acceptodds
Under review as a conference paper at ICLR 2027

FireMark: A Benchmark Dataset for Data-Driven Forecasting of Laminar Flame Dynamics

Abstract

High-fidelity computational fluid-dynamics (CFD) simulations of reacting flows are computationally expensive, limiting their usage in real-time applications. In this context, machine learning (ML) methods offer a promising approach for predicting dynamic flow evolution. However, evaluating them remains an open challenge due to the limited availability of benchmarks that combine high temporal resolution with a diverse range of exogenous inputs. In this work, we introduce FireMark, an open benchmark for evaluating data-driven methods to forecast reacting flows evolution. The benchmark contains multiple simulations of a non-premixed laminar methane/nitrogen coflow flame subjected to time-varying inlet velocities. FireMark includes both thermodynamic and chemical fields, obtained through CFD simulations with a detailed chemical kinetic mechanism. The benchmark is designed to assess the ability of forecasting models to reproduce the evolution of a laminar non-premixed flame and to generalize to unseen input signals. Training data consist of sine-sweep forcing signals at two amplitudes, whereas the test set comprises fixed-frequency sinusoids and step functions, with amplitudes both within and outside the training range. To establish reference performance and demonstrate the benchmark capabilities, we evaluate representative classical reduced-order and ML-based forecasting approaches using a set of quantitative metrics, including measures of physical consistency. A linear or bilinear autoregressive model on a reduced basis is more accurate than every neural forecaster on the same latent space and reproduces the heat-release response best; several neural forecasters fall below the accuracy of a constant initial state predictor; and the Fourier neural Operator (FNO), which has the lowest field error, reproduces the integrated heat release worse than that constant forecast. These results provide a baseline performance against which future forecasting methods can be compared. FireMark is publicly available and provides a standardized framework for the reproducible evaluation of data-driven models for reacting-flow dynamics.

Primary area:
datasets and benchmarks
Keywords:
Scientific machine learning, Flow forecasting, Reduced-order modeling, Reactive flows, Computational fluid dynamics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_ziiSOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces a new dataset of laminar flame dynamics simulations, with the goal of testing the performance of different forecasting methods (including traditional autoregressive models, recurrent neural networks and neural operators) on this complex dataset. The data is obtained by computational fluid dynamics and takes into account chemical reactions. 8 different time rollouts are included with different boundary conditions of the fuel speed. Two such rollouts are used for training and the rest for testing. Testing decouples the latent representation from the forecasting method (except for end-to-end models). The results seem to show that neural operators perform well in terms of forecast accuracy on average, but poorly in terms of predicting the correct heat release. Meanwhile, linear methods outperform several neural networks and achieve the best accuracy in predicting the heat release.

Critical strengths
  • A realistic benchmark for flame dynamics was devised, providing open clean data for others to test their methods on. Several methods were tested with a clear protocol.
  • The evaluation protocol seems to distinguish different qualities of the predictive methods (i.e. just field error or heat release) which may be crucial for final physical accuracy. While the analysis is not fully fleshed out, this could be quite important.
Critical weaknesses
  • The data is limited in scope to a single geometry, to a single flame chemistry, and to laminar flows. Therefore, while filling a gap in existing benchmarks, I believe this data falls short of providing a comprehensive benchmark for flame dynamics.
  • There is very little analysis of the results. Beyond stating that linear models are simpler and thus may overfit the data less, the critical limitation of e.g. recurrent or convolutional models is unclear. Furthermore the very high heat release error of the neural operator remains unexplained given the good performance on the field prediction.
Questions for authors
  • HP selection for neural operators seems to choose a very small NxN_x (2 compared to at least 8 for other methods). Could this be a problem behind the poor results on the heat release error?
  • Could you include a further analysis of the errors for representative methods for each class?
  • Could you expand on the motivation for developing such a benchmark and what are the applications of surrogate models specifically on this problem? In particular I am curious about the standard efficiency motivation: are all tested models faster at inference than the CFD, and by how much?
1 comment

Reviewer ziiS

1. HP selection for neural operators seems to choose a very small NxN_x (2 compared to at least 8 for other methods). Could this be a problem behind the poor results on the heat release error?

