Enterprise AI Analysis

Introduction: The Decarbonization Challenge

Decarbonizing combustion-based energy conversion remains a central engineering and societal challenge because combustion systems are deeply embedded in aviation, marine transport, heavy-duty mobility, industrial heat, and dispatchable power generation. In the near to medium term, material emissions reduction is therefore expected to come from a combination of (i) low-carbon and carbon-free fuels (e.g., hydrogen, ammonia, e-fuels, and oxygenated blends), (ii) device-level efficiency improvements, and (iii) robust emissions control strategies spanning in-cylinder measures, combustor staging, and aftertreatment. Recent syntheses underscore both the promise and complexity of these pathways: hydrogen introduces new safety and operability constraints [1], ammonia offers carbon-free storage advantages but raises ignition and NOx trade-offs [2-3], and oxygenated e-fuels such as oxymethylene ethers can shift soot/NOx behaviour while challenging legacy modelling assumptions [4]. These transitions do not eliminate modelling difficulties; they often amplify them by widening the operating envelope (i.e., pressure, temperature, dilution, and stratification) and introducing additional kinetic pathways for nitrogen chemistry, low-temperature oxidation, and pollutant formation.

At the same time, the modelling problem is becoming more urgent. Regulatory pressure on NOx and particulate matter, which is coupled with climate-driven limits on lifecycle CO2, pushes designers toward multi-objective decisions under uncertainty: stable ignition versus flashback, efficiency versus knock, and low NOx versus low unburned-fuel slip (especially for ammonia). For engines and gas turbines, these objectives are tightly coupled to turbulence-chemistry interaction, heat loss and wall effects, and transient operation. The literature has long emphasized that “single-physics” or purely steady analyses are rarely sufficient at device scale, motivating multiscale and multiphysics simulation strategies spanning detailed kinetics, turbulence modelling, sprays, and wall interactions [5-6]. The implication is straightforward: if the community cannot accelerate the credible predictive simulation and validation pipelines, the design loop will remain too slow to explore the fuel-device-control co-design space that decarbonization demands.

Foundations of AI for Combustion Modelling

Recent advances in AI and ML have introduced new paradigms for modelling complex reacting-flow systems across multiple spatial and temporal scales. Combustion processes are inherently multiscale, spanning atomistic reaction dynamics, mesoscale chemical kinetics, turbulent reacting flows, and device-level performance. The increasing availability of large datasets from high-fidelity simulations (e.g., direct numerical simulation (DNS), large-eddy simulation (LES), molecular dynamics (MD), and density functional theory (DFT)) and experimental diagnostics has created opportunities for data-driven approaches to augment or replace traditional modelling strategies [9, 36].

Within this context, machine learning models can be interpreted as approximators of mappings between physical quantities. For instance, a surrogate model may approximate mappings from thermochemical state variables to reaction source terms, from spatial fields to turbulence closures, or from boundary conditions to full flow solutions. Such mappings may be represented using a parameterized function fθ(x), where the model parameters θ are learned from data. Deep learning architectures provide flexible approximators capable of capturing nonlinear relationships in high-dimensional datasets, making them well suited to the complexity of combustion systems [37].

Despite their flexibility, purely data-driven models often struggle to maintain physical consistency when applied outside their training domain. Consequently, the integration of physical knowledge through constraints, hybrid modelling, or governing equations has emerged as a central theme in scientific machine learning for combustion [38]. This section briefly introduces the principal learning paradigms, model architectures, and physics-constrained approaches that underpin AI-enabled combustion modelling.

AI Across Scales in Combustion Modelling

Multiscale combustion is often discussed in terms of CFD and turbulence-chemistry interaction, but the credibility of any reacting-flow model ultimately rests on chemistry: which pathways are active, how fast they proceed, and how sensitive they are to temperature, pressure, mixture composition, and dilution. At the smallest scales, these questions are governed by the potential-energy surface and the resulting reactive dynamics [51-52]. The past few years have therefore seen rapid growth in AI at the molecular/atomistic level, not as a replacement for combustion modelling, but as a way to (i) approach quantum accuracy at MD cost, and (ii) extract chemically meaningful information (i.e., pathways, intermediates, and rate parameters) that can be carried upward into reduced mechanisms or closures [51].

