🔬 optics

A Fourier Neural Operator for Accelerated Discretization-Invariant Solutions of the Radiative Transfer Equation

This paper presents a Fourier Neural Operator (FNO) surrogate model that achieves up to 26-times computational acceleration over Monte Carlo simulations while providing accurate, discretization-invariant predictions of heat flux in participating media by learning the underlying solution operator rather than a simple image-to-image mapping.

Original authors: Daniel Carne

Published 2026-08-24

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Daniel Carne

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

Thermal radiation is the invisible heat that moves through space, carrying energy from a hot flame to a cold wall or from the sun to the Earth. In many complex environments, such as the swirling gases inside a jet engine or the thick clouds in our atmosphere, this heat does not travel in a straight line. Instead, it bounces off particles, gets absorbed, and is re-emitted in a chaotic dance of energy. To predict how this heat moves, scientists rely on a complex mathematical rule known as the radiative transfer equation. This rule is the gold standard for understanding thermal management in high-tech systems, but it is notoriously difficult to solve. Because the heat depends on where it is, where it is going, and what color it is, the calculations become so massive that even the fastest supercomputers struggle to run them quickly. This bottleneck slows down the design of everything from cleaner engines to better climate models, forcing engineers to either wait too long for answers or use rough approximations that miss the fine details.

A researcher at the United States Naval Academy has found a way to bypass this computational wall using a new type of artificial intelligence. Instead of trying to solve the heavy math every time a new scenario arises, they trained a machine to learn the underlying pattern of how heat moves. They created a vast library of examples using a method called Monte Carlo simulation, which tracks millions of tiny packets of light as they bounce around a virtual room filled with varying materials. In these simulations, the researcher randomly changed the shape of a heat source, the temperature of the walls, and the transparency of the air, generating thousands of unique scenarios. They then fed this data into a specialized neural network called a Fourier neural operator. Unlike standard image-recognition programs that simply match one picture to another, this network learned the fundamental rule that connects the conditions of the room to the resulting heat map. It learned to see the solution as a continuous flow rather than a fixed grid of pixels.

The results of this training were striking. Once the network learned the pattern, it could predict the heat distribution for a new scenario almost instantly. When the researcher compared the speed of their new AI model against the traditional simulation method, they found the AI was up to twenty-six times faster while maintaining the same level of accuracy. This speedup is not just a matter of convenience; it means that problems which previously took hours to solve can now be addressed in seconds, opening the door for rapid design optimization and real-time analysis. Perhaps more importantly, the model proved to be incredibly flexible. Usually, a computer program trained on a specific grid size, like a 64-by-64 pixel map, fails when asked to look at a finer or coarser grid. This new model, however, worked perfectly on grids of all sizes, from coarse 32-by-32 maps to extremely detailed 512-by-512 maps, without needing to be retrained. It had learned the physics of the heat flow itself, not just the resolution of the training images.

The researcher also looked closely at the quality of the predictions to understand what the machine was actually doing. The traditional simulation method, while accurate, often produces a grainy, noisy result because it relies on random sampling, much like static on an old television screen. The new AI model, however, acted as a powerful filter. It successfully ignored the random high-frequency noise inherent in the training data and focused on the smooth, physically meaningful patterns of heat flow. This allowed it to produce clean, clear predictions even when the training data was somewhat noisy. The only place where the model struggled was at sharp edges, such as the boundary of a solid object where the heat changes abruptly. In these specific spots, the model smoothed out the details slightly, a trade-off that is acceptable given the massive gains in speed and flexibility.

This work suggests a new path forward for solving some of the most difficult problems in thermal physics. By teaching a machine to understand the operator—the rule that transforms inputs into outputs—rather than just memorizing specific cases, the researcher has created a tool that generalizes well across different shapes, materials, and resolutions. The model is not a magic wand that solves every problem instantly, but it is a robust and efficient surrogate that can handle the heavy lifting of radiative heat transfer. For engineers designing combustion systems or studying atmospheric heating, this means they can now run thousands of simulations in the time it used to take to run just one, allowing them to explore a much wider range of possibilities and arrive at better solutions faster.

Technical Summary: A Fourier Neural Operator for Accelerated Discretization-Invariant Solutions of the Radiative Transfer Equation

Problem Statement

The Radiative Transfer Equation (RTE) governs thermal radiation in participating media, a process critical to modeling combustion, atmospheric science, high-temperature systems, and radiative thermal management. However, solving the RTE is computationally expensive due to its inherently high-dimensional nature, which depends on spatial position, propagation direction, and wavelength. Traditional deterministic methods (e.g., discrete ordinates) require extensive spatial, angular, and spectral discretization, while stochastic Monte Carlo (MC) methods require massive numbers of photon histories to reduce statistical noise. These costs become prohibitive in coupled thermal-fluid problems, transient simulations, design optimization, and inverse problems where the RTE must be solved repeatedly.

Existing machine learning (ML) approaches for radiative transfer often suffer from limitations:

Parameter-based surrogates (e.g., fully connected networks) are efficient for fixed geometries but lack generalizability to new geometries or problem formulations.
Image-to-image models (e.g., CNNs, U-Nets) used for denoising MC predictions retain the computational cost of the underlying MC simulation and are often tied to specific spatial discretizations and resolutions.

There is a need for a surrogate model that is fast, generalizable across different spatial discretizations, and capable of learning the underlying solution operator rather than a fixed-resolution mapping.