We thank the reviewer for this interesting observation. The FNO value Nx=2N_x=2 was selected by the Optuna hyperparameter search, rather than fixed a priori. The search considered the same range of candidate values used for the other neural models and selected the configuration with the lowest validation MSE. Although we cannot exclude an effect of NxN_x on heat-release prediction, the POD+DMD results Nx=0N_x=0, which also yield one of the lowest ϵQ\epsilon_Q values, suggest that a larger NxN_x does not necessarily lead to better heat-release predictions.

A closer inspection of the FNO predictions (Appendix E, "Additional Result") reveals that larger errors are concentrated mainly in the upper-right region of the domain, where the cells have larger volumes, as visible in the new Figure 1. This helps explain the discrepancy between the field-wise nRMSE and the integrated heat-release error: nRMSE assigns equal weight to every cell, whereas the calculation of total heat release weights each cell contribution by its volume. Consequently, errors in larger cells have a greater impact on the predicted total heat release. We have added this analysis to the Appendix E.

2. Could you include a further analysis of the errors for representative methods for each class?

We thank the reviewer for this suggestion. We added a more detailed error analysis for one representative method per model class in Appendix E ("Additional Results") of the manuscript. The chosen representative methods are POD–ARX (closed-form forecaster on a linear compressor), CAE–ARX (closed-form forecaster on a nonlinear compressor), POD–Transformer (the neural latent forecaster with the lowest nRMSE) and FNO (neural operator on the full fields), with the constant forecast as reference. The analysis includes the error growth along the closed-loop rollout, to assess whether the errors accumulate over the forecast horizon, and the spatial distribution of the heat-release error, to identify where the heat-release error originates and why FNO combines the lowest field error with the largest heat-release error.

3. Could you expand on the motivation for developing such a benchmark and what are the applications of surrogate models specifically on this problem? In particular I am curious about the standard efficiency motivation: are all tested models faster at inference than the CFD, and by how much?

We thank the reviewer for the observation. FireMark addresses the lack of open benchmarks for forecasting spatially resolved reacting-flow fields under controlled, time-varying forcing. Such benchmarks are needed to compare forecasting methods under consistent conditions. In the longer term, surrogate models could support digital twins of combustion systems, enabling real-time prediction and control under changing operating conditions. Potential applications include instabilities and explosions detection, real-time fuel consumption optimization and emission reduction. These applications motivate the benchmark but are not demonstrated by the present single-configuration dataset. We have sytressed this motivation in the introduction, section 1 of the manuscript.

Regarding computational efficiency, the eight CFD simulations required approximately 127–508 wall-clock hours each on 20 cores. This information about the computational cost are reported in section 3 and further details were reported in section B of the appendix.

It is not possible to calculate a speed up factor from CFD simulations to surrogate models as we used two different machines to do the calculations. However, a rough comparison already shows the order of magnitude. The fastest CFD simulation (step, A=0.5A=0.5) required about 4.6×1054.6\times10^{5} s (≈127\approx 127 h on 20 CPU cores) to produce the 2000 snapshots of one trajectory, i.e. about 230 s per snapshot. In contrast, every tested surrogate requires less than 6 ms per step (see the inference-cost table in Appendix F), so a full 2000-step rollout takes 12 s maximum, depending on the model. This corresponds to a speed-up of roughly 4×1044\times10^{4} for the slowest surrogate (exact DMDc) and 4×1054\times10^{5} for the fastest.

We have added this in the paragraph Surrogate model speed-up, Appendix F.

Reviewer_DPx8OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper introduces FireMark, a dataset and benchmark for forecasting forced laminar flame dynamics. It contains eight simulations of an axisymmetric non-premixed methane/nitrogen flame, recording eleven flow and thermochemical fields at 0.5 ms intervals. Models are trained on two sine-sweep trajectories and evaluated through approximately 2,000-step closed-loop rollouts on six sinusoidal and step-forcing trajectories, covering amplitude interpolation and extrapolation. The benchmark compares classical reduced-order models, neural sequence models, and FNO using field errors, integrated heat-release errors, and response gain and phase. The results demonstrate the competitiveness of simple autoregressive models and show that low average field error does not necessarily imply accurate heat-release dynamics.

