Hydrogen reaction rate modeling based on convolutional neural network for large eddy simulation
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.
Tom: I'm Tom.
Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.
Tom: We're looking at the paper: "Hydrogen reaction rate modeling based on convolutional neural network for large eddy simulation".
Paper summary: Tom: So we're looking at this paper now titled "Hydrogen reaction rate modeling based on convolutional neural network for large eddy simulation." It’s about creating a way to model how fast lean hydrogen air burns when you have turbulent flows, which is really tough because the hydrogen molecules diffuse way faster than the heat itself.
Jane: That sounds tricky. So what's the main idea here? What are they trying to achieve with this modeling framework?
Tom: Basically, they’re building a data-driven model using a convolutional neural network, or CNN, to approximate those burning rates for large eddy simulation, or LES of turbulent reactive flows. It’s addressing those huge variations in burning rates and instabilities that happen at the subfilter scale because of how fast the hydrogen diffuses compared to the heat.
Lu: I think they are trying to move away from old methods, you know, flame surface density approaches based only on turbulence wrinkling, because those don't capture these thermodiffusive instabilities caused by the difference between heat and hydrogen mass fluxes.
Meng: That sounds like it needs a high-fidelity simulation to get right. What kind of data are they using to train this CNN? How do you even get the training material?
Tom: They generate the training data by taking five different direct numerical simulations, or DNS, of lean premixed turbulent hydrogen air flames, each with a different global equivalence ratio. Then they filter and downsample those DNS snapshots to make them look like LES data that they know how to handle.
Jane: So they’re using high-fidelity simulation results—the DNS—to create the training set for their machine learning model, and then they use that set to learn the mapping between those simulated flow quantities and the actual burning rates.
Lu: The CNN is designed specifically to learn spatial hierarchies of features, moving from low-level patterns up to higher-level ones, which should help it recognize that complex flame topology and figure out the associated burning rates. They are using a U net architecture inspired by image segmentation for this purpose.
Meng: Okay, if the model is learning from those simulated quantities like progress variable and local equivalence ratio, what’s the actual output we care about? What does it predict?
Tom: It outputs an approximation of the burning rates over that domain, which they call omega over Omega. And one of the most interesting things they found is that this CNN model can approximate burning rates on solutions it has never seen during training with very high accuracy.
Jane: That sounds pretty impressive for a machine learning model. So, how does this new approach compare to what we usually use, like the filtered tabulated chemistry approach?
Paper summary: Tom: The CNN modeling significantly outperforms that filtered tabulated chemistry one, especially in the lean cases—for example, when the fuel-air equivalence ratio is zero point three five and sigma is four and DSF is two. For that specific case, they found the maximum burning rate in their three dimensional field was about seventeen times greater than what they saw in a laminar planar flame <ref:2408.16709#pg1>.
Lu: That suggests the CNN’s ability to consider the full three dimensional flame structure through those successive convolutions is really what gives it that superior performance over simpler models <ref:2408.16709#pg3>.
Meng: Seventeen times greater, huh? That’s a big jump. But how much computational power does training this model actually require? Is it feasible for production use, or is this stuck in simulation labs?
Tom: Well, the paper addresses practical integration challenges related to high performance computing code deployment. It shows that while the model is accurate on three dee full-scale test solutions, deploying it for real-world applications involves some engineering work to make sure it runs efficiently <ref:2408.16709#pg3>.
Jane: So, we’re seeing a tool that can predict complex combustion rates much faster than traditional methods because it learns the underlying physics from high-fidelity simulation data. What does this mean for designing those hydrogen combustion devices they mentioned?
Lu: It means we can get a better fundamental understanding of flame behavior when things get turbulent and lean, which is essential for safety and design in hydrogen systems <ref:2408.16709#pg2>.
Meng: From my side, if I had to deploy this, the main question is generalization. The paper claims it works on solutions never seen during training; that’s good, but how robust is that generalization when we move to a completely different operating condition?
Tom: The authors are confident in its ability to generalize outside the scope of what they trained it on, which is something they highlight as important for this data-driven approach <ref:2408.16709#pg1>.
Jane: So, the main implication seems to be that we can model these extremely challenging lean premixed hydrogen air flames accurately using machine learning instead of relying only on classical chemistry approximations. That’s what they call the hydrogen reaction rate modeling based on convolutional neural network for large eddy simulation.
Lu: It shifts the paradigm toward using data-driven models to capture complex, multi-scale physics in reactive flows, which is a direction that needs more exploration <ref:2408.16709#pg3>.
Meng: For me, I’m still worried about the fidelity of that training data. If those five DNS runs weren't perfectly representative of all possible turbulence states, then the model might be brittle when it hits a real-world scenario we haven't simulated.
