At Dassault Aviation, I notably worked on an experimental project ✈️, aimed at showcasing the capabilities of artificial intelligence 🤖 in the aerospace sector, and more specifically in flight testing.

The goal was to predict the in-flight temperature 🌡️ of the temperature and pressure calibration probe located at the nose of test aircraft, with an average error of 0.5 degrees Celsius 📊. In the future, these models could be used to generate simulated test data or replace physical probes.

The reference model that allowed me to obtain quick results was an XGBoostRegressor ⚙️.

I then used an LSTM (a recurrent neural network, ideal for time series data, like in this case) 🧠, which unsurprisingly achieved the best results. Its ability to retain information for future predictions allowed it to reduce the error to an average of 0.16 degrees Celsius, even during chaotic flights 🌀 (loss of over 3000 meters in one minute!).