Critical strengths
  1. Training and testing are separated by complete forcing trajectories rather than randomly sampled neighboring snapshots. Test rollouts receive the steady-state field only at initialization and subsequently evolve recursively under the known forcing, without repeated resets to reference states. Testing both within and beyond the training amplitude range, together with the transition from sine sweeps to fixed-frequency and step inputs, establishes a practically meaningful forecasting task for forced dynamical systems.

  2. The evaluation includes a constant baseline, classical reduced-order methods, and neural models, with per-forcing results and unstable configurations disclosed in the appendix. Notably, FNO achieves the lowest overall nRMSE of 0.107, yet its integrated heat-release error is 0.611, exceeding the constant baseline’s 0.206; POD–ARX achieves 0.026 on this metric (Table 2). Together with the gain and phase results in Table 6, this demonstrates that selecting models solely by average field error can overlook substantial errors in combustion dynamics. This is the paper’s most informative finding.

Critical weaknesses
  1. The validation split is difficult to reconcile with the stated validation horizon. Section 4 divides each two-second trajectory, sampled at 0.5 ms intervals, into 20 blocks and reserves 30% of each block for validation. This yields approximately 200 samples per block and only 60 validation samples. However, both the main text and Appendix B specify 150-step recursive validation rollouts, while windows are stated not to cross block boundaries. It is therefore unclear how these rollouts remain entirely within held-out portions. This does not establish data leakage, but it prevents readers from verifying the separation between training and validation targets. Since model selection and early stopping depend on this protocol, the ambiguity materially affects confidence in the benchmark.

  2. The numerical accuracy of the reference trajectories is insufficiently established. Section 3 identifies the solver, mesh size, chemical mechanism, and previously studied physical configuration, providing a reasonable foundation for data generation. However, the manuscript does not report mesh or internal time-step convergence for the present forced trajectories, nor clarify which forcing conditions are covered by existing experimental validation. The 0.5 ms recording interval does not establish solver accuracy. Because heat-release gain and phase are central evaluation quantities, representative numerical uncertainty estimates or reference comparisons are needed to assess forecasting errors relative to uncertainty in the CFD reference.

Questions for authors
  1. Please provide the training, context, and validation-scoring indices for a representative block, explaining how approximately 60 held-out samples support a 150-step validation rollout. Does the rollout cross training–validation boundaries? Which samples are used to fit normalization statistics, POD bases, and autoencoders?

  2. What mesh-convergence, internal time-step convergence, or experimental validation evidence is available for the eight trajectories? For representative higher-frequency and larger-amplitude conditions, how sensitive are integrated heat release and its gain and phase to the discretization settings? This evidence would directly affect my assessment of the dataset as a reference benchmark.

1 comment

Reviewer DPx8

1. Please provide the training, context, and validation-scoring indices for a representative block, explaining how approximately 60 held-out samples support a 150-step validation rollout. Does the rollout cross training–validation boundaries? Which samples are used to fit normalization statistics, POD bases, and autoencoders?

We thank the reviewer for this question. It exposed an inaccurate sentence in Section 4: the manuscript states that “30% of each block is held out for validation”. Instead, 30% of the blocks are held out, each one as a whole, so every validation block contains 200 consecutive snapshots rather than ≈60\approx 60. We have corrected the text and added the details below to the appendix. Additionally, to remove any ambiguity we decided to publish the paper's code, which can further clarify the procedure (FireMark code).

Split. Each sine sweep has 4001 snapshots (t=0t=0–22 s, Δt=0.5\Delta t=0.5 ms), divided into 20 consecutive blocks of 200 snapshots (the last one has 201). Six evenly spaced blocks are held out for validation, with snapshot indices [200,400)[200,400), [1000,1200)[1000,1200), [1600,1800)[1600,1800), [2200,2400)[2200,2400), [3000,3200)[3000,3200) and [3600,3800)[3600,3800); the splits are the same for both sweeps and for all models. The remaining 14 blocks form the training set: [0,200)[0,200), [400,1000)[400,1000), [1200,1600)[1200,1600), [1800,2200)[1800,2200), [2400,3000)[2400,3000), [3200,3600)[3200,3600) and [3800,4001)[3800,4001). Each sweep therefore contributes 2801 training and 1200 validation snapshots.

