DeepFlame: Neural Combustion & Flame Detection

• DeepFlame is an interdisciplinary framework integrating deep learning with CFD and computer vision for accelerated reactive-flow simulations.
• It employs neural surrogates to replace stiff ODE integrations and advanced semantic segmentation for real-time flame detection.
• Applications include high-performance rocket engine simulations, industrial fire risk monitoring, and autonomous agents for combustion analysis.

DeepFlame refers to a family of deep learning–empowered frameworks and algorithms at the intersection of reactive-flow simulation, high-fidelity combustion modeling, flame video detection, and nonlinear system identification. The term encompasses open-source computational platforms for machine learning–accelerated CFD, supercomputing applications of flame DNS, advanced neural architectures for flame dynamics, and vision-based flame detection systems. Collectively, DeepFlame systems are designed to exploit neural network speed and representation power for solving, modeling, and monitoring complex reacting flows and fire phenomena, across both physical simulation and computer vision domains (Mao et al., 2022, Guo et al., 26 Aug 2025, Wu et al., 2024, Pérez-Guerrero et al., 2022, Meleti et al., 2023, Aslan et al., 2019).

1. Origins and Major Implementations

The first explicit DeepFlame platform was introduced as a C++ open-source framework integrating OpenFOAM (fluid dynamics), Cantera (detailed thermochemistry), and libTorch (PyTorch C++ API) to enable ML-accelerated reactive-flow simulations (Mao et al., 2022). This foundational DeepFlame provides modular support for traditional ODE integration (CVODE/Cantera) and deep neural network (DNN) chemistry surrogates, facilitating both accuracy and two-orders-of-magnitude computational acceleration in canonical combustion problems.

Subsequent developments include exascale-optimized DeepFlame implementations coupling neural surrogates for both real-fluid EoS and chemical kinetics (e.g., PRNet and ODENet) to reach trillion-cell scales on supercomputers (Guo et al., 26 Aug 2025), and advanced neural architectures such as Dual-Path attention networks for time-domain thermoacoustic response reconstruction (Wu et al., 2024).

In parallel, vision-oriented DeepFlame systems have been developed for flame detection and segmentation, including GAN-based video recognition models and semantic segmentation pipelines for industrial fire risk (Aslan et al., 2019, Pérez-Guerrero et al., 2022).

2. Core Methodologies

2.1 Reacting-Flow Simulation with ML Surrogates

DeepFlame platforms solve the fully compressible Navier–Stokes equations coupled with species and energy conservation:

• Mass: $$\partial\rho/\partial t + \nabla\cdot(\rho\mathbf{u}) = 0$$
• Momentum: $$\partial(\rho\mathbf{u})/\partial t + \nabla\cdot(\rho\mathbf{u}\otimes\mathbf{u}) = -\nabla p + \nabla\cdot\boldsymbol{\tau}$$
• Species: $$\partial(\rho Y_i)/\partial t + \nabla\cdot(\rho Y_i \mathbf{u}) = -\nabla\cdot \mathbf{J}_i + \dot\omega_i$$
• Total energy: $$\partial(\rho E)/\partial t + \nabla\cdot[(\rho E + p)\mathbf{u}] = \nabla\cdot(\lambda\nabla T + \boldsymbol{\tau}\cdot\mathbf{u}) + \sum_i h_i\dot\omega_i$$

The distinguishing element is the replacement of stiff chemical kinetics ODEs and iterative real-fluid equation-of-state evaluations with ML surrogates:

• PRNet: Deep neural surrogate for Peng–Robinson EoS and transport, trained on $$(E, p, \{Y_i\}) \rightarrow (\rho, T, \mu, \alpha, ...)$$ data.
• ODENet: Deep surrogate for chemical source terms, mapping $(T, p, \{Y_i\}) \rightarrow \{\dot\omega_i\}$ (Guo et al., 26 Aug 2025, Mao et al., 2022).

2.2 Semantic Segmentation and Video-Based Flame Detection

Computer vision DeepFlame systems implement encoder–decoder networks (e.g., UNet, UNet++ with attention gates), GANs, and spatio-temporal convolutions for flame semantic segmentation, risk assessment, and obscured fire detection (Aslan et al., 2019, Pérez-Guerrero et al., 2022, Meleti et al., 2023):

• Temporal-slice construction for GAN-based flame/no-flame classification.
• UNet-family architectures for per-frame infrared segmentation, yielding real-time geometric metrics (height, area, lift-off).
• 3D convolutional CNNs with attention for multi-frame, temporally-aware segmentation of wildfire flames, trained via IR-masked RGB data to capture occlusions and motion cues.

2.3 Nonlinear Dynamical System Identification

DeepFlame also refers to specialized neural architectures (Dual-Path LSTM–attention hybrids) for reconstructing nonlinear time-domain thermoacoustic flame responses from limited frequency-sweeping datasets (Wu et al., 2024).

