flamelab

Agenda

flamelab is a measurement instrument, not a video editor: it turns 4K pool-fire footage into per-frame flame geometry (length, width, tip position) in millimetres with propagated uncertainty, classifies the flame’s dynamical regime through its Hopf-bifurcation sequence, and aggregates repeat experiments into ensemble statistics with defensible error bars. Every number carries units, uncertainty and provenance — and where the footage cannot support a trustworthy number (e.g. a frequency near the camera’s Nyquist limit), the pipeline refuses and says so rather than guessing.

Problem

Combustion dynamics are usually measured with a ruler and a stopwatch — or with diagnostics (laser sheet imaging, chemiluminescence filters, intensified cameras) that a small lab cannot run on every experiment. A consumer 4K camera is cheap, but its footage is hard to quantify honestly: veiling glare inflates the apparent flame boundary, 4:2:0 chroma subsampling and unknown ISP processing corrupt colour-based diagnostics, and the 23.976 fps frame rate aliases the flicker frequencies that matter most. flamelab is built around those exact limits: correct the glare, propagate the segmentation uncertainty, and never report an aliased frequency as if it were physical.

Approach

• Sub-pixel, glare-corrected contour — veiling-glare correction followed by a multi-gate flame-component selector (brightness, area, rim-anchoring, axis overlap), anchored to the burner rim measured once per video. On the full 13,127-frame reference recording, dropout frames went from 5.1% to 0.0%, and the detected rim agrees with an independent measurement to 0.45 px.
• Physical quantities with uncertainty — length, width, area, rugosity, tip angle and pinch-off events in mm (ruler-photo calibrated, honesty-flagged as not caliper-confirmed), each with systematic (segmentation) and statistical uncertainty separated.
• Regime classification — a hidden-Markov classifier over sliding-window spectral features labels every frame with the literature Hopf sequence: Ignition Transient → Steady Laminar → Flickering Onset → Periodic Flickering.
• Ensemble statistics across repeat runs — eight repeat n-heptane pool-flame recordings (the same experiment repeated), aggregated with bootstrap confidence intervals, ICC variance decomposition, ANOVA/Levene tests, GUM-combined uncertainty, and an ignition-aligned frame matrix.
• Thirteen analysis methods wired to figures and CSV data — continuous wavelet transform, skeleton/pinch-off precursors, phase-averaged puff-cycle reconstruction, onion-peeling Abel inversion, correlation dimension and largest Lyapunov exponent (each with an explicit reliability flag), Hilbert–Huang/EMD, TV-L1 optical flow cross-checked against kymograph convection velocity, active contours, windowed full-frame POD/DMD, growth-rate fitting at onset, change-point onset detection, and the literature-standard 50%-intermittency height.
• Zonal seismograph instrument — the same real-time tracked boundary and numeric HUD (height, area, volume, heat-release-rate proxy, tip position), but with the flame split into height zones (McCaffrey’s continuous/intermittent/plume regions among the schemes tested) and a scrolling per-zone width trace, so where a fluctuation originates is visible, not just that it happened.
• Evaporation/burning-rate physics — a Zukoski-scaled burning-rate proxy fit per run (rise time to 90% of steady rate, growth-scale, overshoot), pooled ignition-aligned across the ensemble.

Implementation

Python with OpenCV, NumPy/SciPy and scikit-image; a ~300-test suite; and three reproducible drivers — one per run (run-analysis.py), one for the cross-run ensemble (run-ensemble.py), and one for the methods (run-methods.py). Outputs are data-first: per-frame HDF5/CSV tables, a CSV next to every figure containing the exact plotted columns, and standalone plotting scripts so any figure can be edited and re-plotted without the pipeline.

Results

The campaign is eight repeat n-heptane pool-flame runs from one recording session. Detection is successful on 96.8–99.8% of frames per run. The headline measurements, with their honesty caveats on record:

• Ensemble flame height — 110.1 mm (bootstrap 95% CI [75.0, 124.5] across the 8-run campaign; GUM-combined uncertainty ±12.7 mm). Restricting to the three runs with a validated burner-rim anchor gives 117.2 mm [107.9, 123.7] — the two reference-frame definitions are kept separate on purpose.
• Ensemble width — 16.3 mm [14.6, 18.2] across all runs; 13.7 mm [12.3, 14.4] for the anchored subset (width at one burner diameter above the rim — a different physical quantity than the unanchored fallback, documented, not pooled silently).
• Run-to-run structure — variance decomposition: ICC 0.42 (height) / 0.52 (width) — real between-run differences exist on top of within-run flicker, and ANOVA confirms the run means differ.
• Puffing frequency — 0.40 Hz [0.29, 0.52], consistent across all runs, ~35× below the classical pool-fire correlation’s ~13.9 Hz prediction. Two honest candidate explanations are on record: a genuinely different flow regime at this small scale, or aliasing of a true 10–15 Hz oscillation against the 11.99 Hz Nyquist limit. Resolving this needs higher-frame-rate re-acquisition — flagged, not spun.
• Regime sequence — every run follows the same Ignition → Steady → Flickering Onset → Periodic sequence, with the onset time located to a bootstrap confidence interval per run, and positive oscillation growth rates (0.004–0.047 s⁻¹) across all runs at onset — the supercritical-Hopf signature.