Representative block

Consider the validation block of indices [1000,1200)[1000,1200). The input consists of a window of max⁡(Nx,Ni)+1<50\max(N_x,N_i)+1\lt 50 past values followed by K=150K=150 recursively predicted targets, thus there is no training-validation leakage. For POD–ARX (Nx=5N_x=5, Ni=9N_i=9). The model reads max⁡(Nx,Ni)+1=10\max(N_x,N_i)+1=10 snapshots, x1000,…,x1009x_{1000},\dots,x_{1009}, together with the prescribed inputs; it then predicts recursively for 150 steps, scoring the time steps x1010,…,x1159x_{1010},\dots,x_{1159} against the true snapshots. Each validation rollout starts from true snapshots, whereas the test protocol starts from the steady flame.

Samples used for fitting. The normalization statistics, the POD basis, the autoencoder and forecaster weights are all fitted on the training blocks only. The validation blocks are used only for (i) Hyperparameter optimization; and (ii) early stopping. The test trajectories are not used for any fitting or selection.

2. What mesh-convergence, internal time-step convergence, or experimental validation evidence is available for the eight trajectories? For representative higher-frequency and larger-amplitude conditions, how sensitive are integrated heat release and its gain and phase to the discretization settings? This evidence would directly affect my assessment of the dataset as a reference benchmark.

We agree with the reviewer that numerical verification and experimental validation are important for establishing the reliability of a benchmark dataset. FireMark builds on a flame configuration that has been extensively investigated in previous experimental and numerical studies. In particular, Mohammed et al. (1998), ref. 22 of the manuscript, reported good agreement between experimental measurements and numerical predictions for the same axisymmetric coflow flame under sinusoidal forcing at 20 Hz. More recently, D'Alessio et al (2020), ref. 9 of the manuscript, employed the same configuration with laminarSMOKE and selected a mesh of approximately 25,000 cells following a mesh-sensitivity study. Their transient simulations considered forcing frequencies up to 80 Hz, covering the entire frequency range explored in FireMark. These previous studies provide a relevant basis for the numerical setup and mesh resolution adopted here.

For the present dataset, the solver uses an adaptive time step controlled by a maximum Courant number of 0.05. However, we have not performed additional mesh- or time-step convergence studies, nor a dedicated discretization-sensitivity analysis of integrated heat release and its gain and phase for the highest-frequency and largest-amplitude cases. Similarly, we have not carried out a new quantitative comparison with experimental measurements. We clarify this in the dataset description, section 3 of the manuscript.

Reviewer_5E3MOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces a benchmark dataset for flame reaction flow, which may be of interest in certain areas. However, I found the paper was poorly written and lacked enough information to justify the data quality. While the transient flow and additional flow variables may pose a new challenge, the dataset's coverage and size are too limited to serve as a common benchmark (as the authors mention) for the machine learning community to improve models.

Critical strengths
  1. An interesting region where a good benchmark dataset and experiments for ML models are absent.
Critical weaknesses
  1. As a benchmark dataset paper, the reasons for choosing the selected cases, the detailed simulation settings, and the validation/verification of the methods are absent.
  2. The benchmark considers only a 2D axisymmetric laminar flame with a single varying inlet velocity and has very limited diversity (2 simulations for training), which limits the use of the true power of Neural Operators for learning highly non-linear dynamics and biases the performance comparison.
Questions for authors

(1) The paper lacks a schematic plot to show the computational domain, making it hard for readers to understand the flow setting. (2) The choice to use very complex chemical reaction formulas (1698 reactions) but a very simple laminar flow setting is confusing and needs justification. Since there exist simplification methods, why not use them to do the simulation? (3) The cost of simulation for one case was not reported, making it difficult to evaluate the necessity of introducing ML models. (4) The mesh settings are not reported (is it uniform?). How is the mesh convergence? (5) Figure 1 is very unprofessional, as the spots at the top show it is a screenshot. Also, in the figure, particularly in the U1 field, there are strong numerical oscillations near the inlet step, suggesting the numerical method may not be optimal. (6) How general is the conclusion for the benchmark test? Will this hold when having more data points? (7) The discussion section mentioned “The restarted protocol of Section 4, comparing one-step and closed-loop errors, can help distinguish this effect.” But there is no related content in that section.

1 comment

(simpler markdown structure due to the constraint on the number of characters)

1.

