Known virtual leaks exposed errors in source estimates.
A gas sensor detects a plume, but not its origin. Virtual leaks let us test whether a model can work backward from sparse clues.
Simulate ammonia leaks around a container yard. Use a handful of noisy gas sensors and wind readings to infer where a leak started, its rate and timing, and show when the model is uncertain or wrong.
Edited demonstration briefThe result
The featured replay reveals a mistaken leak location.
The model estimates a leak source from sensor histories and wind readings. The known simulated source reveals when that estimate is wrong, as in the featured replay.
Rural dispersion was benchmarked against field data; the urban yard and source-finding system were not validated in the field.
Recorded simulation
The known leak location reveals where the estimate went wrong.
Reveal the plume and its true source behind the sensor readings.
Recorded outputs, replayed in your browser. No new simulation is running.
Wind and buildings separate a gas reading from its source.
The world we created
A dispersion model moves gas through a virtual container yard under changing winds. It records the source and plume as well as sensor readings with delay, noise, calibration errors and saturation.
What this enables
A learned model estimates the source location, release rate and timing from sensor histories and wind readings. The known simulated source lets us check both the answer and its uncertainty.
Simulated leaks paired sensor histories with known sources.
Vary leaks and winds in a virtual container yard. Pair the full plume with imperfect sensor readings.
A learned model estimated the source from sparse readings.
Train a model to estimate the source from sensor histories and changing wind.
Unseen leaks exposed missed locations and overconfidence.
Use unseen leaks to expose missed locations and overconfidence before any field deployment.
Known simulated sources make an otherwise hidden answer available for training, and make errors and overconfidence visible in testing.
Original research artifacts
The original files retain the recorded measurements and model details.