Nested Fourier-enhanced neural operator for efficient modeling of radiation transfer in fires

This paper introduces a nested Fourier-enhanced neural operator (Fourier-MIONet) framework that efficiently and accurately models radiation transfer in 3D fire simulations, achieving global relative errors of 2–4% while significantly reducing computational costs compared to traditional numerical methods.

The Gist

The Big Picture: Predicting Fire Without the Headache

Imagine you are trying to predict how a fire will behave in a building. You need to know exactly how heat moves, especially radiation (the heat you feel on your skin from a flame, even if you aren't touching it).

To do this accurately, scientists use super-computers to run complex simulations. However, calculating how radiation travels through smoke and flames is like trying to count every single grain of sand on a beach while the tide is coming in. It's incredibly accurate, but it takes so long that it slows down the whole simulation. It's the "bottleneck" that stops engineers from running detailed, real-time fire safety tests.

The Solution: The authors of this paper built a "smart shortcut." They created an AI model that acts like a crystal ball for fire radiation. Instead of doing the heavy math every time, the AI looks at the fire's temperature and smoke density and instantly "guesses" the radiation pattern with high accuracy.

---

The Problem: Why Old Methods Are Too Slow

Think of the traditional way of calculating fire radiation like a very thorough, slow-moving librarian.

• The librarian (the computer) has to check every single book (every point in the 3D space) and every single direction the light could travel.
• If the fire is huge or the smoke is thick, the librarian gets overwhelmed.
• To make it faster, people used to use "rough approximations" (like guessing the average temperature of the whole room), but that's like trying to navigate a maze by only looking at the map's edges—you miss the tricky turns and sharp corners where the fire actually behaves dangerously.

The Innovation: The "Fourier-MIONet" (The Super-Reader)

The researchers built a new type of AI called a Fourier-MIONet. Here is how it works, broken down into three simple concepts:

1. The "Fourier" Part: Seeing the Music in the Fire

Fire isn't just smooth; it's turbulent. It has big swirls and tiny, sharp flickers.

• The Analogy: Imagine a song. A standard AI might only hear the bass line (the low, smooth notes). But fire has high-pitched squeaks and sharp cracks too.
• The Fix: The "Fourier" part of their AI is like a musician who can hear every instrument. It uses math (Fourier transforms) to instantly recognize both the big, smooth waves of heat and the tiny, sharp details of the flames. This allows it to see the "fine print" of the fire that other models miss.

2. The "Nested" Part: The Russian Doll Strategy

Real fires are simulated on computer grids that change size. Some areas (like right inside the flame) need a super-detailed grid (like a high-resolution photo), while the rest of the room can be a blurry sketch.

• The Problem: If you try to use one giant AI to look at the whole room at once, it gets confused by the mix of blurry and sharp details. It's like trying to read a newspaper where some words are written in giant font and others are microscopic.
• The Fix: They built a Nested system, like a set of Russian nesting dolls.
  • The Outer Doll (Coarse Level): A small, fast AI looks at the whole room to get the "big picture."
  • The Inner Dolls (Fine Levels): As you zoom in on the fire, a slightly more detailed AI takes the "big picture" from the outer doll and adds the fine details.
  • The Result: The AI doesn't have to do the hard work on the whole room; it only does the heavy lifting where it's needed (inside the flame), while using the "big picture" to guide the rest.

3. The "Generalist" (One Model for All Fires)

Usually, if you train an AI on a small fire, it fails when you show it a huge fire. It's like teaching a student to solve math problems only with the number 5; they get confused when you ask about 500.

• The Breakthrough: The team trained their AI on fires ranging from a small candle (10 kW) to a massive warehouse blaze (60 kW).
• The Result: They created a universal fire expert. This single AI can look at a small fire or a giant inferno and predict the radiation accurately without needing to be retrained. It learned the principles of fire, not just specific examples.

The Results: Fast, Accurate, and Practical

How well does this "crystal ball" work?

• Speed: The traditional method takes about 0.24 seconds to calculate radiation for one moment in a simulation. Their AI does it in 0.04 seconds. That's 6 times faster.
• Accuracy: The AI's guess is off by only 2–4% compared to the slow, perfect calculation. In the world of fire safety, that is incredibly close.
• Why it matters: Because the AI is so fast, engineers can now run simulations that include more detailed physics. They can model how different colors of light (spectral radiation) interact with smoke, which was previously too slow to do. This leads to better building designs, safer sprinkler systems, and more accurate evacuation plans.

Summary

The paper introduces a super-smart, multi-layered AI that acts as a shortcut for calculating fire heat. By using "musical" math to see details and a "nesting doll" approach to handle different scales, it predicts how fire radiates heat 6 times faster than current methods, with almost no loss in accuracy. This allows fire safety engineers to run better, faster, and more realistic simulations to save lives.

Technical Summary

1. Problem Statement

Computational Fluid Dynamics (CFD) is essential for predicting fire behavior, but its accuracy is often limited by the computational cost of modeling radiative heat transfer. In fire scenarios, radiation is the dominant heat transfer mechanism.

• The Bottleneck: Solving the high-dimensional Radiative Transfer Equation (RTE) using traditional numerical methods (e.g., Finite Volume Method, Discrete Ordinates Method) is computationally expensive and often becomes the performance bottleneck in CFD simulations.
• The Challenge: Existing machine learning surrogates (like standard Neural Operators) struggle with:
  1. High-dimensional inputs: CFD meshes contain hundreds of thousands of nodes.
  2. Multi-scale features: Fire plumes exhibit both smooth variations and sharp gradients (turbulence, vortices) that standard Multi-Layer Perceptrons (MLPs) fail to capture due to spectral bias.
  3. Mesh Heterogeneity: Practical fire simulations use locally refined meshes (multiple levels of refinement). A single monolithic neural network cannot efficiently process this non-aligned, hierarchical data structure.
  4. Generalization: Models trained on a specific fire size often fail when the fire grows to a new, unseen size (variable Heat Release Rate, HRR).