We thank the reviewer for the observation. We agree and have added a schematic of the burner and computational domain, showing the fuel nozzle, coflow inlet, and axisymmetric geometry. The figure is in section 3 (current Figure 1).

2.

We thank the reviewer for the interest in the chemical mechanism. In this work, we chose to retain detailed chemistry to provide an accurate representation of the species and their coupling with transport and thermodynamics. The laminar configuration does not make detailed chemistry any less relevant: the level of chemical detail is independent of whether the flow is laminar or turbulent. Although the benchmark releases only a subset of the simulated species and flow variables, the full mechanism is retained in the original simulations. This makes the complete dataset potentially useful for further studies, including the development of ROMs for chemical kinetics. We have added a short justification to the modified manuscript, in section 3.

3.

We thank the reviewer for the observation about the computational cost. Each simulation required approximately 127–508 wall-clock hours on 20 cores, corresponding to a total of approximately 38,250 CPU-hours for the eight trajectories. We have added this information in section 3 and more details can be found in section B of the appendix.

4.

We thank the reviewer for the interest about the mesh. The mesh is a non-uniform, locally refined Cartesian grid with approximately 21,000 cells. Refinement is concentrated near the fuel nozzle, where stronger temperature and species gradients are expected. We have added the mesh description and a visualization to the revised manuscript in section 3 (current Figure 1). We did not perform an independent mesh-convergence study for the present dataset. The mesh resolution follows D'Alessio et al (2020), ref. 9 of the manuscript, who reports a mesh-sensitivity assessment for the same flame configuration.

5.

We thank the reviewer for pointing out possible misunderstandings. Figure 1 (current Figure 2) is not a screenshot, the visible dot in the picture is the point of Q˙\dot{Q} and represents the volumetric derivative of the instantaneous heat QQ released in every cell. Regarding the apparent oscillations in U1, these arise from flow separation and recirculation near the coflow inlet edge, rather than from numerical artifacts. Their prominence in the U1 field shown in Figure 1 (current Figure 2) is mainly due to its narrower colorbar range compared with U3.

6.

We thank the reviewer for letting us clarifying these aspects of the paper; we have added these points in Section 8 of the manuscript. Within the benchmark, the main conclusion holds on every test input, not only on average. Per test input, ARX has a lower nRMSE than every neural forecaster on the same compressor (Table 5). For the heat-release error εQ˙\varepsilon_{\dot Q} (Table 6), ARX performs better in 83 of the 84 comparisons (14 neural forecaster across four different compressor, on 6 test signals →14×6\to 14\times6 comparisons); the only exception is ViTAE on the test consisting in the step function at A=0.5A=0.5 (GRU 0.299 vs. ARX 0.304). The six test inputs cover amplitudes inside (A=0.3A=0.3) and outside (A=0.5A=0.5) the training range, and excitation type never seen in training (step), so the ranking is not tied to a particular test condition.

We do not claim generality beyond the setting considered in this work, and, adding more data points could possibly increase the performance of the neural models. However, data efficiency is itself a relevant criterion in this application. Surrogate models in combustion are developed to support goals such as lower emissions and each additional high-fidelity trajectory costs about 4,780 CPU-hours, on top of the longer training of neural models. Any accuracy gained from more data must therefore be weighed against the computational cost, and the associated energy consumption and emissions, of generating that data and training on it. Thus, methods that perform well from few trajectories are the most useful in practice. The benchmark can nevertheless be extended: the release includes the mesh and the solver setup needed to generate new trajectories.

7.

We thank the reviewer for spotting this. The description of the restarted evaluation protocol was missing from Section 4. Since the protocol is short to describe, we have now included its description directly in the Discussion (Section 8), where it is used. Additionally, we added in Appendix E ("Additional Results") a further analysis of the error growth for representative models throughout the test rollout.

Reviewer_jXVGOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduced a benchmark for reaction flow. The predicted fields include fluid dynamics such as pressure, and chemical state, such as statistics of different chemical species. The authors additionally introduced variability in the forcing value, which can itself be time-dependent. The authors test both classical and neural baselines.

