Ultra-Fast Machine-Learned Interatomic Potential for MoS2 Enabling Non-Equilibrium Molecular-Dynamics Simulation of Epitaxial Growth

This paper presents an ultra-fast machine-learned interatomic potential for multilayer MoS2, developed using the UF3 framework, which accurately reproduces DFT-level structural, elastic, and defect properties while enabling large-scale non-equilibrium molecular dynamics simulations of epitaxial growth that reveal experimentally consistent van der Waals gaps and triangular domains.

The Gist

Imagine you are trying to build a microscopic city out of Lego bricks, but these bricks are made of atoms, and the city is a material called Molybdenum Disulfide (MoS₂). This material is a superstar in the world of future electronics because it's incredibly thin, flexible, and efficient. However, building it atom-by-atom is like trying to construct a skyscraper while blindfolded; you need a perfect blueprint to know exactly how each piece fits.

This paper is about creating that perfect blueprint.

The Problem: The "Goldilocks" Dilemma

Scientists have two main ways to predict how atoms behave:

1. The Super-Computer Method (DFT): This is like hiring a team of genius architects who calculate every single angle and force with perfect precision. It's incredibly accurate, but it's so slow that simulating a tiny city block could take years. You can't use it to watch a whole city grow.
2. The Rule-of-Thumb Method (Empirical Potentials): This is like using a simple sketch. It's fast, but it's often wrong. It might tell you the bricks stick together, but it misses the subtle details, like how the layers of the city slide over each other or how they react when they get hot.

For a long time, scientists couldn't find a "Goldilocks" solution for MoS₂—a method that was fast enough to simulate growth but accurate enough to get the physics right. Existing tools either got the layers wrong or were too slow to be useful.

The Solution: The "Ultra-Fast" AI Architect

The authors of this paper developed a new tool called UF3 (Ultra-Fast Force Field). Think of UF3 as a machine-learning architect that was trained by studying millions of different Lego configurations.

• How it learned: Instead of just looking at perfect, static Lego towers, they threw the AI into a chaotic training camp. They showed it broken towers, twisted shapes, high-heat scenarios, and even how the layers slide apart. This taught the AI to handle the messy, real-world chaos of building a material from scratch.
• The Magic Trick: Usually, when you train an AI to be this smart, it becomes slow. But UF3 is special. It uses a clever mathematical shortcut (like a compressed map) that allows it to run almost as fast as the simple "rule-of-thumb" sketches, but with the accuracy of the genius architects.

What Did They Discover?

Once they built this new AI architect, they put it to the test by simulating the epitaxial growth of MoS₂. In plain English, this means they simulated the process of growing a new layer of MoS₂ on top of an existing one, atom by atom.

Here is what the simulation revealed, which matches what real scientists see in labs:

1. The "Velcro" Gap: MoS₂ is made of layers that stick together loosely, like sheets of paper with a tiny bit of Velcro between them. Many old models failed to capture this "gap." UF3 got it right, showing the layers stacking up with the correct spacing.
2. The Triangular Islands: When the material grows, it doesn't just spread out in a perfect square. It forms triangular islands. Why? Because the edges of the triangle are energetically "cheaper" to build. The AI correctly predicted that the atoms would naturally arrange themselves into these triangles, bounded by specific "zigzag" edges.
3. Speed vs. Reality: The simulation ran fast enough to watch the city grow over time, yet it was accurate enough to predict exactly where defects (broken bricks) would form and how the material would react to heat.

Why Does This Matter?

Think of MoS₂ as the next generation of silicon chips. To make these chips, manufacturers need to grow these materials perfectly on top of other materials. If the growth is messy, the chip fails.

Before this paper, scientists were flying blind, guessing how these materials would grow. Now, they have a high-speed, high-fidelity simulator. They can run thousands of virtual experiments to figure out the perfect temperature, speed, and conditions to grow perfect MoS₂ layers before they ever touch a real lab.

The Bottom Line

This paper is a breakthrough because it finally gave scientists a fast and accurate crystal ball. They can now simulate the birth of these microscopic materials in real-time, ensuring that the next generation of super-fast, flexible electronics can actually be manufactured. It's the difference between guessing how a cake will rise and having a recipe that guarantees a perfect bake every time.

Technical Summary

Here is a detailed technical summary of the paper "Ultra-Fast Machine-Learned Interatomic Potential for MoS₂ Enabling Non-Equilibrium Molecular-Dynamics Simulation of Epitaxial Growth."

1. Problem Statement

The development of next-generation electronic and optoelectronic devices relies heavily on Transition Metal Dichalcogenides (TMDs), particularly Molybdenum Disulfide (MoS₂). A critical bottleneck in the commercialization of MoS₂ devices is the lack of predictive control over manufacturing processes, specifically epitaxial growth (the growth of thin films on crystalline substrates).

• Simulation Challenges: Molecular Dynamics (MD) simulations are essential for understanding epitaxial growth mechanisms (e.g., defect formation, domain propagation). However, existing interatomic potentials face a trade-off:
  • Classical Empirical Potentials (CEPs): Fast but lack the accuracy to capture complex phenomena like non-equilibrium growth, defect energetics, and the specific anisotropy of layered materials.
  • Density Functional Theory (DFT): Highly accurate but computationally prohibitive for the large system sizes and long timescales required for epitaxial growth simulations.
  • Existing Machine-Learned Interatomic Potentials (MLIPs): While offering near-DFT accuracy, most are either limited to 2D monolayers (failing to capture interlayer van der Waals interactions) or are orders of magnitude slower than CEPs, making large-scale growth simulations infeasible.