2. Methodology

The authors propose a Fourier-enhanced Multiple-Input Neural Operator (Fourier-MIONet) framework, extended with a nested architecture to handle multi-resolution meshes and variable fire sizes.

A. Core Architecture: Fourier-MIONet

• Multiple-Input Operator (MIONet): The RTE solution is treated as an operator mapping input fields (absorption coefficient $\kappa$ and temperature $T$) to the output radiative intensity $I(r, s)$. The architecture uses two "branch nets" to encode the input fields and one "trunk net" to encode spatial/angular coordinates.
• Fourier Enhancement: To overcome the spectral bias of MLPs (which struggle with high-frequency details), the authors integrate Fourier layers (inspired by Fourier Neural Operators) after the branch-trunk merger. These layers utilize Fast Fourier Transforms (FFT) to explicitly represent high-frequency components, allowing the model to capture sharp gradients and complex vortices in radiative fields.
• Dimensionality Reduction: Principal Component Analysis (PCA) is applied to the input fields to reduce the dimensionality of the branch inputs, making the high-dimensional CFD data tractable.
• KAN Integration: In some variants, Kolmogorov-Arnold Networks (KAN) replace standard MLP layers in the trunk to better approximate complex functions with learnable activation functions.

B. Nested Architecture for Multi-Resolution Meshes

To address the challenge of locally refined meshes (common in 3D CFD), the authors propose a Nested Fourier-MIONet:

• Hierarchy: The domain is decomposed into nested subdomains (e.g., 4 levels, from coarsest to finest).
• Coarse-to-Fine Inference: Instead of a single model, a hierarchy of Fourier-MIONets is trained.
  • The coarsest level ($N_4$) predicts the solution on the coarse mesh.
  • This prediction is upsampled (via a learnable deconvolution layer) and fed as an additional input to the next finer level ($N_3$).
  • This process repeats recursively up to the finest level ($N_1$).
• Global Assembly: The final global solution is assembled by selecting the finest available resolution for each spatial node from the respective level's prediction.

C. Data Generation

• Solver: Reference data was generated using FireFOAM (an OpenFOAM-based solver) with the Discrete Ordinates Method (fvDOM) and Gray Gas approximation.
• Scenarios:
  1. 2D Pool Fire: Used for architecture benchmarking.
  2. 3D McCaffrey Pool Fire (Fixed HRR): Five datasets with constant HRRs (14 kW to 58 kW).
  3. 3D McCaffrey Pool Fire (Variable HRR): A dataset where HRR ramps continuously from 10 kW to 60 kW to train a unified, generalizable model.
• Preprocessing: Simulation data on unstructured, refined meshes was interpolated onto structured, nested grids using linear barycentric interpolation.

3. Key Contributions

1. Fourier-MIONet for RTE: Demonstrated that Fourier-enhanced neural operators significantly outperform standard MIONet and KAN-only variants in predicting radiative intensity, achieving lower errors by capturing high-frequency features.
2. Nested Multi-Resolution Framework: Developed a novel Nested Fourier-MIONet that successfully handles the complexity of locally refined CFD meshes, enabling accurate predictions across multiple resolution levels without requiring a monolithic model.
3. Unified Variable-HRR Model: Created a single surrogate model capable of generalizing across a continuous range of fire sizes (10–60 kW), eliminating the need for retraining when fire intensity changes.
4. Scalable Efficiency: Provided four model sizes (Tiny, Small, Medium, Large) offering flexible trade-offs between inference speed and accuracy, suitable for different engineering constraints.

4. Results

The models were validated against FireFOAM simulations for 3D McCaffrey pool fires.

• Accuracy:

  • 2D Case: Fourier-MIONet achieved a global relative error of 3.70% for radiative intensity ($I$) and 1.93% for incident radiation ($G$), significantly outperforming baselines.
  • 3D Fixed-HRR: For the 58 kW fire, the "Large" nested model achieved a global relative error of ~1.38% for $I$ and 0.55% for $G$.
  • 3D Variable-HRR: The unified model trained on variable HRRs achieved global errors of 3.91% for $I$ and 2.47% for $G$ on the 58 kW test case, demonstrating robust generalization.
  • Physical Consistency: The models preserved energy conservation and accurately predicted radiative heat loss and radiation fractions (errors < 2%).

• Efficiency:

  • Inference Speed: The nested Fourier-MIONet inference on a single NVIDIA H200 GPU took between 0.016s (Tiny) and 0.044s (Large) for the full multi-level solution.
  • Comparison: This is significantly faster than the estimated cost of a single finite-volume radiation solve in FireFOAM (approx. 0.10s – 0.24s on 48 CPU cores), representing a potential 2x–5x speedup even without direct GPU acceleration of the traditional solver.

5. Significance

• Practical Engineering Application: This work bridges the gap between high-fidelity radiation modeling and computational efficiency. By replacing the expensive RTE solver with a fast, accurate neural surrogate, it enables the use of more spectrally resolved radiation models in large-scale CFD fire simulations.
• Scalability: The nested architecture solves a critical bottleneck in applying neural operators to real-world engineering problems involving adaptive mesh refinement.
• Generalization: The ability to train a single model for variable fire sizes makes the approach viable for dynamic fire scenarios where fire growth is unpredictable.
• Future Impact: The framework paves the way for fully coupled, high-fidelity radiation CFD simulations in industrial applications (e.g., warehouse storage fires, sprinkler performance evaluation), moving beyond proof-of-concept to practical fire safety engineering.