Tom: That’s the inherent trade-off with data-driven modeling, right? You get high accuracy on what you trained it on, but the performance outside that training set is always something to test rigorously.
Jane: Exactly. So, as we look at this paper now, it shows a path forward for simulating these complex carbon-free combustion systems by using neural networks to handle the sheer complexity of those turbulent flows.
Conclusion: Tom: So, to wrap up, this paper is about using a convolutional neural network to model how fast lean hydrogen air burns in turbulent flows for large eddy simulation.
Jane: Right, so the main idea is building this data-driven framework for those tricky combustion rates. It’s trying to solve that problem where the hydrogen molecules diffuse way faster than the heat itself, which messes up traditional models.
Tom: Exactly. The authors use direct numerical simulations from five different setups to train their model on what they know about those turbulent flows, and it learns how to predict the burning rates from flow quantities like progress variable and local equivalence ratio.
Jane: And what’s really interesting is that this network can actually predict burning rates on solutions it has never seen during training with very high accuracy. It's not just fitting the data they showed them.
Tom: Right, because they compared it to a filtered tabulated chemistry approach, and in lean cases, like when the fuel-air equivalence ratio was zero point three five and sigma was four and DSF was two, their model got about seventeen times greater in burning rate than what that simpler approach showed.
Jane: Seventeen times greater. That’s a big number that really highlights how much detail the convolutional network is picking up from the three dimensional flame structure compared to just looking at the surface area.
Tom: Yeah, it shows you can build this kind of machine learning model for lean premixed hydrogen air flames—which is super hard—and it actually generalizes outside the exact conditions they trained it on.
Jane: So what this means for us is that we might be able to get a much better way to simulate these carbon-free hydrogen combustion systems, moving beyond just the classical approximations.
Tom: It opens up a new way of thinking about modeling complex physics in reactive flows using data, and the next step they are taking involves training it on something even more fundamental—the source term of a progress variable instead of individual species.
Quentin Malé, Corentin J. Lapeyre, Nicolas Noiray
CAPS Laboratory, Department of Mechanical and Process Engineering, ETH Zurich · NVIDIA Corporation
cs.CE
Submitted: 2024-08-29
Updated: 2026-11-18
Comments: 18 pages, 9 figures. The journal version is available online at: https://doi.org/10.1017/dce.2025.1
Journal ref: DCE 6 (2025) e11
DOI: 10.1017/dce.2025.1
Project page: https://blastnet.github.io/index.html
License: http://creativecommons.org/licenses/by/4.0/
The gist: This paper establishes a data-driven modeling framework for lean Hydrogen (H2)-air reaction rates for the Large Eddy Simulation (LES) of turbulent reactive flows.
Key concepts
- Lean Premixed H2-air Flames
- These are hydrogen-air mixtures where the hydrogen concentration is low. Modeling them is difficult because the fast diffusion of hydrogen relative to heat causes large variations in burning rates and instability at very small scales, which traditional models often fail to capture.
- Large Eddy Simulation (LES)
- LES is a computational method used to simulate turbulent flows by resolving the largest turbulent eddies directly while modeling smaller, unresolved ones. It is used here to study the complex, three-dimensional behavior of hydrogen combustion in a realistic turbulent environment.
- Convolutional Neural Network (CNN)
- A type of machine learning model inspired by image processing that excels at recognizing spatial features. The CNN was trained on simulation data to learn how different patterns in the flow field relate to the resulting burning rates, allowing it to predict rates with high accuracy even for unseen simulations.
