A Unified Multiscale Auxiliary PINN Framework for Generalized Phonon Transport

This paper introduces MTNet, a unified multiscale auxiliary physics-informed neural network framework that overcomes the computational and numerical limitations of traditional solvers by recasting the generalized phonon radiative transfer equation into a fully differential system, thereby enabling high-fidelity simulation of ballistic-diffusive transport and the solution of geometric inverse problems in nanoscale thermal systems.

The Gist

The Big Picture: Why Heat is Hard to Predict at the Tiny Scale

Imagine you are trying to predict how a crowd of people moves through a hallway.

• In a huge stadium (Macroscopic scale): The crowd moves like a fluid. You can easily predict the flow using simple rules (like "people move from hot to cold"). This is what traditional physics (Fourier's Law) does.
• In a narrow hallway (Nanoscale scale): The hallway is so narrow that the people (which represent phonons, or heat-carrying particles) start bumping into the walls more than they bump into each other. They bounce around chaotically. The simple "fluid" rules break down. You need to track every single person's path to know where the heat goes.

This is the problem scientists face with modern computer chips. As they get smaller, heat gets stuck, creating "hot spots" that can melt the device. The math to describe this chaotic movement is called the Boltzmann Transport Equation (BTE). It's incredibly complex, like trying to solve a puzzle where the pieces keep changing shape and there are billions of them.

The Old Way: The "Grid" Trap

Traditionally, scientists tried to solve this by drawing a grid over the hallway and calculating the movement at every intersection.

• The Problem: At the nanoscale, the "grid" needs to be so incredibly fine that the calculation takes forever, even on supercomputers.
• The Shortcut: To make it faster, they used approximations (like assuming everyone moves at the same speed). But this is like assuming everyone in the crowd walks at a leisurely stroll; it misses the sprinters and the people running into walls, leading to inaccurate predictions.

The New Solution: MTNet (The "Smart Guide")

The authors, Roberto Riganti and Luca Dal Negro, created a new tool called MTNet. Think of it as a super-smart, self-learning guide that doesn't need a grid.

Here is how it works, broken down into three simple concepts:

1. The "Auxiliary" Trick (Turning a Math Monster into a Puzzle)

The main equation involves a "scattering operator," which is a giant math term that calculates how particles bounce off each other. In standard math, this is like trying to add up a list of a billion numbers every time you take a step. It's slow and clunky.

• The Analogy: Imagine you are trying to calculate the total weight of a bag of marbles. The old way is to weigh every single marble one by one.
• The MTNet Way: They invented a "helper" (an auxiliary variable) that acts like a pre-calculated scale. Instead of weighing every marble, the AI learns to predict the total weight directly by looking at the rate at which the weight changes.
• The Result: They turned a difficult "summing" problem into a smooth "rate of change" problem. This allows the computer to use Automatic Differentiation (a super-fast math trick) to solve the equation instantly without needing a grid.

2. The "Multiscale" Vision (Seeing the Forest and the Trees)

In a nanoscale hallway, some heat particles move very slowly (like a slow walker), while others zip by at lightning speed (like a sprinter).

• The Problem: Standard AI models are "spectrally biased." They are good at seeing the slow, smooth trends (the forest) but terrible at seeing the fast, sharp spikes (the trees). They tend to smooth out the details, missing the critical "bounces" off the walls.
• The MTNet Fix: They built the AI with a "multiscale" brain. Imagine a camera that has both a wide-angle lens (to see the whole room) and a high-speed macro lens (to see the tiny details of a bouncing particle) simultaneously. This allows the AI to capture both the smooth flow of heat and the sharp, chaotic bounces at the same time.

3. The "Inverse" Superpower (Guessing the Unknown)

Usually, you know the size of the hallway and the temperature, and you want to know how much heat flows.

• The Inverse Problem: What if you don't know the size of the hallway? What if you only know the temperature at the two ends?
• The MTNet Magic: The AI can work backward. It looks at the temperature difference at the ends and asks, "What size hallway would cause this specific temperature pattern?"
• Real World Use: This is like a doctor diagnosing a broken bone just by looking at the skin temperature, without needing an X-ray. The authors showed that MTNet could accurately guess the thickness of a silicon film just by looking at the temperature at the edges.

Why This Matters

This paper isn't just about better math; it's about building better technology.

• For Engineers: It allows them to design computer chips that don't overheat, even as they get smaller than a human hair.
• For Materials Science: It helps figure out the properties of new, weird materials without having to build them and test them physically first.
• Speed: Because it runs on multiple graphics cards (GPUs) at once and doesn't need a grid, it solves these problems much faster than traditional supercomputers.

The Bottom Line

The authors built a smart, grid-free, multi-lens AI that can predict how heat moves through tiny materials. It solves the "impossible" math of bouncing particles, handles extreme temperature changes, and can even guess the size of a material just by looking at its surface temperature. It's a new toolkit for keeping our future electronics cool and efficient.

Technical Summary

1. Problem Statement

Nanoscale thermal transport is governed by the Phonon Boltzmann Transport Equation (BTE), an integro-differential equation (IDE) describing the evolution of particle distribution functions. Solving the BTE is critical for designing next-generation microprocessors, thermoelectric generators, and quantum materials, where traditional macroscopic models (like Fourier's law) fail due to mesoscopic and quasi-ballistic effects.