The gallery above shows the ignition-aligned frame matrix (same physical field of view in every cell), the seismograph-style tip and width fluctuation charts, the ensemble traces with uncertainty bands, kymographs, POD/DMD modes, the per-regime overlay videos with the detected contour burned onto the real footage, and the zonal seismograph videos described below. The ensemble power-spectral-density figure is the direct spectral evidence behind the puffing-frequency measurement above, and the per-run flickering-onset-time and anchored-subset-robustness figures make the run-to-run repeatability question directly inspectable rather than folded into one summary statistic. A sample of the live measurement:

Zonal instrument and boundary-method validation

The seismograph-style tip/width tracking above is extended into a zonal view: the flame is divided into height bins under several zoning schemes (an 8×15 mm even split, 10 zones per burner diameter, a coarse 4×30 mm split, and McCaffrey’s physically-motivated continuous/intermittent/plume regions), and each zone’s own width fluctuation is tracked and traced live alongside the footage, the numeric HUD, and a shaded systematic-uncertainty band. Pooled across the whole 8-run campaign, RMS width fluctuation climbs steadily with height under every zoning scheme, and the McCaffrey plume-region zone is consistently both the noisiest and the most run-to-run variable — the two facts a single global width-fluctuation number would hide. The same instrument applied independently to three different curated runs (MVI_8036, MVI_8037, MVI_8041) is a direct repeatability check on the technique, not just three more data points.

A dedicated boundary-method validation pass on real footage (2026-09-06) followed a targeted internal audit that found four defects in the method-ranking pipeline itself — a synthetic test flame that was inadvertently too thin, a NaN background region silently disabling every validity gate, an error-metric bug that dropped missed detections instead of penalizing them, and a truth harness that reseeded its own randomness every frame. Every boundary-method ranking on record before that date is superseded by the corrected scoreboard; the gallery’s “segmentation methods head-to-head” figures are the current, post-correction evidence, shown directly on real onset-window and developed-puffing footage rather than only on synthetic test cases.

Per-run quality-control contact sheets back the campaign’s scale-consistency check: independently fitted burner-rim scale anchors disagree noticeably run to run (one run’s fit was flagged as likely locking onto the wrong reference feature), while the flame’s own detected length in pixels agrees much more tightly across runs — the QC sheets are the visual evidence behind trusting the length-based cross-run comparison over the individually-fitted scale anchors.

Validation

Every method is validated against synthetic truth before touching real footage: the forward-model oracle (rolling shutter, exposure, noise, H.264 round-trip), known-answer nonlinear diagnostics (Lyapunov on the logistic map, Abel round-trip on Gaussian profiles, aliasing fold-back algebra), and real-data cross-checks that don’t share an assumption (50%-intermittency height vs. time-averaged length; TV-L1 flow vs. kymograph convection velocity; rim anchor vs. modal observed base). The full suite passes — currently 309 tests (up from 299 as the zonal instrument and evaporation-physics modules picked up their own coverage).

My contribution

The instrument is mine: the measurement pipeline, the uncertainty and refusal policy, the regime classifier, the ensemble statistics and every figure in the gallery. The experimental rig (tube burner, pool-fire configuration) was designed and commissioned jointly with Gagan Verma.

Calibration, video stitching and an interactive course

A take recorded in several files is declared in the run registry as a list of video parts, and flamelab reads the parts as one continuous frame stream: it checks that the parts match in frame size, frame rate and rotation, refuses a recording with a time gap unless that is declared, counts the frames it can actually decode rather than trusting the container, and records the part boundaries in its output. On a real clip split in two, every stitched frame was verified to equal an original frame. A separate calibration clip (a ruler beside the unlit burner) measures the pixels-per-millimetre scale from the periodic ticks of the ruler, and shows that a scale cannot be carried from one clip to another when the camera has moved.

The ideas behind the measurements (aliasing against the camera’s frame rate, veiling glare and thresholding, sub-pixel edges, the kymograph, flame height, the onset of puffing, onion-peeling inversion, onset detection and stitching) are also available as nine interactive models in the Math Visualization Studio, opened by the Explore in Studio action of this page, and as an eight-session written course with figures drawn from these runs.

Where this is going

• High-frame-rate re-acquisition (≥240 fps) — the single highest-value next step: it removes the Nyquist ambiguity outright, and the pipeline accepts it unchanged.
• Multi-fuel campaign — the same instrument applied to Methanol-to-Gasoline (MtG) surrogate and ethanol blends, where the fuel-dependent sooting behaviour is itself a measurable.
• Bifurcation analysis — the growth-rate fit at onset is the seed of a supercritical/subcritical Hopf classification, the strongest publishable claim this footage can support.
• Physics-informed boundary dynamics — a PINN-consistent surrogate for the boundary dynamics, once the calibrated multi-fuel dataset exists.