🔬 materials science

Fine-Tuned Machine-Learned Interatomic Potentials for Structural and Vibrational Properties of Twisted 2D Materials

This study demonstrates that fine-tuning universal machine-learned interatomic potentials, specifically the MACE model, is essential for achieving DFT-level accuracy in modeling the structural reconstruction and vibrational properties of twisted 2D van der Waals bilayers, successfully capturing strain landscapes and phonon spectra that align with experimental observations.

Original authors: Viet-Anh Tran, Viet-Hung Nguyen, Wei Chen, Gian-Marco Rignanese, Jean-Christophe Charlier

Published 2026-06-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Viet-Anh Tran, Viet-Hung Nguyen, Wei Chen, Gian-Marco Rignanese, Jean-Christophe Charlier

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have two sheets of incredibly thin, atom-thin paper (like graphene or other 2D materials). If you stack them perfectly on top of each other, they look like a single sheet. But if you twist one sheet slightly relative to the other, something magical happens: a giant, repeating pattern called a "moiré superlattice" appears, like the rippling interference pattern you see when you hold two window screens at a slight angle.

This paper is about figuring out exactly how these twisted sheets behave, but there's a catch: calculating the physics of these twisted sheets using standard supercomputer methods is like trying to count every grain of sand on a beach one by one. It takes too long and costs too much computing power.

Here is the story of how the authors solved this problem, explained simply:

1. The "Universal Translator" vs. The "Specialist"

The researchers started with a powerful, pre-trained AI model (called a "foundation model"). Think of this model as a universal translator who has read every book in the library. It knows a little bit about almost everything—chemistry, materials, and physics.

However, when they asked this universal translator to predict the behavior of these specific twisted sheets, it failed. It was like asking a general encyclopedia to perform heart surgery; it knew the basics of biology, but it didn't know the specific, delicate nuances required for this specific job. The model couldn't predict the tiny energy differences that decide how the atoms rearrange themselves.

The Solution: The team took this "universal translator" and gave it a crash course (fine-tuning) specifically on twisted sheets. They showed it thousands of specific examples of how these atoms behave. Now, the model isn't just a generalist; it's a specialist surgeon for twisted materials.

2. The "Rubber Sheet" Effect (Atomic Reconstruction)

When you twist these sheets, the atoms don't just stay in a rigid grid. They are like people in a crowded room trying to find the most comfortable spot.

The Goal: Atoms want to sit in specific, comfortable "stacking" positions (like sitting in a chair rather than standing on a table).
The Conflict: To get into these comfortable spots, the atoms have to stretch and squeeze the material, which costs energy (like stretching a rubber band).

The paper shows that the atoms compromise. They form large, comfortable "neighborhoods" (domains) where they sit perfectly, separated by narrow "highways" (soliton lines) where the stretching happens.

Graphene and Boron Nitride: These are stiff materials. They stretch a little bit to find comfort.
Molybdenum Disulfide (MoS2): This material is much softer (more compliant). It's like a rubber sheet compared to a stiff plastic sheet. Because it's softer, it stretches much more to find those comfortable spots, creating a much larger distortion.

3. The "Strain Map"

The researchers used their new, fine-tuned AI to draw a map of the stress (strain) in these materials.

The Neighborhoods: In the comfortable "domains," the material is almost perfectly relaxed, like a calm lake.
The Highways: In the narrow lines separating these domains, the stress is intense. It's like the traffic jam between two calm neighborhoods.
The Twist: As you twist the sheets tighter (smaller angle), these "comfortable neighborhoods" get bigger, and the "traffic jams" (strain lines) get more intense.

They also found that if you press down on these sheets from the top (vertical pressure), you force the atoms to squeeze together. This makes the "traffic jams" even worse, concentrating even more stress in those narrow lines.

4. The "Musical Instrument" (Vibrations)

Materials don't just sit still; they vibrate. These vibrations create sound waves (phonons) that can be measured with lasers (Raman spectroscopy).

The Problem with Old Models: Previous computer models predicted that these vibrations would be wildly different, suggesting the material was vibrating like a broken drum.
The New Reality: The fine-tuned AI predicted that the vibrations change in a very specific, subtle way that matches real-world experiments perfectly.
The Analogy: Imagine a guitar string. If you press on it in different spots (strain), the note it plays changes slightly. The paper shows that the "notes" (vibrations) of these twisted sheets change in a pattern that perfectly matches what scientists actually hear in the lab. This proves their new AI model is accurate.

Summary

The paper claims that to understand these twisted, atom-thin materials, you can't just use a general-purpose AI. You must fine-tune a general model with specific data. Once they did this, they could accurately predict:

How the atoms rearrange themselves into "neighborhoods" and "highways."
How much the material stretches (which depends on how soft the material is).
How the material vibrates, matching real experimental data.

This gives scientists a fast, accurate, and cheap way to study these complex materials without needing to run impossible supercomputer simulations.