Methodology

Fourier Neural Operator (FNO) Architecture

This work implements a Fourier Neural Operator (FNO) to approximate the solution operator of the RTE. Unlike standard neural networks that map fixed-dimensional arrays, the FNO learns mappings between function spaces, making it "discretization-invariant."

Inputs: Five discrete input fields representing the scattering coefficient ( $\mu_s$ ), absorption coefficient ( $\mu_a$ ), asymmetry parameter ( $g$ ), geometry, and temperature ( $T$ ).
Output: The absorbed heat flux field ( $q''$ ).
Structure: The architecture consists of:
1. A lifting layer ( $P$ ) that projects inputs to a higher-dimensional latent space.
2. Four Fourier layers, each containing a local pointwise path and a nonlocal Fourier integral path. The integral path operates in Fourier space, parameterizing the integral operator via convolution to efficiently model nonlocal interactions. High-frequency modes are truncated to reduce cost.
3. A projection layer ( $Q$ ) that maps the latent representation back to the output space.
Training: The model is trained to minimize the mean squared error (MSE) between the predicted heat flux and the ground truth generated by Monte Carlo simulations.

Dataset Generation

A benchmark 2D radiative transport problem was constructed to generate training and testing data:

Domain: A square domain with periodic boundary conditions containing a solid rectangular emitting element embedded in a participating medium.
Property Variation: Spatially varying optical properties ( $\mu_a, \mu_s, g$ ) were generated as smooth continuous fields using the superposition of Gaussian radial basis functions (background and localized blobs).
Ground Truth: Monte Carlo simulations were used to generate the absorbed heat flux maps. The training set used 500,000 photons per simulation, while the test set used 5,000,000 photons to ensure high-fidelity ground truth.
Spectral Analysis: A discrete Fourier transform was applied to both MC solutions and FNO predictions to analyze spatial frequency content and noise characteristics.

Key Results

Performance and Scaling

Data and Model Scaling: The model exhibits power-law scaling with dataset size up to ~120 simulations, after which performance saturates, indicating a transition to a model-size-limited regime. Increasing the FNO width (number of parameters) beyond 8 yields diminishing returns, suggesting the model is data-limited at higher widths.
Noise Robustness: Training with low photon counts leads to overfitting to stochastic MC noise. However, as photon counts increase, the test loss stabilizes. The FNO demonstrates an ability to filter high-frequency noise present in the training data while preserving the physically meaningful low-to-mid frequency spectrum.

Computational Acceleration

The trained FNO achieves up to 26× computational acceleration compared to Monte Carlo simulations at equivalent error levels.
The FNO provides lower error at comparable computational costs or significantly lower costs at comparable errors compared to MC.

Discretization Invariance

Generalization: A model trained on a specific grid resolution (e.g., $64 \times 64$ ) can accurately predict heat flux on coarser ( $32 \times 32$ ) and finer ( $128 \times 128$ , $512 \times 512$ ) grids without retraining.
Resolution Requirements: While the FNO is discretization-invariant, the training resolution must be sufficiently refined to capture the underlying physical features. Models trained on coarse grids ( $32 \times 32$ ) showed significant error increases when tested on finer grids. In contrast, models trained on higher resolutions ( $64 \times 64$ and $128 \times 128$ ) exhibited significantly lower increases in prediction error when tested on higher grid resolutions.
Energy Conservation: The FNO maintained high energy conservation accuracy (maximum error of 0.63% for the $128 \times 128$ trained model) across different grid resolutions, despite no explicit enforcement of energy conservation during training.

Spectral Analysis

Spectral analysis revealed that the FNO preserves low-mid frequency energy (physically meaningful structures) identical to high-photon-count MC solutions.
The FNO significantly suppresses high-frequency energy (Nyquist region) associated with Monte Carlo statistical noise. This indicates the model learns the true solution operator rather than memorizing the noise realization of the training data.
Limitation: The FNO struggles to capture high-frequency events at known discontinuities, such as geometry boundaries or rapid changes in volumetric optical properties.

Significance and Claims

The author claims that this work demonstrates the potential of Fourier Neural Operators as fast, accurate, and discretization-invariant surrogate solvers for radiative transfer. By learning the underlying solution operator rather than a fixed image-to-image mapping, the FNO approach moves toward the development of generalizable surrogate models capable of handling various geometries, discretizations, and boundary conditions.

The study highlights that the FNO architecture offers a promising foundation for reducing the computational burden of radiative transfer simulations, particularly for applications requiring repeated evaluations such as coupled thermal-fluid problems and optimization. The ability to generalize across spatial resolutions without retraining addresses a key limitation of previous CNN-based approaches.

Future Directions

The paper outlines specific future directions without overextending claims:

Angular and Source Expansion: Extending the architecture to include angular inputs/outputs (e.g., radiative source terms, boundary conditions, and intensity) to handle diverse boundary conditions (diffuse, specular).
Superposition of Sources: Leveraging the linearity of the RTE (a linear Boltzmann equation) to train separate networks for different source types (e.g., surface emission vs. volumetric emission) and superimpose their outputs, potentially reducing model size and training data requirements.
Physics Integration: Incorporating physics constraints or error estimation by training models to predict discretized radiative intensity and applying the discrete ordinates method to enforce physics and calculate residuals.