Modeling subgrid scale production rates on complex meshes using graph neural networks

This paper presents a graph neural network (GNN) that accurately predicts filtered species production rates for large-eddy simulations of turbulent premixed flames on complex, non-uniform meshes, demonstrating superior generalization across varying fuel compositions and filter widths compared to traditional unclosed models and convolutional neural networks.

The Gist

Imagine you are trying to predict the weather for a massive city. You have a supercomputer that can track every single raindrop and gust of wind, but that would take too long and require too much power. So, instead, you decide to look at the weather in big, 10-mile-wide blocks. You know the average temperature and wind speed for each block, but you don't know exactly what's happening inside the tiny streets and alleys.

This is the problem scientists face when simulating fire and explosions (like in jet engines or power plants). They use a method called Large-Eddy Simulation (LES). It's like looking at the "big blocks" of the fire. It captures the big swirls of flame, but it misses the tiny, super-fast chemical reactions happening in the gaps between those swirls.

Because the computer can't see the tiny details, it has to guess what's happening inside those gaps. This guess is called a "closure." If the guess is wrong, the whole simulation of the engine could fail or predict a fire that never happens.

The Old Way: Trying to Force a Square Peg into a Round Hole

Traditionally, scientists tried to fix this guess using two main methods:

1. The "No-Model" Guess: Just assume the average conditions are the whole truth. It's like saying, "The average temperature in the city is 70°F, so it must be 70°F everywhere." This is usually wrong because the fire burns hottest in tiny, specific spots that get averaged out.
2. The "CNN" (Convolutional Neural Network) Guess: This is a type of AI that is very good at looking at pictures. But, it only works on perfect, grid-like pictures (like graph paper). Real fire simulations happen on messy, irregular grids (like a city with winding streets and dead ends). To use this AI, scientists had to remap their messy data onto a perfect grid, run the AI, and then map it back.
   • The Analogy: Imagine trying to fit a jigsaw puzzle with irregular pieces onto a square table. You have to cut the pieces to fit the table, solve the puzzle, and then try to glue them back to their original shapes. In the process, you distort the picture and lose important details.

The New Solution: The "Graph Neural Network" (GNN)

This paper introduces a new AI called a Graph Neural Network (GNN). Instead of forcing the data into a perfect grid, the GNN understands the data exactly as it is: a messy, irregular network of connections.

Here is how the authors made it work, using some fun analogies:

1. The Neighborhood Chat (Message Passing)

Imagine the computer simulation is a giant neighborhood. Each house (a point in the mesh) has neighbors.

• Old AI: Only talks to the houses directly to its North, South, East, and West on a perfect grid.
• The GNN: Understands that in a real city, neighbors might be slightly closer or further away, or the streets might curve. It sends "messages" along the actual connections (roads) between houses.
  • The Magic: The AI learns that if a house is hot and its neighbor is cold, the heat is moving between them. It does this by "chatting" with its immediate neighbors, then their neighbors, and so on. This allows it to understand the shape of the fire without needing a perfect grid.

2. The "No-Remeshing" Superpower

The biggest breakthrough is that the GNN doesn't need to be forced onto a grid.

• The Analogy: Think of the GNN as a smart tour guide who knows the city by its actual streets. The old CNN was like a tourist who only knows how to walk in straight lines on a grid. The tourist has to constantly stop and ask, "Which way is North?" and get confused by the winding streets. The tour guide (GNN) just walks naturally, knowing exactly how the streets connect, preserving the true shape of the fire.

3. The "Blind Taste Test" (Generalization)

To prove their new AI was smart, the scientists did a tricky test:

• They trained the AI on fires with 10% hydrogen and 80% hydrogen.
• They then asked the AI to predict a fire with 50% hydrogen—a mix it had never seen before.
• The Result: The GNN guessed correctly! It figured out that the 50% mix was somewhere in the middle of the two it knew. The old methods (the "No-Model" and the "CNN") failed miserably, either guessing wildly wrong or blurring the details.

4. The "Zoom Out" Test

They also tested what happens if you make the "blocks" (the mesh) much bigger (coarser resolution).

• The Analogy: Imagine looking at a fire through a telescope. If you zoom out too far, the fire looks like a blurry blob.
• The GNN was able to look at these blurry blobs and still guess the chemical reactions accurately, without needing to be retrained. It's like a master chef who can guess the recipe of a soup just by tasting a spoonful, even if the spoonful is tiny or the soup is very hot.

Why Does This Matter?

This new method is a game-changer for designing cleaner, safer, and more efficient engines.

• Accuracy: It predicts chemical reactions much better than before.
• Speed: It doesn't waste time remapping data to perfect grids.
• Flexibility: It works on the messy, complex shapes of real-world engines (like the "backward-facing step" geometry they tested, which mimics the tricky corners inside a jet engine).

In short: The scientists built a new AI that learns to predict fire chemistry by understanding the actual, messy shape of the fire, rather than forcing the fire into a box. It's smarter, faster, and more accurate, paving the way for better engines and cleaner energy in the future.

Technical Summary

1. Problem Statement

Large-Eddy Simulations (LES) are essential for simulating turbulent combustion but face a fundamental closure problem: the filtered chemical source terms (production rates) cannot be accurately calculated by simply evaluating reaction rates at the filtered thermochemical state ($\dot{\omega}_k \neq \dot{\omega}_k(\tilde{\phi})$). This is because subgrid-scale (SGS) turbulence-chemistry interactions significantly alter reaction pathways.

