Internship : Developing A Neural Network For Exospheric Temperature Forecasting

For satellites in Low Earth Orbit, the atmospheric drag is the main disruptive force, opposed to movement. The lower the satellite, the higher the air density, the stronger the atmospheric drag. As the atmospheric drag decreases with altitude over time, this implies an altitude loss and therefore orbit perturbations.

The air density is a function of altitude and a semi-empirical model is developed at CNES. It relies on the assumption that the main constituents of the thermosphere exist in independent diffuse equilibrium. Based on this assumption, the model reconstructs the total density by combining the partial densities of each constituent, first determined at a reference altitude of 120 km and then extrapolated to higher altitudes (up to about 1500 km) using height functions dependent on the thermal structure.

The state of the thermosphere is strongly influenced by several environmental drivers, including geographic location, local solar time and season, solar activity, and geomagnetic activity. To represent the effect of these drivers, an empirical function is used, expressed as a spherical harmonic expansion which includes:

• Non-periodic terms, which capture, for instance, latitudinal dependence (through Legendre polynomials), solar variability, and geomagnetic disturbances
• Periodic terms, which describe seasonal variations (annual and semi-annual) as well as diurnal, semi-diurnal, and ter-diurnal oscillations,…

The model is adjusted according to empirical observations in order to best match said observations but with a low spatial and temporal resolution, due to the small number of objects that can be used to create a reliable database. There are few satellites allowing for sufficient precision at the disposal of the CNES and only them can be used to adjust the model thanks to empirical observations. This means that the empirical observation responsible to model the entire atmosphere is only relying on six satellites, which explains the low spatial and temporal resolution. As a result, these models are only updated once or twice per decade. That’s why a new model based on neural network has been created in a previous internship and substantially improves the prediction accuracy of the DTM model for 24-hours predictions. For this project, many architectures were used to produce temperature predictions. The final choice was an autoencoder GRU (also known as seq2seq GRU) and trained on a validation period sampled through periodic intervals of the whole dataset. All the code has been developed in python, using Pytorch and panda. The purpose of this following internship is to continue this work by introducing new datasets now that the architecture and the choice of the NN has been made.

This internship could start in March 2026.

Curiosité, autonomie et rigueur. Connaissances en IA et en python.