Contrail Prediction and Avoidance with ML

Register with the Copernicus Climate Data Store (CDS) and use the CDS API to download ERA5 pressure-level data for a region and time period of your choice (start with a single month over the North Atlantic or continental US). Request: temperature, relative humidity, specific humidity, u/v wind components, and geopotential at pressure levels from 150 to 350 hPa (the cruise altitude band, roughly 30,000–40,000 feet).

Also download surface-level data (solar radiation, outgoing longwave radiation) as these affect whether contrails have a warming or cooling effect. The data arrives in NetCDF format — xarray handles this natively with xr.open_dataset().

Implement the Schmidt-Appleman criterion (SAC) — the thermodynamic threshold for contrail formation. This requires calculating whether the mixing line between the engine exhaust plume and ambient air crosses the liquid saturation curve on a temperature-water content diagram. The key parameters are ambient temperature, ambient humidity, fuel-specific energy, and the overall propulsive efficiency of the engine.

Apply the SAC to every grid point in your ERA5 data to create a binary "contrail possible" field. Then compute relative humidity with respect to ice (RHi) from ERA5's temperature and specific humidity to determine where contrails would persist (RHi > 100%). This physics-based prediction is your baseline.

For supervised learning, you need ground truth. Use Google's Open Contrails dataset (released on Kaggle), which contains GOES-16 satellite images labeled with human-annotated contrail masks. Alternatively, use the CoCiP (Contrail Cirrus Prediction) model outputs available from Breakthrough Energy as a softer label source.

Align the contrail observations with your ERA5 data spatially and temporally. This registration step is tricky — satellite pixels don't align perfectly with reanalysis grid cells, so you'll need to interpolate ERA5 data to satellite observation locations and times using xarray's .interp() method.

Build a U-Net or lightweight CNN in PyTorch that takes multi-channel atmospheric fields (temperature, RHi, wind shear — stacked as channels at each pressure level) and predicts a contrail probability map. The architecture should capture spatial context because contrail-favorable regions are spatially coherent — a 5x5 patch of humid cold air is more likely to produce contrails than an isolated grid cell.

Use binary cross-entropy loss for training since this is fundamentally a pixel-wise classification problem. Apply class weighting or focal loss to handle the severe class imbalance — contrails cover only a small fraction of the sky at any given time. Train with the Adam optimizer, starting with a learning rate of 1e-4, and monitor validation Brier score.

Compare your ML model against the SAC baseline using precision, recall, Brier score, and reliability diagrams. The SAC baseline will have high recall (it identifies most contrail-possible regions) but low precision (many SAC-positive regions don't actually produce visible contrails because of sub-grid variability). Your ML model should improve precision by learning patterns the threshold criterion misses.

Create spatial maps showing where the two approaches disagree. Does the ML model learn that contrails are less likely near the boundaries of ISSRs? Does it capture the diurnal cycle (nighttime contrails have different climate impact than daytime ones)?

Using your contrail probability maps, simulate a simple avoidance strategy: for a set of flight trajectories (from OpenSky Network ADS-B data), check whether each flight segment passes through a high-contrail-probability region. If so, compute the altitude change needed (typically a shift of 2,000–4,000 feet) and estimate the extra fuel burn from flying at a non-optimal altitude.

Calculate the trade-off: the additional CO2 from extra fuel vs. the warming avoided by preventing contrails. Research suggests the climate benefit of contrail avoidance outweighs the fuel penalty by a factor of 10–100x, but this depends heavily on the accuracy of the prediction model — false positives mean wasted fuel with no climate benefit.

Quantify the uncertainty in your predictions using Monte Carlo dropout or an ensemble of models trained on different data splits. Map the uncertainty — is it highest at the edges of ISSRs (as expected) or in regions where ERA5 resolution is insufficient? Discuss the operational implications: an airline would only reroute if the contrail probability exceeds some threshold — what threshold balances climate benefit against fuel cost?

Write a research-style report with an abstract, methodology, results, and discussion. Include a section on limitations: ERA5's humidity biases in the upper troposphere, the satellite detection limitations (thin contrails are invisible to GOES), and the simplifying assumptions in your routing simulation.