Current data-driven solutions, such as Convolutional Neural Networks (CNNs), typically require mapping non-uniform or unstructured computational meshes (common in practical combustors) onto uniform Cartesian grids via interpolation. This process:

• Distorts the thermochemical structure.
• Introduces interpolation errors.
• Loses the native spatial connectivity of the solver mesh.
• Limits the applicability of LES to complex geometries.

The paper addresses the need for a mesh-native, interpolation-free closure model that can predict filtered species production rates directly on complex, non-uniform meshes while capturing spatial correlations.

2. Methodology

The authors propose a Graph Neural Network (GNN) framework designed to operate directly on the discrete connectivity of the computational mesh.

• Graph Formulation:
  • The computational domain is partitioned into subdomains.
  • Nodes: Represent mesh cell centers.
  • Edges: Defined by one-hop connectivity (six-point stencil plus self-connection) along mesh coordinate directions.
  • Features:
    • Node Features: Filtered mass fractions of a compact set of species (reactants, intermediates, products) and filtered temperature.
    • Edge Features: Differences in filtered scalars, relative coordinate offsets, and Euclidean distances between connected nodes. This encodes local spatial gradients and non-uniform mesh spacing.
• Architecture:
  • An Encoder-Processor-Decoder architecture with five message-passing (MP) layers.
  • Message Passing: Information is exchanged between neighboring nodes. Each layer aggregates messages from the neighborhood (using mean aggregation to handle varying neighbor counts) and updates node features via Multi-Layer Perceptrons (MLPs).
  • Receptive Field: Five layers allow nodes to access information from a five-hop neighborhood, capturing spatial correlations comparable to the integral length scale at LES resolution.
  • Normalization: Min-max normalization is applied to inputs/outputs to handle the wide dynamic range of species production rates. Sigmoid activation ensures predictions remain within the normalized [0, 1] range.
• Training Data:
  • Source: Direct Numerical Simulation (DNS) of turbulent premixed hydrogen-methane ($H_2/CH_4$) jet flames.
  • Compositions: 10% $H_2$ (H10), 50% $H_2$ (H50), and 80% $H_2$ (H80).
  • Filtering: DNS fields are Favre-filtered with filter widths matching the target LES mesh spacing.
  • Strategy: The model is trained on H10 and H80 blends and tested on the unseen H50 blend to evaluate cross-composition generalization.
• Baselines for Comparison:
  1. No-Model (Unclosed): Evaluates chemistry directly at the filtered state.
  2. CNN Baseline: A convolutional network with matched complexity (5 layers, 48 channels) that requires interpolating data to a uniform grid and back.

3. Key Contributions

• Mesh-Native Closure: Introduces a GNN that operates directly on non-uniform meshes, eliminating the need for interpolation or remeshing, thereby preserving the exact solver topology.
• Multi-Species Prediction: Unlike many prior works focusing on a single progress variable, this model predicts the production rates of an extensive set of species simultaneously, capturing complex thermochemical coupling.
• Robust Generalization: Demonstrates the ability to generalize to unseen fuel compositions (H50) and varying filter widths (coarser resolutions) without retraining.
• Geometric Versatility: Validates the approach on a complex geometry (backward-facing step combustor) with different flow physics (recirculation zones) and fuel (ethylene-air).

4. Results

The study evaluates the GNN against the no-model and CNN baselines across several metrics:

• Prediction Accuracy (H50 Unseen Blend):
  • No-Model: Exhibits massive errors (>100%) in thin reaction zones due to ignoring SGS correlations.
  • CNN: Reduces peak errors but introduces artificial smearing due to interpolation, distributing errors over a broader spatial extent.
  • GNN: Maintains prediction errors below 10% for most species. The largest errors (~20%) are observed for the highly reactive $OH$ radical, which is sensitive to minute perturbations.
• Statistical Agreement:
  • Joint Probability Density Functions (PDFs): The GNN tightly tracks the reference distribution, whereas the no-model diverges significantly and the CNN remains statistically diffuse.
  • $R^2$ Scores: The GNN achieves $R^2 > 0.89$ for most species (e.g., $H_2O$, $CH_2O$), compared to $R^2 = -5.53$ for the no-model and $0.29$ for the CNN.
• Resolution Robustness:
  • When tested on coarser meshes (downsampling factors of 12$\times$ and 16$\times$) without retraining, the GNN maintains bounded errors (<15% for most species), demonstrating resilience to significant loss of resolved gradients.
• Complex Geometry (Backward-Facing Step):
  • Applied to a lean premixed ethylene-air flame in a backward-facing step configuration. The GNN successfully restricted errors to the reacting shear layer (<10%) and predicted near-zero production in unburned regions, preventing spurious ignition.

5. Significance

This work establishes Graph Neural Networks as a superior, scalable approach for data-driven subgrid chemistry modeling in LES.

• Elimination of Interpolation Artifacts: By operating natively on the mesh, the model avoids the distortions inherent in CNN-based approaches, making it suitable for industrial-grade simulations on complex, unstructured meshes.
• Generalization: The model's ability to handle unseen fuel blends and varying filter widths suggests it can be deployed in practical combustors where operating conditions fluctuate.
• Pathway to Practical LES: The framework provides a computationally efficient, interpolation-free route to integrating detailed finite-rate chemistry into LES, bridging the gap between high-fidelity DNS data and practical engineering simulations.