Critical strengths
  • The text is well written and clear.
  • The inclusion of classical baselines is interesting.
Critical weaknesses
  • The neural baseline models do not seem to work well. While divergence in rollout can be expected, I am not sure whether technique such as time-bundling [1] is tested, which can significantly reduce the actual rollout steps. Besides, the commonly adopted Markov training is only tested for FNO (i.e., identity encoder in the paper). Since the proposed training (latent evolution and rollout loss) does not seem to work well, it may be worth testing the more commonly-adopted training, like in PDEArena [2]'s codebase. UNet models (such as UNets with ConvNext blocks) may also worth testing.
  • There is no visualizations about the forecasted fields, which can help understand the prediction error level qualitatively.
  • There is only two training trajectories, which is small compared to other benchmarks. Also since the forcing in test is not covered in training (such as step function), I would doubt the transferability given the small number of samples.

[1] Message Passing Neural PDE Solvers [2] Towards Multi-spatiotemporal-scale Generalized PDE Modeling

Questions for authors
1 comment

Reviewer jXVG

1. Why only 8 trajectories? What is the computation cost for data generation?

We thank the reviewer for the question. The eight trajectories were designed to assess model performance across both in-distribution (ID) and out-of-distribution (OOD) test conditions, while probing different aspects of the flame's dynamic response. The training sine sweeps cover a broad frequency range, allowing models to learn how the flame responds to different frequencies. This is particularly relevant for forecasting, as sinusoidal components provide fundamental building blocks for characterizing the response to more complex time-varying inputs. Step signals complement this by probing the response to abrupt changes in operating conditions. Together, these inputs provide a structured basis for assessing forecasting performance under different excitation dynamics.

We varied frequency rather than amplitude during training because amplitude changes are generally easier to interpolate, acting primarily as a scaling factor on the response. The test signals were chosen to assess complementary aspects of prediction. Fixed-frequency sinusoids provide a standard forecasting task and test interpolation within the frequency range covered during training (see DOI: 10.1002/nme.70058). Step inputs represent abrupt changes in operating conditions, relevant transients in practical combustion systems such as furnaces or engine burners.

We also selected two test frequencies (10 and 40 Hz) to probe the flame response in different frequency regimes. The simulations required approximately 127–508 wall-clock hours each, running on 20 cores. The total dataset generation cost was approximately 38,250 CPU-hours (about 4,780 CPU-hours per trajectory). We have added this information in Section 3 of the revised manuscript and in Section B of the Appendix.

2. Are the auto-encoders trained together with the propagator or they are trained separately?

We thank the reviewer for exposing an unclear aspect of the paper. To eliminate any ambiguity we decided to release the paper's code (FireMark code). The training is done in three consecutive phases, each with its own hyperparameter search (Section 5.3). The autoencoder is first trained alone to minimize the reconstruction error, then the forecaster is trained on the frozen latent representations produced by the autoencoder from the previous phase. Finally, when both autoencoder and forecaster are neural networks, they are fine-tuned together. For the latter stage, the loss L(θF,θE,θD)\mathcal{L}(\theta_F,\theta_E,\theta_D), depends on θF,θE,θD\theta_F,\theta_E,\theta_D, which are respectively the parameters of the forecaster, the encoder and decoder, so when it is backpropagated, it updates θF,θE,θD\theta_F,\theta_E,\theta_D simultaneously. For closed-form propagators, no joint fine-tuning is performed. This procedure is described in paragraph "Training and model selection." of Section 4, Sections 5.1 and 5.3 of the manuscript. Nevertheless, we decided to release the code to remove any ambiguity.

3. Are these related?

arXiv:2512.18595, OpenReview

We thank the reviewer for pointing us to REALM and Open-CK benchmarks. Although both works address fire-related dynamics, they target different applications from FireMark. REALM evaluates surrogates for the autonomous dynamics of flames across a wide range of applications—without any control inputs. Similarly, Open-CK focuses on the free propagation of large-scale industrial fires across different fire scenarios. Instead, FireMark is designed to support the development of surrogate models that act as digital twins of a combustion burner. Therefore, it focuses on predicting the dynamic response of a laboratory-scale coflow flame to controlled input forcing. This distinction is reflected in the modeling approaches: FireMark investigates how the flame responds to prescribed input signals, whereas REALM and Open-CK addresses flame behavior under prescribed conditions. We have added both works in the Related work section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.