Experimental combustion remains the primary arbiter of credibility for multiscale modelling because it supplies the measurable targets, i.e., flame stabilisation limits, ignition delay, burning velocity, pollutant markers, and space-time resolved fields, against which chemistry models and CFD closures are ultimately judged. In practice, however, many quantities of interest are only partially observable: optical access is limited, signals are noisy at high pressure or high speed, multi-parameter laser diagnostics are expensive and difficult to synchronise, and three-dimensional (3D) reconstructions are often ill-posed. Recent work therefore uses machine learning not as a substitute for diagnostics but also as an additional “inference layer" that (i) improves the signal fidelity (denoising/deconvolution), (ii) reconstructs the unmeasured quantities from correlated measurements (virtual sensing), and (iii) accelerates the inverse problems such as tomography so that they become feasible on experimental timescales. These developments are particularly relevant for emerging fuels (e.g., NH3/H2 blends) because the most decision-relevant outputs can be emissions- or stability-related, where the experimental coverage across operating space is typically sparse [74].

CFD remains the principal predictive framework for combustor and engine design as it resolves the coupled transport of momentum, heat, and chemical species within complex geometries over engineering-relevant time scales. Despite its maturity, two persistent bottlenecks continue to constrain both accuracy and computational efficiency: (i) turbulence-chemistry closure modelling in RANS and LES formulations and (ii) the numerical cost associated with detailed chemical kinetics, particularly stiff ODE integration and large reaction mechanisms. Consequently, recent research has increasingly focused on integrating artificial intelligence techniques into reacting-flow CFD—not as a substitute for conservation-law-based solvers, but as a structured augmentation strategy aimed at improving the closure fidelity, accelerating chemistry evaluation, and enabling rapid-response surrogate models for parametric studies and optimisation tasks [90-92].

Quantitative Synthesis of ML Surrogate Performance

This section draws on a curated corpus of 125 peer-reviewed studies published between 2020 and 2026 that investigate machine-learning surrogates for combustion and thermal decomposition modelling across scales. The papers were compiled through manual screening with the aim of capturing both methodological advances and application-driven deployments relevant to multiscale combustion—spanning chemistry and kinetics, turbulence-chemistry interaction, reacting-flow transport, and device-level configurations where performance and emissions are key outcomes. The corpus is therefore intentionally broad: it includes work focused on accelerating detailed-chemistry calculations, learning closures for continuum solvers, building reduced-order surrogates for reacting flows, and enabling near real-time inference for control, diagnostics, or optimisation tasks.

Since the literature uses heterogeneous assumptions, datasets, and validation protocols, the following analysis is framed as a quantitative synthesis rather than a single benchmark comparison. Each paper is assigned a Ref. number (used consistently throughout the tables and figures), and each study is categorised along three complementary dimensions: Surrogate model family (method taxonomy), Application domain (where the surrogate is used), and Evaluation signals (how performance is quantified).

Combustion surrogate models are rarely evaluated along a single dimension. In most multiscale workflows, surrogates are introduced to reduce the cost of repeatedly evaluating chemistry, closures, or reacting-flow response surfaces inside larger simulation or optimisation loops. Consequently, the model value is determined jointly by predictive fidelity and computational gain, and the literature consistently frames performance as a trade-off between these objectives.

Current Limitations and Future Directions

Despite the rapid progress surveyed in Sections 3 and 4, the current AI-combustion ecosystem remains fragmented, with many of the most eye-catching speed-ups achieved under narrow conditions, non-standard baselines, or limited validation envelopes. The next step is therefore not only to invent new models but to address the structural limitations in data, metrics, workflows, and trust so that AI becomes a reliable, scale-bridging component of combustion science rather than an ad-hoc accelerator [109].