Terminology
Summary
This paper establishes a data-driven modeling framework for lean Hydrogen (H2)-air reaction rates for the Large Eddy Simulation (LES) of turbulent reactive flows. p. 0 2NVIDIA Corporation, Santa Clara, CA, United States. Keywords: data-driven reacting flow modeling; hydrogen combustion; machine learning; artificial intelligence; computational
fluid dynamics
Abstract
This paper establishes a data-driven modeling framework for lean Hydrogen (H2)-air reaction rates for the Large
Eddy Simulation (LES) of turbulent reactive flows. This a priori study shows that the proposed data-driven machine learning
framework is able to address the challenge of modeling lean premixed H2-air burning rates. It paves the way for a
new modeling paradigm for the simulation of carbon-free hydrogen combustion systems.Where the paper says it
Where the paper says it
Modeling Challenges in Turbulent Combustion
The challenge in modeling lean premixed H2-air burning rates stems from the high molecular diffusion of H2, which is much faster than heat, leading to large variations in burning rates and thermodiffusive instabilities at the subfilter scale. These phenomena invalidate usual Flame Surface Density (FSD) approaches based on turbulence wrinkling alone. p. 2 [54]
reports a four-fold enhancement in flame speed compared to an unstrained flame. These phenomena invalidate usual Flame Surface Density (FSD) approaches [43, 34, 11, 51] based
on the assumption that subfilter-scale effects are limited to an increase in flame surface density through
wrinkling due to turbulence alone. First, lean H2 flames are prone to thermodiffusive instabilities. The disparity between heat and H2 mass
fluxes (called differential diffusion) amplifies small flame front perturbations, inducing significant flame
wrinkling and strong variations of the reaction rates along the flame front [42, 5, 6]. The disparity between heat and H2 mass
fluxes (called differential diffusion) amplifies small flame front perturbations, inducing significant flame
wrinkling and strong variations of the reaction rates along the flame front [42, 5, 6]. Second, burning
rates increase significantly when lean H2 flame is subjected to a strain [39, 54], up to a certain limit.Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
The data generation strategy involves filtering and downsampling data from Direct Numerical Simulations (DNS) to emulate LES data with known filtered quantities. The process includes the following steps:
-
Five different lean premixed turbulent H2-air flame Direct Numerical Simulations (DNSs) are computed each with a unique global equivalence ratio. p. 0
Where the paper says it
Our data-driven approach leverages a Convolutional Neural Network (CNN), trained to approximate filtered burning rates from emulated LES data. First, five different lean premixed turbulent H2-air flame Direct Numerical Simulations (DNSs) are computed each with a unique global equivalence ratio.
-
DNS snapshots are filtered and downsampled to emulate LES data. p. 0
Where the paper says it
First, five different lean premixed turbulent H2-air flame Direct Numerical Simulations (DNSs) are computed each with a unique global equivalence ratio. Second, DNS snapshots are filtered and downsampled to emulate LES data.
-
A Convolutional Neural Network (CNN) is trained to approximate the filtered burning rates as a function of LES scalar quantities: progress variable, local equivalence ratio and flame thickening due to filtering. p. 0
Where the paper says it
Second, DNS snapshots are filtered and downsampled to emulate LES data. Third, a CNN is trained to approximate the filtered burning rates as a function of LES scalar quantities: progress variable, local equivalence ratio and flame thickening due to filtering.
Where the gains come from
The CNN is designed to learn spatial hierarchies of features, from low- to high-level patterns, making it well fitted for recognizing complex flame topology and learning associated burning rates. The CNN architecture is inspired by the U-net architecture originally developed for image segmentation. p. 4 The
application of CNNs to the modeling of flame features has already been successfully performed in
Refs [32, 31, 49, 36] for fuel-air mixture without significant preferential diffusion effects, unlike the
H2-air mixtures considered in this paper. The CNN architecture used is inspired by the U-net architecture originally developed for image
segmentation [44] (Fig. The filtered equivalence ratio can be
approximated as
𝜙e=
𝜉e(1 − 𝜉𝑠)
𝜉𝑠
1 − 𝜉e
. 5)
The CNN takes as input the field of e𝑐 and 𝜙eover a domain Ω, performs successive convolutions and
operations, and outputs an approximation of 𝜔¤ over the same domain Ω. From the emulated LES fields, a Convolutional Neural
Network (CNN) is trained to estimate the LES burning rates based on known LES scalar quantities such
as the filtered mixture fraction and progress variable. Once trained, the model is found to approximate
the burning rates on solutions never seen during training with high accuracy.Where the paper says it
Where the paper says it
Where the paper says it
Comparison with a filtered tabulated chemistry approach
The CNN-based modeling significantly outperforms the filtered tabulated chemistry one, especially for the lean cases (e.g., phig = 0.35 in Fig. 9). For the case phig = 0. p. 0 Sustainably produced hydrogen (H2) is a promising energy carrier
for the decarbonization of combustion systems where there is no viable technology to be a substitute. This is the case, for example, with gas turbines for power generation or aviation, as well as some
industrial burners. It is chosen
so as to have a LES grid with a resolution index RI = 𝛿
1
𝑐,𝐿/Δ𝑥 + 1 of about 6. 𝛿
0
𝑐,𝐿 and 𝛿
1
𝑐,𝐿 are the
laminar flame thickness based on the progress variable of a non-filtered and filtered 1D flame
𝛿
0
𝑐,𝐿 = (max |∇𝑐|)−1
and 𝛿
1
𝑐,𝐿 = (max |∇e𝑐|)−1
.Where the paper says it
Where the paper says it
Conclusion and future work
The trained CNN model retrieves burning rates with very high accuracy on the 3D full-scale test solutions. The ML-based modeling outperforms a priori a classical filtered tabulated chemistry approach with correction for flame surface density due to subfilter-scale flame wrinkling. p. 2 [54]
reports a four-fold enhancement in flame speed compared to an unstrained flame. These phenomena invalidate usual Flame Surface Density (FSD) approaches [43, 34, 11, 51] based
on the assumption that subfilter-scale effects are limited to an increase in flame surface density through
wrinkling due to turbulence alone. The phenomena involved in H2 turbulent combustion being extremely complex, not well understood,
and still subject to discussion, we propose to change the paradigm by developing a data-driven method
based on ML instead of trying to model turbulent H2 burning rates by physical description. This work
is motivated by the positive results of such an approach for the approximation of the subfilter flame
surface density both a priori and a posteriori for methane-air combustion [32, 31, 26]. Final implementation will depend on the given LES modeling framework. For example,
the ML model could replace the tabulation method proposed in Refs [29, 30] for the chemical source
term of the progress variable.Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Data-driven modeling of lean premixed hydrogen air reaction rates can be achieved by training a Convolutional Neural Network (CNN) to approximate filtered burning rates from emulated Large Eddy Simulation (LES) data. This allows the system to retrieve burning rates with very high accuracy on test solutions never seen during training, including cases where interpolation is performed between training conditions.