3. High-Performance Computing and System Design

DeepFlame platforms leverage exascale architectures through several techniques (Guo et al., 26 Aug 2025):

• Block-sparse matrix structures and two-level (MPI, thread) domain decomposition (SCOTCH, Cuthill–McKee) for CFD.
• Batched, mixed-precision neural inference kernels (FP16/FP32), double buffering, and vectorized SpMV.
• Exascale I/O schemes: in-memory mesh refinement, grouped parallel I/O using leader ranks, and foam file index precomputation.
• Dynamic load balancing (DLB) and adaptive mesh refinement (AMR) built on OpenFOAM's infrastructure, extended for 1D–3D, minimize computational costs associated with spatial heterogeneity in reaction zones (Mao et al., 2022).

On Sunway and Fugaku, DeepFlame achieves up to 1.18 EFlop/s (21.8% of peak) on 618B cells (Sunway), with time-to-solution reductions of $10^4$–$10^5\times$ over legacy supercritical simulation codes (Guo et al., 26 Aug 2025).

4. Applications in Combustion Modeling and Fire Risk

4.1 Virtual Experiments for Rocket Propulsion and Energy

High-fidelity DeepFlame simulations provide predictive digital twins for rocket engines, allowing full-chamber, multi-injector design studies and virtual testing of mixing, instabilities, and cooling strategies with true real-fluid, detailed-chemistry fidelity (Guo et al., 26 Aug 2025). One example is the unprecedented simulation of a LOX/\ce{CH4} rocket engine with 127 injectors at trillion-cell resolution.

4.2 Industrial Fire Monitoring and Risk Management

Semantic segmentation DeepFlame systems process live IR imagery to extract key geometric and radiative features from jet flames for safety dashboards and automated emergency protocols, offering IoU ≈ 0.87, area errors <8%, and sufficient throughput for sub-Hz fire dynamics (Pérez-Guerrero et al., 2022).

4.3 Autonomous Research Agents

LLM-enabled orchestration for combustion modeling (e.g., via the FlamePilot agent) treats DeepFlame as a back-end solver, automating setup, simulation, error analysis, and optimization of reactive-flow cases (Xiao et al., 4 Jan 2026). The system can parse literature, prepare DeepFlame-compatible case files, execute simulations, and autonomously perform corrective or parameter studies.

5. Neural Surrogates, Model Training, and Validation

DeepFlame chemical surrogates typically employ fully connected feed-forward networks (matrices of widths 512–4096) with normalization and Box–Cox preprocessing. Models are trained from datasets spanning $T \in [300, 3000]$ K, $p \in [0.1, 50]$ atm and randomized initial species, with target increments computed via CVODE/Cantera integrations (Mao et al., 2022, Guo et al., 26 Aug 2025). PRNet/ODENet match reference tables to $O(1\%)$ accuracy and reduce chemistry evaluation cost by $10$–$50\times$.

Flame CNNs and GANs use temporal-slice, 3D-convolution, or attention-based modules, with segmentation or classification accuracy validated on curated datasets (e.g., FLAME2, FLAME3) and benchmarked on IoU, Dice, Hausdorff, and real-time throughput metrics (Aslan et al., 2019, Meleti et al., 2023, Pérez-Guerrero et al., 2022).

Thermoacoustic networks are trained on frequency-sweep simulation data, employing dual-path (chronological LSTM and temporal-detail attention) feature extractors. These architectures achieve mean time-domain relative errors below 7% on unseen single-frequency signals, with robust generalization to strong nonlinearities (Wu et al., 2024).

6. Limitations and Future Directions

• Current DeepFlame platforms are limited to gas-phase, single-GPU/DCU surrogates, and 0D/1D–3D CFD flows; distributed ML inference and multiphase (spray) combustion are planned extensions (Mao et al., 2022, Guo et al., 26 Aug 2025).
• Computer vision systems may require large, scenario-diverse annotated datasets—IR-derived ground truth and narrow camera setups are limiting factors (Meleti et al., 2023).
• Nonlinear system-identification DeepFlame models are currently 0D; extension to spatiotemporal field mappings and incorporation of physical priors are natural progression points (Wu et al., 2024).
• For LLM-based workflow orchestration, handling arbitrary user-defined reactor models and mesh topologies is not yet fully automated (Xiao et al., 4 Jan 2026).

7. Representative System and Performance Table

DeepFlame Variant	Application Domain	Key Algorithms/Features
DeepFlame (Mao et al. 2022)	CFD, combustion DNS/LES	OpenFOAM+Cantera+libTorch, CVODE/DNN chemistry
DeepFlame Exascale (Zhang et al. 2025)	Supercritical flames, HPC	PRNet/ODENet surrogates, block-sparse, FP16, I/O
Vision DeepFlame (Turgay 2019 et al.)	Flame detection/video	DCGAN, spatiotemporal CNN, temporal slices
DeepFlame–Risk (López-Marrón 2022 et al.)	Jet fire segmentation	UNet/Att-UNet/UNet++, IR imaging, geometric ext.
DeepFlame Dual-Path (Liu et al. 2024)	Thermoacoustics	LSTM+cross-attention, frequency-sweep training

Each variant adheres to domain-specific data, algorithms, and validation protocols but is united under the principle of leveraging modern neural surrogates and integration strategies for physical modeling, perception, or nonlinear identification.