A Self-Evolving Machine-Learning-Based Kinetic Monte Carlo Method for Modelling Thin-Film Growth

This paper presents a self-evolving kinetic Monte Carlo framework that dynamically trains machine-learning models on-the-fly to efficiently predict atomic diffusion rates for thin-film growth simulations, achieving high computational efficiency and accuracy by replacing expensive calculations with uncertainty-guided learning, as demonstrated in Ag/Ag{111} sub-monolayer growth.

The Gist

Imagine you are trying to predict how a crowd of people will move through a giant, shifting maze. In the world of thin-film manufacturing (like making solar panels or computer chips), scientists need to understand how individual atoms move and stick together to form layers.

This paper introduces a new, smart computer program designed to simulate exactly that: how atoms dance, jump, and build islands on a surface. Here is how it works, explained simply:

The Old Way: The Exhaustive Librarian

Traditionally, scientists used two main methods to study this, but both had flaws:

1. The "Slow Motion" Method (Molecular Dynamics): This is like watching a movie of every single atom vibrating. It's incredibly accurate, but the movie plays so slowly that you can only watch a few seconds of "real time" before the computer crashes. It's like trying to watch a whole year of a person's life by watching every single second of it in real-time.
2. The "Rulebook" Method (Standard Kinetic Monte Carlo): This skips the vibrations and just looks at the jumps. It's fast, but it relies on a pre-written rulebook. The problem is that the "rules" for how an atom jumps depend on exactly what its neighbors are doing. Since there are infinite ways neighbors can arrange themselves, writing a rulebook for every single possibility is impossible. It's like trying to write a dictionary for every possible sentence a human could ever speak.

The New Way: The Self-Learning Apprentice

The authors created a Self-Evolving Machine-Learning KMC method. Think of this as a smart apprentice who learns on the job.

1. The Starting Point: The computer starts with a basic map of how atoms should behave (based on physics equations), but it doesn't know the specific "cost" (energy) of every possible jump yet.
2. The "Guess and Check" Loop:
   • When the simulation needs to know the energy cost of a specific jump, the apprentice first guesses using a Machine Learning (ML) model.
   • The ML model also says, "I'm pretty sure about this guess," or "I'm not sure at all."
   • If the model is confident: It uses the guess. This is fast and efficient.
   • If the model is unsure: It pauses, does a rigorous, slow, high-precision calculation (called NEB) to find the exact answer, and then adds that new fact to its memory bank.
3. The Evolution: As the simulation runs, the apprentice encounters new situations. Every time it gets confused, it learns the answer and stores it. Over time, its "memory bank" grows, and it needs to do the slow, hard calculations less and less. It becomes faster and faster while staying accurate.

The Specific Experiment: Silver on Silver

To test this, the team simulated Silver (Ag) atoms landing on a Silver {1 1 1} surface.

• The Challenge: Atoms don't just land in a perfect grid. They can land in slightly different spots, form little islands, and those islands can grow in weird shapes (triangles, jagged lines, or smooth circles) depending on the temperature.
• The Result: The self-learning model successfully predicted how these islands would form.
  • At low temperatures, the atoms were sluggish and formed messy, jagged clusters (like a pile of leaves).
  • At higher temperatures, the atoms had enough energy to move around and form neat, triangular islands (like a well-organized stack of coins).
  • The shapes and sizes of these islands matched what scientists have seen in real experiments and what other complex theories predicted.

Why This Matters (According to the Paper)

The paper claims this is a breakthrough because it solves a major bottleneck: It allows for "full atomistic accuracy" without needing to know every possible rule beforehand.

• No Pre-Programming: You don't need to tell the computer every possible way an atom can jump. The computer figures it out as it goes.
• Dynamic Growth: The simulation can handle atoms piling up to form new layers and new angles (facets) automatically, without the computer crashing or needing a rigid, pre-defined grid.
• Efficiency: It starts slow (learning the rules) but gets faster as it learns, eventually running much quicker than traditional methods while keeping the same level of detail.

In short, the authors built a digital "apprentice" that learns the rules of atomic movement by doing the work, rather than being handed a manual. They proved it works by watching silver atoms build tiny, perfect islands, matching real-world physics perfectly.

Technical Summary

Technical Summary: A Self-Evolving Machine-Learning-Based Kinetic Monte Carlo Method for Modelling Thin-Film Growth

Problem Statement
The theoretical design of thin films and nanoscale materials requires computational models that achieve atomic-scale accuracy over experimentally relevant timescales. Existing methods face a fundamental trade-off: Molecular Dynamics (MD) offers detailed deterministic descriptions but is limited to nanosecond timescales due to the rarity of diffusion events compared to thermal vibrations. Conversely, traditional Kinetic Monte Carlo (KMC) extends simulation times to seconds but often sacrifices atomistic accuracy. Standard KMC relies on Bond-Counting Schemes (BCS) or pre-tabulated activation barriers, which fail to capture the complex dependence of diffusion rates on the Local Atomic Environment (LAE). While "on-the-fly" KMC methods calculate barriers as needed, they struggle with the vast, heterogeneous space of LAEs in multi-component systems and the presence of unstable atomic configurations (e.g., overhangs) or intermediate potential energy minima. Currently, no model exists that simulates the growth of multi-elemental films with full atomistic accuracy in the multilayer regime where LAEs cannot be predicted a priori.

