🔬 materials science

NanoBTE: Fast Iterative Solution of the Phonon Boltzmann Transport Equation for Nanoscale Heat Transport

The paper introduces NanoBTE, a deterministic finite-volume solver for the phonon Boltzmann transport equation that handles complex 2D/3D geometries and boundary conditions while leveraging synthetic iterative schemes and hybrid CPU/GPU parallelization to efficiently simulate nanoscale heat transport.

Original authors: Hongjiang Chen, Hai-Xuan Lin, Xiaole Tian, Hongyu Chen, Juan Zhang, Shengyao Yan, Hezhu Shao, Yu Wu, Junjie Liu, Hao Zhang

Published 2026-07-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Hongjiang Chen, Hai-Xuan Lin, Xiaole Tian, Hongyu Chen, Juan Zhang, Shengyao Yan, Hezhu Shao, Yu Wu, Junjie Liu, Hao Zhang

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 understand how heat moves through a tiny, microscopic city inside a computer chip. In the big, macroscopic world (like a hot cup of coffee cooling down), heat moves predictably, like a crowd of people slowly shuffling through a wide hallway. This is what scientists call "diffusive" heat flow, and it's described by a simple rule known as Fourier's Law.

But in the nanoscale world of modern transistors (the switches that power our phones and computers), things get weird. The "hallways" are so narrow that the heat carriers—tiny particles called phonons (which are essentially vibrations of atoms)—can't shuffle slowly anymore. Instead, they zoom around like sprinters in a stadium, bouncing off walls and rarely bumping into each other. This is called the ballistic regime. When you mix these two behaviors, the old rules break down, and predicting where the heat goes becomes a nightmare.

This is where the paper introduces NanoBTE. Think of NanoBTE as a super-smart, high-speed traffic simulator designed specifically for these microscopic cities.

The Problem: The "Traffic Jam" of Heat

As computer chips get smaller, they pack more transistors into less space. This creates "hotspots" where heat gets trapped because the phonons (the heat messengers) are too small and fast to follow the usual rules. If you can't predict exactly where the heat goes, the chip can overheat, slow down, or even break.

The Solution: NanoBTE

The authors built a new software tool called NanoBTE to solve the Boltzmann Transport Equation (BTE). Don't let the name scare you; think of the BTE as the ultimate "traffic law" for phonons. It tracks not just where the heat is, but also which way it's moving, how fast it's going, and what kind of vibration it is.

Here is how NanoBTE works, using some everyday analogies:

1. The "Non-Gray" Map
Old simulators treated all heat like it was the same color (gray). But in reality, different phonons act like different types of vehicles: some are fast sports cars, some are slow trucks, and some take different routes. NanoBTE is "non-gray," meaning it tracks every specific type of phonon separately, giving a much more accurate map of the heat flow.

2. The "Fast-Forward" Button (Synthetic Iterative Scheme)
Simulating millions of tiny particles moving around is computationally heavy. It's like trying to count every single grain of sand on a beach one by one.

The Old Way: You count every grain, check the total, and repeat. This takes forever, especially when the sand is spread out (the "diffusive" regime).
NanoBTE's Trick: It uses a "Synthetic Iterative Scheme." Imagine you have a drone that flies over the beach and estimates the total sand volume based on the density, while a ground crew counts a few specific spots to correct the estimate. NanoBTE couples a detailed microscopic count with a fast, broad-brush macroscopic estimate. This allows the simulation to "fast-forward" through the slow parts and converge on the answer much faster.

3. The "Parallel Play" Strategy (CPU & GPU)
To make this run fast, NanoBTE splits the work like a massive construction crew.

The Strategy: Instead of one person doing all the work, the software breaks the problem into thousands of tiny, independent tasks (tracking specific phonon bands and directions).
The Hardware: It sends these tasks to both CPUs (the general workers) and GPUs (the specialized, high-speed assembly lines). Because the tasks are independent, they can all happen at the same time without getting in each other's way. This is like having 32 people counting sand grains simultaneously instead of one person doing it alone.

What They Tested It On

The team tested NanoBTE on three "cities" to prove it works:

A 2D Sheet (MoS2): Like a flat piece of graphene. They showed that when the sheet is tiny (10 nm), heat behaves like a sprinter bouncing off walls. When it's larger (100 microns), it behaves like a slow shuffle. NanoBTE accurately captured this transition.
A FinFET (The "Fin"): A common 3D transistor shape that looks like a vertical fin. They simulated heat generated in a tiny spot on the fin. As the fin got thinner, the heat got "stuck" and the temperature rose sharply because the phonons couldn't escape easily.
A GAAFET (The "Stack"): A newer, more complex transistor with multiple stacked layers. This is like a multi-story building where heat is generated on different floors. NanoBTE successfully tracked how heat moved between these stacked layers and the surrounding material.

The Results

Accuracy: NanoBTE matched the results of an existing, trusted code (GiftBTE) perfectly, proving it gets the physics right.
Speed: It was significantly faster. On a standard computer, it was up to 14 times faster than the old code when using many processors.
GPU Power: When they used powerful graphics cards (GPUs), the simulation time dropped to just 2–3 minutes for complex 3D devices, making it practical for engineers to use in real design cycles.

In a Nutshell

NanoBTE is a new, high-speed, super-accurate traffic controller for heat in the microscopic world. By treating heat particles individually and using modern computer power to run millions of calculations at once, it allows scientists to predict exactly how hot a tiny computer chip will get, even when the heat behaves in strange, non-intuitive ways. This helps engineers design better, cooler, and more reliable next-generation electronics.