The corpus analysis in Section 4 highlights that "speed-up" is often defined heterogeneously, with studies variously quoting wall-clock ratios to detailed ODE solvers, comparisons to reduced-mechanism tables, or gains relative to low-order surrogates, frequently on different hardware and without counting training cost. As a result, nominally similar numbers can reflect profoundly different trade-offs, making it difficult to construct a robust Pareto frontier of accuracy versus efficiency across AI architectures and application regimes.

A second limitation is that most benchmarks remain task-local: models are typically tuned and evaluated on a single canonical flame, reactor configuration, or fuel blend, with only modest probing of distribution shift across facilities, diagnostics, fuels, or operating envelopes. Sections 3.2 and 3.3 already illustrate how the measurement manifolds and solver behaviour can change under soot-loading, pressure, or geometry variation, but only few studies in the corpus explicitly quantify robustness to these shifts or report uncertainty estimates aligned with operational decision-making.

Looking forward, one of the most promising directions is the emergence of agentic and virtual-laboratory paradigms for combustion science, in which autonomous but supervised AI systems help design simulations, monitor execution, analyse outputs, quantify uncertainty, and propose the next most informative cases. If developed with strong physics constraints, transparent logging, and human oversight, such frameworks could help address many of the current bottlenecks identified in this review, including inconsistent workflows, slow campaign turnaround, and limited reproducibility. Ultimately, the real promise of AI in combustion lies not simply in making existing simulations faster but in enabling a more rigorous, reproducible, and predictive science of clean reactive systems across scales.

Nomenclature

Abbreviations

AI: Artificial intelligence
ML: Machine learning
ANN: Artificial neural network
MLP: Multilayer perceptron
LES: Large-eddy simulation
MD: Molecular dynamics
DFT: Density functional theory
PINN: Physics-informed neural network
DNN: Deep neural network
DeepONet: Deep operator network
CNN: Convolutional neural network
FNO: Fourier neural operator
GNN: Graph neural network
NODE: Neural ordinary differential equation
RL: Reinforcement learning
GPR: Gaussian process regression
CFD: Computational fluid dynamics
RANS: Reynolds-averaged Navier-Stokes
DNS: Direct numerical simulation
TCI: Turbulence-chemistry interaction
FPV: Flamelet/progress-variable
POD: Proper orthogonal decomposition
FGM: Flamelet generated manifold
PCA: Principal component analysis
VAE: Variational autoencoder
OH-PLIF: OH planar laser-induced fluorescence
CH-PLIF: CH planar laser-induced fluorescence
CH2O-PLIF: Formaldehyde planar laser-induced fluorescence
PIV: Particle image velocimetry
FES: Flame emission spectroscopy
NNP: Neural network potential
MLIP: Machine-learned interatomic potential
ReaxFF: Reactive force field
DPMD: Deep Potential molecular dynamics
NOX: Nitrogen oxides
NH3: Ammonia
CA50: Crank angle at 50% heat release
RFR: Random forest regression
SVM: Support vector machine
LSTM: Long short-term memory network
U-Net: Encoder-decoder convolutional neural network with skip connections
CombML: Combustion machine learning

Symbols

fθ(x): Parameterized model/surrogate mapping
k(T): Temperature-dependent reaction-rate coefficient
X: Input state or feature vector
T: Temperature
θ: Trainable model parameters
T̃: Filtered or resolved temperature
Ea: Activation energy
Ŷco: Filtered carbon monoxide mass fraction
ΔH: Reaction enthalpy
Z̃: Filtered mixture fraction
ν: Characteristic frequency
ώ: Reaction or burning rate
Φ: Equivalence ratio
ώNN: Neural-network-predicted filtered burning rate
R²: Coefficient of determination
ώF: Filtered tabulated-chemistry burning-rate estimate
DSF: Downsampling factor
ώFC: Alternative filtered tabulated-chemistry burning-rate estimate
SNR: Signal-to-noise ratio
ώ*: Reference filtered burning rate

Subscripts

α: Activation-related quantity in Ea
F: Filtered/tabulated reference quantity
CO: Carbon monoxide
FC: Alternative filtered/tabulated combustion quantity
g: Global quantity, where used
NN: Neural-network prediction
0: Initial condition, where used
ref or *: Reference value