Technical Summary: Fine-Tuned Machine-Learned Interatomic Potentials for Twisted 2D Materials

Problem Statement
Twisted van der Waals bilayers form moiré superlattices where structural and vibrational properties are highly sensitive to local stacking registry and atomic reconstruction. Accurately modeling these systems requires resolving energy differences between competing stacking configurations on the order of a few meV per atom. While Density Functional Theory (DFT) provides the necessary accuracy, its computational cost scales prohibitively with system size, making it infeasible for the large supercells required at small twist angles (where reconstruction is strongest). Conversely, classical interatomic potentials (e.g., Stillinger–Weber, REBO, Tersoff) combined with interlayer terms (e.g., Kolmogorov–Crespi) lack the flexibility to capture the subtle, stacking-dependent interlayer energetics, leading to inaccurate predictions of local deformation and vibrational properties. Although recent machine-learned interatomic potentials (MLIPs) based on foundation models (trained on broad chemical datasets) offer a promising middle ground, it remains unclear whether these general models possess the requisite precision for the specific energetics of twisted bilayers without further adaptation.

Methodology
The authors investigate the efficacy of fine-tuning universal atomistic foundation models, specifically the MACE (Many-body Atomic Cluster Expansion) architecture, for twisted bilayer systems. The study focuses on three representative materials: twisted bilayer graphene, hexagonal boron nitride (h-BN), and molybdenum disulfide (MoS₂), which span a range of mechanical compliance and interlayer energy corrugation.

The methodology involves a targeted fine-tuning procedure:

Dataset Construction: Starting from the MACE-OMAT foundation model (trained on the Open Materials 2024 dataset), the authors generate thermally distorted configurations via molecular dynamics (MD) at 300 K. These are supplemented with structures spanning relevant high-symmetry registries and intermediate translations.
Selection and Labeling: Representative configurations are selected using Smooth Overlap of Atomic Positions (SOAP) descriptors to ensure structural diversity. These structures are labeled with high-accuracy DFT energies and forces (PBE functional with D3 dispersion correction).
Fine-Tuning: The foundation MACE model is fine-tuned on this specific dataset by minimizing a combined energy–force loss function, prioritizing force accuracy (weighting ratio $\lambda_E : \lambda_F = 1:10$ ) to ensure high-quality structural relaxation.
Validation and Application: The fine-tuned models are validated against DFT for interlayer separations, stacking energetics, and phonon dispersions. They are then applied to simulate large-scale moiré superlattices at small twist angles.
Analysis Techniques: The study employs local strain tensor analysis to map deformation fields and a phonon band unfolding technique (using spectral-weight formalism) to analyze vibrational properties and compare them with experimental Raman data.

Key Contributions and Results

Necessity of Fine-Tuning: The study demonstrates that broadly trained foundation models are insufficient for twisted bilayer systems. While they capture general chemistry, they systematically overestimate interlayer distances (by up to 1.05 Å for h-BN) and fail to resolve the meV-scale energy differences between stacking registries. This leads to qualitatively incorrect reconstruction patterns. In contrast, the fine-tuned MACE models achieve near-DFT accuracy, reducing deviations in interlayer distances to <0.1 Å and accurately reproducing the energy landscape.
Reconstruction-Induced Strain Landscape: Using the fine-tuned potentials, the authors reveal a consistent strain landscape across all three materials at small twist angles:
- Domains: Energetically favorable stacking domains (e.g., AB/BA in graphene) relax toward ideal registries, remaining nearly strain-free with negligible bond distortion.
- Soliton Lines: Narrow regions separating these domains (soliton lines) concentrate intense deformation gradients and complex multi-component strain fields.
- AA Regions: High-symmetry AA stacking regions, topologically constrained at the intersections of soliton lines, undergo large local rotations (opposite to the surrounding domains) but retain negligible bond distortion.
- Material Dependence: The amplitude of reconstruction scales with mechanical compliance. MoS₂, having lower in-plane stiffness and deeper interlayer energy corrugation, exhibits significantly larger reconstruction amplitudes and strain fields compared to graphene and h-BN.
External Tuning: The models show that vertical compression selectively amplifies strain at the registry-transition (soliton) regions, demonstrating that the strain landscape is tunable by external mechanical perturbations.
Vibrational Properties and Experimental Agreement: The fine-tuned models successfully predict phonon spectra that align with experimental observations, unlike classical force fields which overestimate strain and predict unobserved large G-band splittings.
- Graphene/h-BN: The models predict a symmetry-breaking splitting of the G-band (E $_{2g}$ mode) consistent with micro-Raman experimental bounds (~~10 cm $^{-1}$ ), whereas classical potentials predicted unphysically large splittings (~~70 cm $^{-1}$ ).
- MoS₂: The models reproduce the emergence of the interlayer shear mode and the splitting of the layer-breathing mode as the twist angle decreases, matching experimental trends reported by Quan et al.
- Spatial Localization: The analysis of local phonon density of states reveals that different spectral features correspond to specific stacking regions (e.g., soliton lines vs. AB domains), providing a theoretical basis for nano-Raman hyperspectral imaging.

Significance
The paper establishes that fine-tuning foundation models on compact, system-specific datasets is an essential and data-efficient strategy for achieving DFT-level accuracy in modeling large-scale moiré systems. This approach overcomes the computational barriers of DFT and the accuracy limitations of classical potentials. By providing a reliable description of the delicate interlayer energetics governing atomic reconstruction, the methodology enables the quantitative investigation of strain landscapes and vibrational properties in twisted 2D materials. The results offer a validated theoretical framework that bridges the gap between atomistic modeling and experimental observations of reconstructed van der Waals interfaces, serving as a scalable blueprint for studying the complex energetics of twisted bilayers and heterostructures.