However, solving the BTE presents three major computational challenges:

1. High Dimensionality: The phase space involves spatial, momentum, and temporal dimensions, making analytical solutions intractable.
2. Integro-Differential Nature: The collision operator contains integral terms representing scattering. Traditional Physics-Informed Neural Networks (PINNs) struggle with these because they rely on deterministic numerical quadrature, which couples spatial and momentum points, reintroducing grid-based dependencies and limiting parallelization.
3. Multiscale Complexity & Spectral Bias: Phonon mean free paths (MFPs) span several orders of magnitude. Standard neural networks suffer from "spectral bias," converging quickly on low-frequency (macroscopic) features while failing to resolve high-frequency (sharp, localized) phenomena near boundaries or interfaces.
4. Nonlinearity: Under large temperature gradients ($\Delta T$), the standard linearization of the Bose-Einstein distribution fails, and scattering rates become highly nonlinear, causing traditional solvers to break down.

2. Methodology: The MTNet Framework

The authors propose MTNet (Multiscale Thermal Network), a unified framework that combines an Auxiliary formulation with a Multiscale PINN architecture to solve the Generalized Equation of Phonon Radiative Transfer (GEPRT).

A. Auxiliary Formulation (Converting IDE to ODE/PDE)

To eliminate the need for numerical quadrature, the authors recast the GEPRT into a purely differential system. They introduce three auxiliary variables ($\hat{f}, \hat{v}, \hat{g}$) defined by differential equations that approximate the integral terms:

• Scattering Term: The integral over the scattering phase function is replaced by an auxiliary variable $\hat{f}$ governed by a differential equation with respect to the angular coordinate.
• Radiative Equilibrium: The integral over frequency and angle required to close the system (energy balance) is replaced by auxiliary variables $\hat{v}$ and $\hat{g}$, governed by differential equations with respect to frequency and angle, respectively.
• Result: The original integro-differential equation becomes a system of coupled partial differential equations (PDEs) that can be solved entirely using Automatic Differentiation (AD), removing grid-based discretization errors.

B. Multiscale Architecture

To address spectral bias and capture sharp discontinuities (e.g., at $\mu=0$ in the angular distribution), the network employs a Multiscale Sine Feature Mapping.

• Intensity Network: A deep network with sine feature mapping and $x \cdot \sigma(x)$ activation functions to resolve high-frequency variations in the phonon intensity.
• Temperature Network: A shallow network with hard constraints to ensure the temperature $T$ remains physically valid (avoiding division by zero in the Bose-Einstein distribution).
• Auxiliary Network: Mirrors the intensity network but embeds boundary conditions via hard constraints to reduce loss terms.

C. Distributed Training

The framework is designed for multi-GPU parallelization. By decoupling the collocation points from the integral operator, the training workload can be distributed across GPUs using TensorFlow's MirroredStrategy. This allows for massive batch sizes and efficient handling of the high-dimensional phase space.

3. Key Contributions

1. Quadrature-Free Solver: Successfully transforms the GEPRT into a fully differential system, enabling analytical evaluation of scattering operators via AD and eliminating the computational bottlenecks of numerical integration.
2. Multiscale Resolution: Demonstrates the ability to resolve sharp angular discontinuities and ballistic-diffusive transitions that standard PINNs miss due to spectral bias.
3. Generalized Scattering: Unlike the Relaxation Time Approximation (RTA), which treats all scattering as dissipative, MTNet explicitly distinguishes between elastic (directional redistribution, energy conserving) and inelastic (dissipative) scattering, leading to more accurate thermal conductivity predictions.
4. Inverse Problem Capability: Shows the ability to solve geometric inverse problems (retrieving unknown slab thickness) using only boundary temperature data, without requiring internal volumetric measurements.

4. Results

The framework was validated on a silicon thin film under various conditions:

• Validation (RTA vs. Literature): In the diffusive regime ($L=10\mu m$) and mesoscopic regime ($L=100nm$), MTNet reproduced known temperature profiles and boundary slips, matching literature values for effective thermal conductivity ($k_{eff}$).
• GEPRT vs. RTA: When solving the full GEPRT (without RTA simplification), MTNet predicted a higher effective thermal conductivity ($42$ W/m/K vs. $30$ W/m/K in RTA). This is because the GEPRT correctly conserves energy during elastic scattering, whereas RTA artificially thermalizes these carriers.
• Large Temperature Gradients ($\Delta T = 100$ K): MTNet successfully simulated transport under extreme non-equilibrium conditions where linearization fails. It captured:
  • Asymmetric temperature profiles due to temperature-dependent scattering rates.
  • "S-shaped" curvature and boundary slips in the mesoscopic regime.
  • The model converged monotonically, proving physical consistency without relying on surrogate linearizations.
• Inverse Problem: The network successfully retrieved the unknown film thickness ($L$) using only the measured temperatures at the two boundaries. This demonstrates the framework's utility for non-destructive thermal metrology where internal data is inaccessible.

5. Significance

This work establishes a robust, fully differentiable foundation for kinetic transport modeling.

• Scalability: The mesh-free, multi-GPU approach makes high-fidelity BTE solving accessible on standard workstations, not just supercomputers.
• Engineering Impact: It provides a tool for optimizing thermal management in nanoelectronics and thermoelectrics by accurately predicting heat dissipation in regimes where traditional models fail.
• Metrology: The inverse problem capability offers a new pathway for characterizing nanomaterials (e.g., extracting MFPs or layer thicknesses) using only surface-level experimental data, bridging the gap between simulation and non-destructive testing.
• Future Applications: The framework is generalizable to anisotropic materials and complex heterostructures, paving the way for the design of advanced nanocomposites and quantum materials.