Hypersonic Heat-Flux ML Diagnostics
Predicting IR heat flux from Schlieren video — DEVCOM ARL
- Mach 6.25
- Test regime
- 70+
- Labeled samples
- 8
- Test conditions
Problem
Hypersonic surface heating decides whether a vehicle survives its trajectory, but instrumenting a wind-tunnel model with dense heat-flux sensors is impractical. The question: can heat-flux fields be inferred from the flow imagery the tunnel already captures?
Motivation
At the UMD hypersonic wind tunnel, both Schlieren video and IR thermography are recorded every run. If a model could map the optical flow features to surface heat flux, each test would yield a full thermal field with no added instrumentation.
Constraints
- Mach 6.25 flow with limited, hard-won test time across 8 conditions
- Sparse labeled data — samples had to be generated from the runs available
- Heat flux recovered indirectly from transient IR temperature, not measured directly
Process
- 01
Pipeline
Built a 5-stage Python pipeline that extracts shock angle, intensity, and position from Mach 6.25 Schlieren video, automating bow-shock detection with OpenCV polynomial fitting.
- 02
Labeling
Generated 70+ labeled training samples across 8 test conditions, pairing extracted flow features with the corresponding thermal response.
- 03
Heat-flux recovery
Computed surface heat flux from FLIR IR thermography using the Cook-Felderman inverse method, porting the original MATLAB implementation to Python.
- 04
Model
Trained a multimodal model — a ResNet-18 image backbone, a 1-D CNN, and an SVD-based decoder — to predict full IR heat-flux fields from the Schlieren inputs.
Key decisions
Multimodal architecture over a single image model
Why · Pairing a ResNet-18 visual backbone with a 1-D CNN and an SVD decoder let the model exploit both the spatial shock structure and the lower-dimensional structure of the heat-flux fields.
Port Cook-Felderman from MATLAB to Python
Why · Bringing the inverse heat-flux method into the same environment as the pipeline made label generation reproducible and kept the whole workflow in one toolchain.
Results
- ✓End-to-end pipeline from raw Schlieren video to predicted heat-flux fields.
- ✓Automated bow-shock detection producing 70+ labeled samples across 8 conditions.
- ✓Reusable Python port of the Cook-Felderman inverse method.
Lessons learned
- When data is scarce and expensive, the feature-extraction pipeline matters more than the model.
- Indirect measurement — flow imagery to heat flux — can replace instrumentation you can't physically add.