Technical Summary: NanoBTE

Problem Statement

As semiconductor device dimensions shrink into the deep-submicrometer and nanometer regimes (e.g., FinFETs and GAAFETs), self-heating has emerged as a critical bottleneck. In these scales, the characteristic device length becomes comparable to or smaller than the phonon mean free paths (MFPs). Consequently, heat transport deviates significantly from the classical Fourier description, entering the Casimir-Knudsen regime where transport is non-local and involves a coexistence of diffusive, quasi-ballistic, and ballistic phonons.

Accurately modeling this regime requires solving the phonon Boltzmann transport equation (BTE). However, the numerical solution of the BTE is computationally demanding due to its high-dimensional phase space (dependence on physical space, propagation direction, frequency/band, and polarization). Existing methods face trade-offs: stochastic Monte Carlo (MC) methods handle complex geometries but suffer from statistical noise and slow convergence in near-diffusive regimes, while deterministic methods like the discrete ordinate method (DOM) avoid noise but struggle with high-dimensional discretization and ray effects. Furthermore, achieving convergence across the ballistic-to-diffusive transition and efficiently handling complex 3D device geometries remain significant challenges.

Methodology

The authors present NanoBTE, a deterministic finite-volume solver designed for the steady-state, non-gray phonon BTE under the relaxation-time approximation (RTA). The solver is built to handle complex two- and three-dimensional geometries using unstructured meshes generated by COMSOL and Gmsh.

Key methodological components include:

Phase Space Discretization: The solver employs band-resolved phonon properties (derived from first-principles calculations via Phonopy and ShengBTE) and discrete-ordinates angular quadrature to resolve the phonon distribution function.
Boundary Conditions: It supports thermalizing, diffuse, and specular boundary conditions, allowing for the direct capture of surface scattering and localized self-heating effects.
Iterative Schemes: To address convergence issues in multiscale regimes, NanoBTE implements two strategies:
1. Sequential Iteration: Solves the steady BTE and updates the lattice temperature via local energy conservation. While robust, this can be slow in near-diffusive regimes.
2. Synthetic Iterative Scheme (SIS): Accelerates convergence by coupling the microscopic BTE with a macroscopic diffusion-type temperature equation. This allows long-range temperature information to propagate efficiently, overcoming the slow convergence of purely local updates in diffusive limits.
Parallelization Strategy: The solver exploits the natural separability of the discretized BTE in the band-direction phase space. Independent transport tasks for each (band, direction) pair are distributed among MPI ranks. This "band-direction task decomposition" avoids complex spatial domain decomposition, simplifying the handling of unstructured meshes.
Hardware Acceleration: Both CPU and GPU implementations are provided. The dominant sparse transport operations are offloaded to GPUs, while the CPU manages geometry, boundary conditions, and residual construction.

Key Results

The NanoBTE solver was validated against the established GiftBTE code and tested on several benchmark cases spanning different transport regimes:

2D Benchmarks (MoS2 and Diamond):
- Simulations of in-plane MoS2 and cross-plane diamond demonstrated the solver's ability to capture the transition from ballistic to diffusive transport as the characteristic length ( $L$ ) increased from 10 nm to 100 $\mu$ m.
- At small scales ( $L=10$ nm), the results showed clear temperature jumps at boundaries and flattened interior profiles, characteristic of quasi-ballistic transport. As $L$ increased, the profiles converged to linear Fourier-like distributions.
3D Device Simulations (FinFET and GAAFET):
- FinFET: Simulations of silicon FinFETs with fin widths of 10 nm, 100 nm, and 1 $\mu$ m revealed strong size dependence. The 10 nm case exhibited a peak temperature rise of ~~110 K due to suppressed effective thermal conductivity from boundary scattering, whereas the 1 $\mu$ m case showed a much smoother, lower temperature rise (~~6 K).
- GAAFET: A three-stacked nanosheet Gate-All-Around FET was simulated to test complex geometry and multi-hotspot coupling. The results confirmed that reducing nanosheet dimensions enhances boundary scattering and local temperature rise, with the 5 nm case showing a peak temperature of ~340 K. The solver successfully handled the thermal coupling between vertically stacked channels.
- In all 3D cases, NanoBTE results showed good agreement with GiftBTE in both temperature and heat-flux distributions.
Performance Benchmarks:
- CPU Scaling: On a dual-socket Intel Xeon Gold 6133 platform, NanoBTE demonstrated superior scalability compared to GiftBTE. For the 10 nm MoS2 case, runtime decreased from 629 s (1 core) to 100 s (32 cores), achieving a speedup of ~14.43 $\times$ relative to GiftBTE at 32 cores.
- GPU Acceleration: On a workstation with two NVIDIA RTX 4090 GPUs, the GPU-accelerated version achieved runtimes of approximately 2–3 minutes for all benchmarks. The optimal configuration was found to be 16 MPI ranks sharing two GPUs; beyond this, oversubscription led to performance degradation due to memory contention.

Significance and Claims

The paper claims that NanoBTE provides an efficient and versatile framework for simulating phonon heat transport in realistic nanoscale semiconductor devices. Its primary contributions are:

Deterministic Efficiency: It offers a deterministic alternative to MC methods, avoiding statistical noise while maintaining accuracy in complex 3D geometries.
Multiscale Convergence: The implementation of the synthetic iterative scheme effectively accelerates convergence in near-diffusive regimes, a known bottleneck for standard BTE solvers.
Scalability: The band-direction task decomposition strategy enables efficient parallelization on both CPUs and GPUs, making high-fidelity BTE simulations feasible for complex transistor architectures like FinFETs and GAAFETs.

The authors position NanoBTE as a tool capable of addressing the critical challenge of self-heating in next-generation semiconductor devices, where classical Fourier models fail and microscopic transport effects dominate. The solver is validated to be accurate across a wide range of characteristic length scales, from the ballistic limit to the diffusive regime.