A Self-Evolving Machine-Learning-Based Kinetic Monte Carlo Method for Modelling Thin-Film Growth Journal: Computational Materials Science

Jyri Kimari, Flyura Djurabekova, Kostas Sarakinos · 2026

Read A Self-Evolving Machine-Learning-Based Kinetic Monte Carlo Method for Modelling Thin-Film Growth Journal: Computational Materials Science by Jyri Kimari, Flyura Djurabekova, Kostas Sarakinos (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-2605-31144.txt.

Introduction

Kimari, Djurabekova and Sarakinos simulate thin films atom by atom with a kinetic Monte Carlo method that teaches itself as it runs. A machine-learning model estimates jump rates from local atomic surroundings, calling in costly exact calculations only where its uncertainty runs high. As confidence grows, fast estimates take over, and a silver-on-silver test reproduces island shapes and densities matching theory and experiment.

Catalog record

CREATORJyri Kimari, Flyura Djurabekova, Kostas Sarakinos
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE39889
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2605.31144

Browse the full library · About the Tower of Babel

Full text

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

A Self-Evolving Machine-Learning-Based Kinetic Monte Carlo Method for Modelling Thin-Film Growth Journal: Computational Materials Science Jyri Kimari, Flyura Djurabekova, Kostas Sarakinos Abstract We present a kinetic Monte Carlo (KMC) simulation framework parameterized by automatically sampling machine-learning (ML) for modeling thin-film growth atom by atom. Given an interatomic potential energy function, the KMC algorithm builds an ML-based regression model for rate parameters on runtime, being trained on the local atomic environments encountered during the system evolution. New environments are continuously added to the training set in a self-evolving manner at points where the ML model estimates high uncertainty. As the simulation progresses, the ML model gains confidence, and the quick estimation of rates increasingly overtakes the relatively-expensive nudged elastic band calculations, promoting computational efficiency while retaining high fidelity description of the atomic diffusion kinetics. As a test case, we simulate the sub-monolayer growth of Ag on Ag { 1 1 1 } \{1\,1\,1\} , where we demonstrate adatom islands forming in shapes and densities in accordance with the underlying atomistic interaction model, the theoretical framework, and available experimental results related to thin-film nucleation and growth. 1 Introduction Theory-informed design and synthesis of thin films and nanoscale materials requires computational modeling of surface and bulk diffusion processes with atomic-scale accuracy at experimentally-relevant timescales. State-of-the-art methods cannot meet those two requirements simultaneously. Molecular Dynamics (MD) provides access to a detailed deterministic description of atomic-scale processes, but at timescales that do not exceed a few tens of nanoseconds 62 . This is because atomic diffusion is a rare event that occurs at frequencies many orders of magnitude smaller than that of thermal vibration of atoms, thereby limiting the length of the timestep required for accurate integration of the equations of motion. Kinetic Monte Carlo (KMC) models forego the description of atomic vibrations allowing for the simulation timescale to be extended to a few seconds and beyond 95 . However, this comes at the expense of atomistic accuracy and completeness as KMC models are parameterized directly in terms of the activation barriers, which have a complex dependence on the local atomic environment (LAE) of the diffusion events. Due to this complexity, full tabulation of every barrier is practically impossible. One common approximation to overcome this inherent weakness of KMC is the use of bond-counting schemes (BCS), where the description of the LAE is simplified down to the number of first- and second-nearest neighbour bonds that the diffusing atom breaks and forms along its migration path 72 . The activation energy is thus a function of the bonds, and the coefficients of this function are fitted to a sample of representative barriers obtained, e.g., by molecular statics calculations. BCS parameterizations have been used to simulate deposition and diffusion on solid surfaces already before Bortz, Kalos and Lebowitz introduced the nowadays standard n -fold KMC algorithm in 1975 12 . Abraham and White modeled adsorption, evaporation and migration on cubic close-packed surface in 1970, describing all event rates in terms of single bond energies 2 . Gilmer and Bennema carried out a similar study on crystal growth from solution in 1972 42 . The first efforts in developing a BCS approach for specific materials, instead of generic crystals, were taken by Voter in 1986 for the case of Rh clusters diffusing on the Rh \{1\,0\,0\} surface 94 . Since then, the development of BCS-reliant KMC and lattice gas (LG) models has branched both in the simple, generic-system 28 , 29 , 87 , 30 , 93 , 58 , 59 , 60 , 97 and material-specific 27 , 84 , 46 , 45 , 14 , 70 , 89 , 11 , 69 , 38 , 56 , 10 , 47 , 48 , 39 directions. The BCS approach was also taken in Refs. 63 , 41 , 40 to simulate Ag deposition and diffusion on Ag \{1\,1\,1\} and generic weakly-interacting substrates with the purpose of studying the dynamics of three-dimensional morphological evolution of thin films. Even after almost 40 years, Voter’s vision of using simple, or at least tractable, models to capture “classically accurate” dynamics at the fraction of the cost of MD simulations, has not been fully realized. Instead, it was understood rather early on 87 that the number of necessary parameters to accurately represent the microscopic evolution of real-material surfaces in KMC is effectively infinite. This realization, along with the increasing computational power at researchers’ disposal, has spawned the class of on-the-fly KMC models that are launched without an a priori list of events and associated barriers. Instead barriers are calculated on runtime only for the environments they encounter, using, e.g., saddle point search for identifying diffusion events 6 , 43 . For computational efficiency, any found events and barriers may be saved in a database for later reuse 91 , 82 , 37 , 57 , 9 , 61 . Concurrently, the so-called “hybrid” KMC methods, where the system is treated partially by the stochastic event-based algorithm and partially by deterministic MD, have shown some popularity 100 , 99 , 79 . The vast majority of thin-film applications concerns the deposition of multi-component layers on chemically heterogeneous substrates. The implementation of BCS-based models in such growth scenarios is challenging as film-forming species are expected to encounter numerous and a priori unknown LAEs, each of which may require a separate BCS function for barrier calculation. These limitations have been circumvented chiefly by either limiting the growth to two-dimensional, submonolayer regime 89 , 58 , 59 , 60 , 31 , or (in the case of three-dimensional growth) by coarse-graining sites to represent multiple atoms 13 and by imposing a rigid NaCl 4 , 77 , 13 , 66 , 67 , 65 or wurtzite 50 , 26 lattice structure with predetermined sites for different atomic species. Additionally, the diffusion of impurity species 31 , or even all deposited atoms 49 may also be restricted. In some cases, hybrid KMC-MD approaches have been applied to molecular deposition of thin films 1 , 78 . To our knowledge, there is no model to simulate with full atomistic accuracy growth of multi-elemental films in the multilayer growth regime where the LAEs cannot be predicted a priori . This includes the growth of metals and alloys on weakly-interacting substrates which exhibit a pronounced 3D morphological evolution. These systems are relevant for applications such as photovoltaics 96 , energy-saving windows 34 , and biosensors 98 . Specific challenges also include facet formation and large anisotropy in the diffusivity of the film-forming species. The growth of new facets with off-lattice sites was captured by Lü et al. using a BCS model 63 , but attempts to extend this approach to the multi-elemental case have been unsuccessful. Breakthroughs in machine-learning (ML) methods in the last two decades have paved the way for development of new, efficient and accurate atom-level simulation techniques. Bartók et al. pioneered the Gaussian approximation potential (GAP) approach 7 and the smooth overlapping atomic positions (SOAP) atomic environment descriptors 8 . ML interatomic potentials have since been well-established in MD simulations 35 , 76 . Djurabekova, Castin and others have applied artificial neural networks to KMC simulations of bulk diffusion in Fe-based alloys in a rigid lattice 36

Read the complete text