FlamePilot: Combustion, CFD & LES

• FlamePilot is a multifaceted concept encompassing pilot flame stabilization in stratified swirl burners, AI-driven CFD workflow automation, and LES for transcritical real-fluid combustion.
• It employs advanced numerical methods, such as the double-flux scheme and FPV approach, to accurately simulate complex reactive flows under extreme conditions.
• The integrated AI agent autonomously processes literature data to generate, optimize, and audit simulation workflows, achieving superior performance metrics in combustion research.

FlamePilot encompasses three distinct paradigms in combustion and computational sciences: the stabilizing pilot flame in stratified swirl burners, an LLM-enabled autonomous workflow agent for combustion CFD, and a large-eddy simulation (LES) framework for transcritical real-fluid reacting flows leveraging the flamelet progress variable (FPV) approach. This article systematically covers their definitions, mechanisms, methodologies, performance, and interrelationships as established in primary research literature.

1. Definition and Contexts of "FlamePilot"

The term "FlamePilot" has three principal referents in contemporary literature:

1. In experimental combustion, it denotes the rich, centrally injected pilot flame in stratified swirl burners, critical for flame stabilization and ignition characteristics (Han et al., 2021).
2. In AI-enabled CFD, FlamePilot designates a modular LLM agent system capable of translating scientific literature into executable, self-correcting simulation workflows for OpenFOAM and DeepFlame (Xiao et al., 4 Jan 2026).
3. In numerical modeling, FlamePilot is a finite-volume code for real-fluid, transcritical combustion, based on the FPV method and double-flux numerics (Ma et al., 2017).

Each thread is rooted in the need for robust ignition stabilization, automated workflow management, or thermodynamically consistent simulation at extreme conditions.

2. FlamePilot in Stratified Swirl Burners: Stabilization and Dynamics

In the Beihang Axial Swirler Independently-Stratified (BASIS) burner, the pilot flame ("FlamePilot") issues from an inner axial swirler (swirl number $S \approx 0.68$; main $S \approx 0.50$) as a premixed methane–air jet. It serves as the primary anchor in lean premixed operation by:

• Establishing a hot PRZ (primary recirculation zone) for reliable main flame ignition and extending blow-off limits.
• Operating at a richer equivalence ratio ($\phi_p$) than the main stage, thus creating a spatially stratified thermal field.

Three canonical operating regimes are:

Mode	Configuration	Notable Flame Behavior
Pilot-only (PfMc)	Main closed, $\phi_p = 0.85$	V-shaped, anchored on center body
Pilot+Air (PfMa)	Main air only, $\phi_p = 0.85$	M-shaped, thickened by shear/dilution, strong 159 Hz oscillations
Stratified (PfMf)	Both fueled, $\phi_p = 0.85$, $\phi_m = 0.63$	Beating oscillations at $f_1 \approx 198$ Hz, $f_2 \approx 223$ Hz

Key transitions include V-to-M shape transformation under main air dilution, and quasi-periodic beating oscillations in stratified operation, quantified by a beating index $B \approx 0.86$. LES with ReactingFoam and OpenFOAM, using the WALE turbulence model and a 4-step global methane scheme, accurately reproduces all observed modes and recirculation structures. A two-section acoustic model (OSCILOS) captures coupled- ($f_A$) and decoupled- ($f_B$) modal frequencies matching experimental data (Han et al., 2021).

3. FlamePilot: Autonomous Literature-Aware CFD Agent

The LLM-enabled FlamePilot system automates and manages complex combustion simulation workflows:

• Architecture: A single LLM orchestrator decomposes high-level user goals into executable sub-tasks, leveraging atomic shell tools (e.g., file read/write, log parsing, mesh inspection). It employs domain skills for solver interfaces and PDF-to-parameter extraction.
• Literature ingestion: PDFs are converted to Markdown, then parsed with schema-driven extraction to generate JSON records (geometry, mesh, models, BCs). These are inserted into OpenFOAM dictionary templates.
• Self-correction: Upon execution errors, FlamePilot analyzes logs, proposes fixes (e.g., for negative cell volumes, missing syntax), and iterates until success or max iterations.

This transparent workflow is fully auditable by researchers and amenable to extension via new domain skills (e.g., for other solvers or problem domains) (Xiao et al., 4 Jan 2026).

