🔬 materials science

Phonon-Mediated Thermal Transport in Nanocrystalline Silicon Using Machine-Learning Interatomic Potentials

This study develops a machine-learning interatomic potential framework combining GAP and MACE models with lattice-dynamical and non-equilibrium molecular dynamics simulations to provide a more accurate and internally consistent description of phonon-mediated thermal transport and grain boundary resistance in nanocrystalline silicon compared to classical potentials.

Original authors: Houssem Rezgui, Catalina Coll Benejam, Miguel Pruneda, Clivia M. Sotomayor Torres

Published 2026-07-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Houssem Rezgui, Catalina Coll Benejam, Miguel Pruneda, Clivia M. Sotomayor Torres

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 are trying to move a crowd of people (heat) through a hallway. In a perfect, empty hallway (pure crystal silicon), the people can run freely and quickly. But in a real building, the hallway is often filled with obstacles, broken walls, and messy crowds (grain boundaries and defects). These obstacles slow the people down, making it hard to move heat efficiently. This is a huge problem for tiny computer chips, which get hot and need to cool down fast to work properly.

This paper is like a team of architects and engineers trying to figure out exactly how those obstacles slow down the crowd, and which set of blueprints (computer models) gives the most accurate prediction.

Here is the breakdown of their work:

1. The Problem with Old Blueprints

For a long time, scientists used "old blueprints" (classical physics models like Stillinger-Weber and Tersoff) to simulate how heat moves through silicon. Think of these old blueprints as simple, hand-drawn sketches. They are fast to draw, but they miss the fine details.

The Flaw: These old models assumed the obstacles were smoother and less chaotic than they really are. They predicted that the crowd could run faster than they actually can, leading to an overestimation of how well silicon conducts heat.

2. The New "AI" Blueprints

The authors of this paper decided to use a smarter tool: Machine-Learning Interatomic Potentials (MLIPs).

The Analogy: Imagine instead of a hand-drawn sketch, you use a high-definition 3D scanner combined with an AI that has studied millions of real buildings. This AI (specifically models called GAP and MACE) learns the exact shape of every brick and every crack.
The Result: When they used these AI models, they found that the "crowd" (heat) gets stuck much more easily at the messy boundaries than the old models suggested. The AI models showed that the obstacles are rougher and more chaotic, causing the heat to scatter and slow down significantly.

3. Testing the Models on "Pure" Silicon

Before testing the messy parts, they first checked if their new AI blueprints worked for a perfect, empty hallway (pure bulk silicon).

They compared the AI predictions against real-world experiments.
The Verdict: The AI models (GAP and MACE) matched the real-world data almost perfectly. They correctly predicted how fast the "people" run and how long they can keep running before getting tired. The old models, however, were too optimistic and predicted the heat would travel too far and too fast.

4. The Messy Hallway: Grain Boundaries

Next, they built a simulation of a "nanocrystalline" silicon block. This is like a wall made of many small tiles (grains) glued together. The lines where the tiles meet are called grain boundaries.

The Experiment: They created these walls with different angles of misalignment (some slightly crooked, some very crooked) and different levels of roughness (some smooth seams, some jagged, bumpy seams).
The Finding: They discovered that the rougher and more chaotic the seam, the harder it is for heat to cross.
The AI Advantage: When they used the old models, they didn't see much difference between a smooth seam and a rough one. But the AI models showed a dramatic difference: as the seam got rougher, the heat resistance skyrocketed. The AI models were sensitive enough to see that a bumpy wall blocks heat much more effectively than a smooth one.

5. The "Temperature Jump"

To measure the difficulty of crossing the boundary, they looked at the "temperature jump."

The Metaphor: Imagine a river flowing toward a dam. If the dam is easy to cross, the water level stays roughly the same. If the dam is a massive, clogged wall, the water piles up on one side, creating a huge difference in water level (temperature) across the dam.
The Result: The AI models predicted a much larger "temperature jump" (a bigger pile-up of heat) at the grain boundaries than the old models did. This confirmed that the AI models understand that these boundaries are major traffic jams for heat.

Summary

In short, this paper shows that to understand how heat moves through tiny, complex silicon structures, we can no longer rely on simple, old-fashioned math models. They are too optimistic and miss the chaos of the real world.

By using Machine Learning to create highly detailed, accurate "blueprints" of how atoms interact, the authors proved that grain boundaries act as massive roadblocks for heat. The rougher the boundary, the more heat gets stuck. This new, more accurate way of modeling is essential for designing better, cooler, and more efficient tiny electronic devices in the future.

