>90%snapshots within ±15° of POD phase, intermediate J
5:1geared hot-wire rotation, decoupled from translation
1.0
Objective
The upstream shear layer of a jet in crossflow rolls up into vortices at more than 1 kHz. The lab's PIV system samples at 5 Hz, so each snapshot lands at an unknown phase of the cycle. I built a machine-learning pipeline that finds the shear-layer vortices in each snapshot, estimates the snapshot's phase from where they sit, and sorts the snapshots back into one reconstructed cycle. I also redesigned the hot-wire handle and the laser-sheet optics holder so alignment is repeatable.
532 nm light sheet in the jet's centre planeVortex cores in each frame (vorticity, red and blue)The newest frame, filed by its phase ΦJason's hardware: geared optics holder (1:4) and hot-wire handle (5:1)A 5 Hz strobe rebuilds a shear layer that rolls up at more than 1 kHz. The upstream shear layer of the jet sheds vortices more than 200 times between two PIV pulses, and the roll-up is no perfect clock, so each pulse of the 532 nm sheet (120 mJ, from a dual-cavity Nd:YAG rated for 15 Hz) freezes the layer at an effectively random phase. Jason's pipeline finds the vortices in each frame, weights each by the classifier's confidence w (0 to 1, 98.3% accurate) and takes the frame's phase as Φ = arg Σ w·exp(i2πs/λ), with s the distance along the layer and λ the vortex spacing. Filed by Φ into twelve 30° bins (±15°, the tolerance the poster scores against), the frames add up to one whole cycle, which the monitor plays back. Shown 5× slower than the tunnel: one pulse a second. At intermediate J the ML phase agreed with POD within ±15° for more than 90% of frames (99% near J = 7); at J = 1 and J ≥ 20 frames hold two or fewer clear vortices and agreement falls below about 60%. The jet path is the Pratte–Baines centreline; vortex spacing and confidence are an illustrative model of those findings.
3.0
Work performed
3.1Built the vortex-detection pipeline: flow-field preprocessing to reject noise, Q-criterion computation, candidate-blob extraction, and a neural-network classifier trained on labelled blobs from J = 1, 2, 5, 8, 11, 16, 20 and 30. Each candidate gets a confidence weight between 0 and 1.
3.2Estimated each snapshot's phase from the confidence-weighted positions of its vortices along the shear layer, then sorted all snapshots by phase to rebuild ~1 kHz shear-layer dynamics from 5 Hz data.
3.3Validated the ML phase against POD-derived phase snapshot by snapshot: more than 90% agreement within ±15° across the intermediate-J range, close to 99% near J ≈ 7. The method is now used in Ph.D. work to cross-check POD-based reduced-order models.
3.4Redesigned, printed, tested and installed the hot-wire handle (three iterations) and the laser-sheet optics holder (two iterations), after printing calibration coupons to measure the printer's real fits.
4.0
Results
Lab data, not illustrations: everything in this section is drawn from the run files my pipeline produced.
Fig 4-1Phase reconstruction · lab PIV data
J
0°90°180°270°360°
J keys · drag phase · ← →J 5 · Φ 000.0° · t 0.000 ms
Fig 4-1The real reconstruction from my MATLAB pipeline. At each J the same 200 PIV snapshots (S = 1.0, Uj = 6.7) are sorted by ML phase (left) and by POD phase (right), then averaged in 48 bins of 7.5°. When the two panels advance together, the ML phase agrees with POD. Time comes from each run's estimated roll-up period: about 0.85 kHz at J = 1, up to 1.66 kHz at J = 9.
Fig 4-2Stereo PIV · ω* and (u, w) · lab data
Loading lab data…
0°90°180°270°360°
ω*
−60+6
5 m/s◯ Real (w) · ⬚ Noise
Hover the field to probe it—
Fig 4-2The numbers behind the reconstruction, exported from the lab's stereo PIV (S = 1, Uj = 6.7 m/s, jet diameter 4.04 mm; the first 200 snapshots of each run). Colour is the fluctuating vorticity ω*: Gaussian-filtered ωy minus the run's mean field, normalised by Uj/Dj. Arrows are the raw in-plane velocity (u, w), every 7th vector (every 14th on narrow screens). Phase average bins the snapshots by ML phase into 48 bins of 7.5°; time uses the hot-wire fundamental. Raw snapshots are six single frames with every candidate the classifier scored: solid rings were classified Real (labelled with their weight w), dashed rings Noise.
Fig 4-3ML vs POD phase agreement · all J
100%75%50%25%0
90%
Momentum-flux ratio J
Hover a bar · click to load that J aboveJ 7 · 99.2% within ±15° · mean |Δφ| 4.6° · PLV 0.995
Fig 4-3For each J, the share of 120 paired snapshots whose ML phase lands within ±15° of the aligned POD phase (from pod_ml_all_j_summary). Agreement is highest at intermediate J, 99.2% at J = 7, and falls toward the ends of the range, where frames hold too few clear vortices to fix a phase. Solid bars reach 90%; hatched bars do not.
