eighty . Various methods for simulating materials, including a-Si, can be categorized into dynamic and non-dynamic approaches lewis2022fifty , madanchi2024future . Dynamic methods, traditionally based on density functional theory (DFT) sholl2022density or semiempirical force fields torrens2012interatomic , primarily use melt-quench processes. However, these methods face challenges: DFT simulations are limited in system size and quenching rate, and semi-empirical force fields are often compromised by limited accuracy and transferability issues. Non-dynamic methods include the Reverse Monte Carlo (RMC) approach mcgreevy1988reverse , mcgreevy1992rmc , gereben2007new , gurman1990reverse , mcgreevy2001reverse . In RMC, the computer-generated a-Si structure is randomly modified in order for its simulated structure factor and radial distribution function g(r) to match experimental data. RMC’s main drawback is that it can yield multiple configurations for the same g(r) — in other word the solution is degenerate. To overcome this, hybrid RMC methods incorporate force field interactions, ensuring generated configurations are both consistent with experimental data and physically realistic. Variants like the force-enhanced atomic refinement pandey2015force , igram2018large and experimentally constrained molecular relaxation biswas2004inclusion utilize semi-empirical and DFT interactions, respectively. Other key non-dynamic methods include the Wooten-Winer-Weaire (WWW) bond-switching algorithm wooten1985computer , which optimizes atomic configurations by minimizing energy, and the Activation-Relaxation Technique nouveau (ARTn) barkema1996event , malek2000dynamics , mousseau2012activation , jay2022activation , which combines activation and relaxation steps to discover low-energy structures. ARTn can be paired with various force fields and ab initio methods, making it suitable for applications that include relaxation of disordered systems and diffusion processes mousseau1999exploring , levesque2022internal , ganster2012first , el2007ab , salles2017strain , joly2012optimization , sun2022statistical , sun2023calculation , trochet2020off , beland2011kinetic . One of the key components in both dynamical and non-dynamical methods is the interaction potential. The ongoing challenge of balancing accuracy and computational cost in these methods has led to the development of machine learning interatomic potentials (MLIPs) behler2007generalized , deringer2019machine , wang2024machine . MLIPs combine descriptors of the atomic environment, regression methods, and DFT data. Many MLIPs frameworks have been developed, including artificial neural networks behler2007generalized , khorshidi2016amp , kernel-based methods bartok2010gaussian , bartok2018machine , and linear regression approaches thompson2015spectral , shapeev2016moment , among others. For a detailed description of MLIPs, including their implementation and testing, the reader is referred to the existing literature wang2024machine . In the case of a-Si, structures have been modeled using artificial neural networks li2020unified , fan2021neuroevolution , li2019dependence and Gaussian Approximation Potentials (GAP) bartok2010gaussian , bartok2018machine . In initial studies, the GAP potential was applied to a-Si systems containing between 512 and 4,096 atoms deringer2018realistic . Later, the GAP model was extended to larger systems of up to 100,000 atoms to investigate the the atomistic mechanisms underlying various structural transitions in disordered silicon deringer2021origins . Other MLIPs, such as the Moment Tensor Potential shapeev2016moment (MTP), have also been applied to a-Si zongo2024unified . In these studies, melt-quench MD simulations were employed to generate a-Si structures. A natural next step is to couple ARTn barkema1996event , malek2000dynamics , mousseau2012activation , jay2022activation with machine learning potential to generate a-Si models. The objective of this paper is to generate a-Si models using ARTn-ML, which could enhance our understanding of the a-Si model and open new avenues for further research. The manuscript is organized into three main sections. The Methods section outlines the principles of the