Technical Summary: Phonon-Mediated Thermal Transport in Nanocrystalline Silicon Using Machine-Learning Interatomic Potentials

Problem Statement
Thermal management in nanoscale devices, particularly those involving nanocrystalline silicon, is fundamentally limited by phonon scattering at grain boundaries (GBs) and interfaces. While classical interatomic potentials (such as Stillinger–Weber and Tersoff) have enabled large-scale simulations, their restricted functional forms often fail to accurately capture vibrational properties, anharmonic interactions, and interfacial disorder. This leads to significant inaccuracies in predicting phonon dispersions, lifetimes, and thermal boundary resistance (TBR), especially in structurally complex environments where local atomic coordination deviates from ideal bulk symmetry. There is a critical need for interatomic models that offer near first-principles accuracy while retaining the computational efficiency required for large-scale, non-equilibrium simulations.

Methodology
The authors developed a unified machine-learning (ML) framework integrating Gaussian Approximation Potentials (GAP) and equivariant neural-network-based MACE models with lattice-dynamical calculations and non-equilibrium molecular dynamics (NEMD).

Model Training and Validation:
- A unified dataset of energies and forces was generated from bulk silicon configurations (diamond cubic structure) using Density Functional Theory (DFT) reference data.
- Four potentials were benchmarked: a retrained GAP model, a MACE model, and the empirical Stillinger–Weber (SW) and Tersoff potentials.
- Harmonic and anharmonic force constants were computed using the finite-displacement method (Phonopy/Phono3py) on 3×3×3 supercells. These were used to calculate phonon dispersions, densities of states (DOS), lifetimes, and lattice thermal conductivity ( $\kappa$ ).
Grain Boundary Modeling (Lattice Dynamics):
- Symmetric tilt grain boundaries were constructed with misorientation angles of 10°, 20°, 30°, and 40°.
- The system was partitioned into "near-GB" and "GB core" regions to analyze spatially resolved vibrational properties.
- A MACE potential, specifically trained on GB configurations, was employed to compute phonon lifetimes, mean free paths (MFPs), and group velocities within the disordered interface regions.
Interfacial Transport (NEMD):
- ML-NEMD simulations were performed in LAMMPS using the QUIP/QUIPY interface.
- Bicrystal systems with controlled interfacial roughness (sinusoidal displacements with amplitudes $A = 1, 2, 3$ Å) were subjected to a temperature gradient ( $T_{hot}=400$ K, $T_{cold}=300$ K).
- Thermal boundary resistance (TBR) was extracted from the steady-state temperature discontinuity ( $\Delta T_j$ ) and heat flux ( $J_x$ ).

Key Results

Bulk Validation: The MLIPs (GAP and MACE) accurately reproduced experimental phonon dispersions and DOS, correctly separating acoustic and optical branches. In contrast, SW and Tersoff potentials systematically overestimated optical-mode frequencies and underestimated anharmonic scattering. Consequently, MLIPs predicted phonon lifetimes and lattice thermal conductivity ( $\kappa$ ) in excellent agreement with experimental data and first-principles calculations, whereas classical potentials significantly overestimated $\kappa$ due to artificially long phonon lifetimes.
Grain Boundary Vibrational Properties: In nanocrystalline silicon, the GB core exhibited significant spectral modifications, including peak broadening and spectral weight redistribution, particularly in low-frequency acoustic modes. Phonon lifetimes in the GB core were drastically reduced to the picosecond scale across all frequencies, indicating strong scattering. While misorientation angles influenced the degree of disorder, the primary scattering mechanism was identified as local atomic disorder rather than the specific tilt angle.
Thermal Boundary Resistance (TBR): NEMD simulations revealed that TBR is highly sensitive to both interfacial roughness and the choice of interatomic potential.
- Roughness Dependence: Increasing roughness amplitude ( $A$ ) systematically increased TBR for all potentials.
- Potential Dependence: The GAP model predicted significantly higher TBR values (up to 1.66 m² K GW⁻¹ for $A=3$ Å) compared to SW (1.08) and Tersoff (1.17). This discrepancy arises because MLIPs capture stronger anharmonic scattering, leading to larger temperature jumps ( $\Delta T_j$ ) and lower effective heat fluxes.
- Experimental Consistency: The TBR values predicted by the GAP model for rough interfaces align well with experimental ranges (0.3–2.3 m² K GW⁻¹), whereas classical potentials tend to underestimate resistance in highly disordered morphologies.

Significance and Claims
The paper claims to establish a unified, internally consistent framework for modeling thermal transport that bridges the gap between accuracy and scale. By demonstrating that MLIPs provide a more physically realistic description of both harmonic and anharmonic lattice dynamics, the authors argue that predictive modeling of interfacial heat transport in low-dimensional materials requires moving beyond empirical potentials. The study highlights that the choice of interatomic potential significantly alters predictions of thermal boundary resistance, particularly for complex, rough interfaces. The work validates that ML-derived force constants can reliably capture the suppression of thermal conductivity in nanocrystalline silicon, offering a robust tool for the thermal engineering of next-generation nanomechanical and electronic devices.