Table 5-1 · Measured design clearances (in.) for the HeyGear Reflect RS printer
Shape
Transition fit
Tight fit
Cylinder
0.003
0.000
Spur gear
0.006
0.002
Square
0.002
0.001
6.0
Plates
Pl. 1JICF wind tunnel and optical diagnostics: a dual-cavity Nd:YAG laser (Litron Nano L PIV, 15 Hz) feeds sheet-forming optics above the test section, imaged by two stereo PIV CCDs and a PLIF ICCD; red circles mark the laser-sheet optics holder and the hot-wire handle.
Pl. 2Hot-wire handle, old to new: the original clamp moved in x and rotated in the y–z plane through the same joint (photo and CAD, left); the new geared handle separates translation from rotation (photo and CAD, right).
Pl. 3The original hot-wire handle installed on the tunnel: a black clamp block on a pair of rods above a micrometer traverse stage.
Pl. 4The new geared hot-wire handle installed: a grey 3D-printed gear housing on the probe rod, held in a metal C-clamp.
Pl. 5Laser-sheet holder, old to new: the three-screw ring mount with two coupled degrees of freedom (left) replaced by a 3D-printed holder with a single geared rotation knob (right), shown installed and in CAD.
Pl. 6The original laser-sheet optic mount: a black ring positioned by three knurled screws, on a post below the sheet-forming optic.
Pl. 7The new 3D-printed laser-sheet holder installed on its post mount: a white printed ring with a single knurled knob for geared rotation.
Pl. 8Earlier draft concept (not built as shown): the geared laser-sheet holder driven by a stepper motor through a spur gear.
Pl. 9Step-1 flowchart in four panels: flow-field preprocessing (Q field at J = 5 to a thresholded planar-Q map), ML training on blob measurements, and vortex identification with a confidence weight on each candidate.
Pl. 10Candidate-vortex extraction at J = 5, frame 41: (a) transverse vorticity fluctuation, (b) planar Q above threshold, (c) neural-network-classified candidates, with accepted vortices on the shear-layer path.
Pl. 11Manual labelling mode used to build training data (J = 1, frame 1): smoothed vorticity with candidate blobs marked (left) beside the planar Q field (right).
Pl. 12Planar Q distribution at J = 5, 8, 10, 12, 20 and 30 (recent NEWLENS S = 1 campaign; black contours mark Q = 0): the vortex train steepens as J increases.
Pl. 13Validation confusion matrix for the vortex classifier: 4,594 noise and 1,646 real candidates classified correctly, 49 noise called real and 59 real called noise (98.3% accuracy).
Pl. 14Phase sorting: in acquisition order the estimated phases jump randomly around the cycle (left); sorted by ML phase the same snapshots form a smooth 0–360° sequence (right).
Pl. 15Phase-advancing vorticity at J = 5 (S = 1.0, same 200 snapshots, phase 0°): the ML-derived phase (left) and the POD-derived phase (right) give closely matching vortex trains along the shear layer.
Pl. 16Phase-averaged roll-up at J = 5, tracked: the classifier marks each shear-layer vortex on the dashed layer path with its confidence (AI 1.00 = certain vortex), and follows V1–V3 as they advance through the cycle.
Pl. 17ML–POD phase agreement within ±15° across momentum-flux ratio J: about 52% at J = 1, roughly 90% or higher from J = 3 to 11 with a peak near 99% at J = 7, then about 57% at J = 20 and 62% at J = 30; dashed line marks 90%.
7.0
Anomaly report
Closed
Observed
Hot-wire and laser-sheet adjustments coupled several degrees of freedom. Alignment was slow and hard to reproduce between runs.
Root cause
The hot-wire handle moved in x and rotated in the y–z plane through the same clamp. The laser-sheet holder was set by three coupled screws.
Corrective action
The new hot-wire handle separates translation from rotation and adds 5:1 geared angular control. The laser-sheet holder has a single geared rotation axis (1:4). Fits came from printed test coupons, not nominal tolerances.
Lesson
Measure the printer before designing for it. Cylinder fits needed 0.003 in of transition clearance; spur gears needed 0.006 in.
8.0
Limits & next
Operating range. Agreement drops at the ends of the tested range (52.5% at J = 1, 57.5% at J = 20, 62.5% at J = 30), where snapshots often hold two or fewer well-defined vortices. That sets the method's current operating range.
Next. Use the identified vortex cores to estimate the fundamental shear-layer frequency and compare it directly with hot-wire spectra.