The core problem is the absence of an interatomic potential that combines near-DFT accuracy with empirical-potential speed to simulate the highly non-equilibrium, large-scale process of MoS₂ homoepitaxial growth.

2. Methodology

The authors developed a new MLIP using the Ultra-Fast Force Field (UF3) framework. The development process involved three main stages:

• Dataset Generation:

  • A comprehensive dataset of 28,579 structures was generated using the Genetic Algorithm for Structure and Phase Prediction (GASP) and DFT calculations (VASP with PBE-D3 corrections).
  • The dataset spans a vast range of configurations: stoichiometric and non-stoichiometric Mo-S compositions, polymorphs (1H, 2H, 3R, 1T'), strained structures, surfaces, nanoribbons, and ab-initio molecular dynamics (aiMD) trajectories at temperatures up to 3000 K.
  • Rigorous filtering ensured physical realism, removing spurious high-energy or short-distance configurations.

• Interaction Representation Design:

  • The UF3 framework utilizes cubic B-spline interpolations for two-body (2B) and three-body (3B) terms.
  • Cutoffs: A long-range 2B cutoff (0.001–9 Å) was employed to capture van der Waals (vdW) interactions essential for layered stability. 3B terms were defined to capture in-plane covalent bonding and angular correlations within the trigonal prismatic coordination.
  • Efficiency: The use of compact support B-splines ensures that the computational cost remains fixed and low, comparable to empirical potentials, regardless of the learned potential's complexity.

• Hyperparameter Optimization and Validation:

  • Over 7,000 candidate potentials were generated and optimized using UF3Tools (employing the Optuna TPE algorithm).
  • Validation Stages: Candidates were screened against DFT benchmarks for:
    1. Structural parameters and phase stability.
    2. Phonon spectra and finite-temperature stability (up to 1500 K).
    3. Defect and edge formation energies (critical for nucleation).
    4. Epitaxial Growth Capability: The final selection criterion was the ability to successfully simulate the crystallization of a new layer on a substrate while maintaining the vdW gap.

3. Key Contributions

• Development of UF3-MoS₂: The first ultra-fast MLIP specifically designed for multilayer MoS₂ that accurately captures both strong in-plane covalent bonding and weak interlayer vdW forces.
• Bridging the Speed-Accuracy Gap: The potential achieves near-DFT accuracy while running only ~2x slower than the fastest empirical potentials (REBO, SW), making it orders of magnitude faster than other MLIPs (like MACE or CHGNet) and DFT.
• First Non-Equilibrium Epitaxial Simulation: The paper reports the first successful MD simulation of MoS₂ homoepitaxial growth that reproduces experimental observations, a feat previously unattainable with existing potentials.

4. Key Results

A. Structural and Energetic Accuracy

• Lattice Constants & Phase Stability: The potential reproduces lattice parameters for 1H, 2H, and 3R phases with <2% error relative to DFT. It correctly predicts the relative stability ordering of phases (2H < 3R < 1H < 1T').
• Interlayer Binding: It accurately captures the basal plane surface energy (within ~2% error for 2H-MoS₂), a metric where most existing potentials fail significantly.
• Elasticity & Phonons: The model reproduces the highly anisotropic elastic tensor and phonon spectra with RMS deviations of ~0.5 THz, ensuring correct thermal and mechanical behavior.

B. Defect and Edge Energetics

• Defect Hierarchy: The potential shows a strong correlation ($R^2 = 0.91$) with DFT for the formation energies of ten different defect types (vacancies, antisites, complexes).
• Edge Stability: It accurately reproduces the energy difference between zigzag (ZZ) and armchair (AC) edges (within 5% of DFT), predicting that ZZ edges are energetically favored. This is crucial for predicting the equilibrium shape of growing domains.

C. Computational Performance

• Speed: In a benchmark of a 54,000-atom system, UF3 completed 40,000 timesteps in ~1.7 hours. In contrast, universal MLIPs (MACE, CHGNet) failed due to memory limits or took orders of magnitude longer (32–63 minutes for just 1,000 steps on a smaller 6,000-atom system).

D. Non-Equilibrium Homoepitaxial Growth Simulations

• Van der Waals Gap Preservation: Simulations of Mo and S atom deposition on a 2H-MoS₂ substrate successfully maintained the characteristic vdW gap between the substrate and the growing epilayer, preventing unphysical intermixing.
• Triangular Domain Formation: The simulations spontaneously formed triangular domains bounded by zigzag edges, consistent with experimental observations. This confirms the potential correctly captures the anisotropic edge energies driving morphology.
• Kinetic Dependence: Higher temperatures and slower deposition rates yielded higher quality, less defective triangular domains, demonstrating the potential's ability to model kinetic processes.

5. Significance

This work represents a major advancement in computational materials science for 2D materials:

1. Enabling Predictive Modeling: It provides the first tool capable of simulating the full epitaxial growth process of MoS₂ at the atomic scale, bridging the gap between theoretical mechanisms and experimental manufacturing.
2. Foundation for Heterostructures: By successfully modeling homoepitaxy, this potential lays the groundwork for developing MoS₂/substrate potentials, which are essential for designing heterostructure devices.
3. Methodological Impact: It demonstrates that the UF3 framework, by prioritizing physical interpretability and transferability over minimizing raw energy/force errors, can achieve the robustness required for highly non-equilibrium processes.

In summary, the UF3-MoS₂ potential solves the long-standing challenge of simulating layered material growth, offering a unique combination of DFT-level fidelity and empirical-level speed, thereby enabling the design and optimization of next-generation 2D semiconductor devices.