Methodology
The authors present a novel, self-parameterizing Machine-Learning augmented Kinetic Monte Carlo (ML-KMC) framework. The method operates on a dynamic, refined lattice that generates new adsorption sites (both fcc and hcp) as needed, allowing for the organic formation of facets at arbitrary angles and the handling of off-lattice positions.

Key methodological components include:

• Self-Evolving Training: The algorithm builds a Gaussian Process Regression (GPR) model on runtime. It uses Smooth Overlapping Atomic Positions (SOAP) descriptors to characterize the LAE up to the fourth-nearest neighbor.
• Uncertainty-Driven Sampling: The GPR model estimates migration barriers and their associated uncertainties. If the uncertainty exceeds a user-defined threshold (0.05 eV), a Nudged Elastic Band (NEB) calculation is triggered using the LAMMPS software and a Finnis-Sinclair potential to generate a new data point for the training set. This ensures the model gains confidence over time, reducing the frequency of expensive NEB calculations.
• Handling Complex Kinetics: To address challenges such as intermediate energy minima and overhang positions, the model allows hops to multiple distances (up to 3nn) and includes exchange-displacement events. It explicitly avoids artificial stabilization techniques (like tethering) by permitting atoms to jump to true potential energy minima, including metastable hcp sites and intermediate states.
• Data Stratification: To reduce training data heterogeneity, separate ML models are trained for specific jump lengths and coordination numbers of the jumping atom at initial and final positions.
• Efficiency Optimizations: The code includes a "flickering" resolution scheme that temporarily lowers jump rates when the system is trapped in low-barrier cycles, and a pattern recognition database to avoid redundant minimizations.

Key Contributions

1. Dynamic Lattice Generation: The development of an algorithm that dynamically generates fcc and hcp sites for new layers and arbitrarily oriented facets, enabling the simulation of 3D morphological evolution without rigid lattice constraints.
2. Self-Evolving ML-KMC: A framework that learns migration barrier distributions on-the-fly without requiring a priori knowledge of diffusion events or specific training data sampling schemes.
3. Robustness to Complex Environments: The ability to handle unstable atomic environments (overhangs) and intermediate potential energy minima (e.g., fcc-hcp-fcc pathways) by allowing multi-distance hops and avoiding artificial barrier adjustments.
4. Open-Source Implementation: The provision of a fully functional, open-source codebase (licensed under GPL v3) capable of simulating any element or alloy supported by LAMMPS potentials.

Results
The method was validated through simulations of sub-monolayer homoepitaxial growth of Silver (Ag) on an Ag {1 1 1} surface.

• Barrier Accuracy: The ML model demonstrated high fidelity in predicting migration barriers for key processes (terrace diffusion, dimer migration, edge diffusion, and step crossing) when compared to NEB calculations and literature values. While some barriers (e.g., dimer bond energies) were slightly underestimated, the relative stability of fcc vs. hcp dimers and the correct ordering of edge-to-corner barriers were preserved.
• Morphological Evolution: The simulations successfully reproduced the temperature-dependent evolution of island shapes. At low temperatures (<200 K), the model produced numerous small clusters; at intermediate temperatures (200–250 K), dendritic islands formed; and at higher temperatures (>300 K), compact triangular islands bounded by B-edges emerged. This aligns with experimental observations and theoretical expectations regarding corner-crossing anisotropy.
• Island Density Scaling: The scaling of island density with temperature followed theoretical predictions. The model correctly identified a transition in the dominant mobile species (from monomers to mobile dimers) around 110 K. The fitted activation energy sum (0.275 ± 0.006 eV) agreed within error margins with the sum of the monomer and effective dimer barriers derived from the ML model (0.255 ± 0.015 eV), confirming the internal consistency of the diffusion kinetics.
• Computational Efficiency: The self-evolving nature of the model proved efficient. While NEB calculations were initially the bottleneck, their frequency decreased as the model learned. At lower temperatures, the computational cost shifted to LAE construction and SOAP descriptor generation, but the overall approach remained viable for reaching saturation island densities.

Significance and Claims
The paper claims that this work represents a crucial step toward realizing A.F. Voter's vision of fast and accurate modeling of surface dynamics for multicomponent thin films on chemically heterogeneous substrates. The authors assert that their model is the first to simulate the growth of multi-elemental films in the multilayer regime with full atomistic accuracy, overcoming the limitations of BCS and rigid lattice approaches. By demonstrating feasibility on Ag homoepitaxy, the study establishes a foundation for extending the method to complex, multi-elemental systems relevant to applications such as photovoltaics, energy-saving windows, and biosensors. The authors remain modest, noting that while the current proof-of-concept uses Gaussian Process Regression, the workflow is adaptable to other uncertainty-aware ML models, and the primary computational bottlenecks (NEB and descriptor generation) offer avenues for future optimization.