-
The improved AI system can perform high-fidelity simulation of hydrogen combustion systems by addressing the challenge posed by the high molecular diffusion of H2 through a data-driven machine learning framework. This enables the simulation to account for complex turbulence-chemistry interactions and thermodiffusive instabilities at the subfilter scale, which are challenging for traditional physical modeling approaches.
-
The CNN can be trained to recognize complex flame topology and learn associated burning rates by utilizing spatial convolutions on fields of filtered progress variable and filtered equivalence ratio. This capability allows the model to accurately reproduce extremely complex turbulent flame topologies, including burning pockets detached from the main flame brush, across various LES filter sizes and global equivalence ratios.
-
The system can generalize its predictive capabilities to new LES parameters and global equivalence ratios by training on a diverse database of five different DNS cases across varying global equivalence ratios. This demonstrates the ability to approximate burning rates with low error even for interpolation cases not included in the training dataset, paving the way for application in engineering simulation software.
-
The ML-based modeling outperforms a priori classical filtered tabulated chemistry approaches by accounting for subfilter-scale flame wrinkling via a wrinkling factor X. This is particularly effective for very lean flames where differential diffusion effects are major, as it captures the significantly higher burning rates of the turbulent 3D field compared to freely propagating planar flames.
Abstract
This paper establishes a data-driven modeling framework for lean Hydrogen (H2)-air reaction rates for the Large Eddy Simulation (LES) of turbulent reactive flows. This is particularly challenging since H2 molecules diffuse much faster than heat, leading to large variations in burning rates, thermodiffusive instabilities at the subfilter scale, and complex turbulence-chemistry interactions. Our data-driven approach leverages a Convolutional Neural Network (CNN), trained to approximate filtered burning rates from emulated LES data. First, five different lean premixed turbulent H2-air flame Direct Numerical Simulations (DNSs) are computed each with a unique global equivalence ratio. Second, DNS snapshots are filtered and downsampled to emulate LES data. Third, a CNN is trained to approximate the filtered burning rates as a function of LES scalar quantities: progress variable, local equivalence ratio and flame thickening due to filtering. Finally, the performances of the CNN model are assessed on test solutions never seen during training. The model retrieves burning rates with very high accuracy. It is also tested on two filter and downsampling parameters and two global equivalence ratios between those used during training. For these interpolation cases, the model approximates burning rates with low error even though the cases were not included in the training dataset. This a priori study shows that the proposed data-driven machine learning framework is able to address the challenge of modeling lean premixed H2-air burning rates. It paves the way for a new modeling paradigm for the simulation of carbon-free hydrogen combustion systems.
Sources
Related papers
- Constrained Sensing and Reliable State Estimation with Shallow Recurrent Decoders on a TRIGA Mark II Reactor
- DREChanNet: Patient-Relative iEEG Model for Channel-Wise Epileptogenic Zone Localization
- A matrix-free, differentiable PyTorch solver for phase-field fracture: Formulation, benchmarks, and inverse analysis
- BizAgentBench: Unveiling Hidden Challenges in Deploying Long-Horizon Agents for Real-World Business
- Colour me shocked: Exact Molecular Hessians from local MLIPs in O(N) time using sparse differentiation!
- In-context learning as a retraining-free paradigm for constitutive modeling