4. Performance Metrics and Validation in AI-CFD

On the FoamBench-Advanced suite ($N=16$), FlamePilot achieved:

• Executability score $M_{\rm exec} = 1.0$ (all cases ran, surpassing prior scores of 0.625 and 0.750)
• Success rate $S = 0.438$ (higher than prior best of 0.250)
• Mean squared error $M_{\rm NMSE} = 0.469$ (comparable to best prior at 0.406)

A detailed MILD combustion case study demonstrated the pipeline: after ingesting literature data (e.g., geometry, models), running baseline RANS k–ε simulations, and comparing with experiment, the agent autonomously identified underprediction of peak temperature, proposed literature-supported parameter changes (e.g., increasing $C_{1\epsilon}$, refining shear-layer mesh), conducted a 5×5 parameter grid, and converged to within ±5% error of experiment—all under minimal human intervention. All steps, decisions, and corrections are logged explicitly (Xiao et al., 4 Jan 2026).

5. FlamePilot for Transcritical Real-Fluid Simulations

FlamePilot, as described in (Ma et al., 2017), refers to a compressible, finite-volume LES formulation for transcritical combustion, combining:

• Double-flux scheme: Mitigates spurious pressure oscillations arising from nonlinearities in the Peng–Robinson equation of state (EoS), via local freezing of effective specific heat ratio $\gamma_e$ during inviscid flux calculation.
• Entropy-stable hybrid flux: Blends 4th-order central and 2nd-order ENO stencils with a density-gradient sensor, adding dissipation only when needed to ensure $\Delta S \geq 0$.
• Thermochemistry: Employs a flamelet/progress variable (FPV) approach, pretabulating scalar fields $Y_k(Z,C)$, $T(Z,C)$, $e(Z,C)$ from 1D counterflow flames, then correcting for local $(p,T)$ via analytic departure functions according to the cubic EoS.

During simulation, transport equations for mixture fraction $\widetilde Z$, variance $\widetilde{Z''^2}$, and progress variable $\widetilde C$ are solved. Output fields are reconstructed using the $\beta$-PDF (for $Z$) and $\delta$-PDF (for $C$) turbulence–chemistry closure.

Typical applications include LOX/GH₂ mixing and reacting layers at $p = 10$ MPa and density ratios $\sim 80$, with mesh resolutions $5 \times 10^5$ cells. The scheme yields time-averaged and instantaneous profiles in line with AVBP/RAPTOR benchmarks and avoids pressure spikes even at sharp real-fluid gradients (Ma et al., 2017).

6. Methodological Principles and Key Mathematical Definitions

Across all FlamePilot applications, precise mathematical definitions are central:

• Equivalence ratio: $$\phi \equiv \dot m_{\rm fuel}/(\dot m_{\rm air} S_{st})$$, where $S_{st}$ is stoichiometric.
• Swirl number: $$S = \frac{\int_0^R \rho u_z u_\theta r^2 dr}{R \int_0^R \rho u_z^2 r dr}$$
• Acoustic modal frequency: $f_n = \frac{n c}{2 L}$
• Rayleigh criterion: $\int_V \int_0^T p'(x,t) \dot q'(x,t) dt\, dV > 0$ signals instability growth.
• Governing equations for real-fluid combustion: Mass, momentum, energy, $\widetilde Z$, $\widetilde{Z''^2}$, and $\widetilde C$, with details as outlined in (Ma et al., 2017).
• Performance metrics in AI-CFD: Executability $M_{\rm exec}$, success rate $S$, and normalized mean squared error $M_{\rm NMSE}$, as defined above.

7. Transparency, Extensibility, and Broader Implications

The FlamePilot agent is intrinsically interpretable—every model input, workflow change, and corrective action is logged in human-readable patches. Its modular "skills" design facilitates rapid adaptation across scientific domains by swapping in solver conventions and literature-extraction templates. The reliance on atomic shell tools, rather than opaque APIs, supports rigorous auditability. A plausible implication is straightforward porting to other computational science codes—such as LAMMPS, COMSOL, or finite-element analysis—with similar gains in automation, self-correction, and auditability.

FlamePilot’s approaches to stabilization (in experiment), automation (in AI), and consistency (in numerics) collectively underscore the convergence of physics-based architectures and AI-driven workflows in advanced combustion research (Han et al., 2021, Xiao et al., 4 Jan 2026, Ma et al., 2017).