Your application must include a recommendation letter from your Ph.D. supervisor, a detailed CV including university education and work experience, a list of publications, a 2-page description of the work undertaken during the course of your PhD.

For more Information, contact : Directeur de Recherche  - sean.bruinsma@cnes.fr

Submit the complete application online (Apply) before March 13th, 2026 Midnight Paris time

===================

The construction of very large constellations of commercial LEO satellites began in about 2018 when the private company SpaceX launched its first Starlink satellite prototypes, leading to an increasingly crowded environment in the low Earth orbit (LEO) region. In LEO, aerodynamic drag is by far the largest error source associated with orbit propagation with a defining role in orbit and reentry prediction, and collision avoidance – which is becoming a challenging activity due to the sheer number of objects in LEO. The main uncertainty in aerodynamic drag calculation and forecasting of objects in LEO is due to the highly variable, in both space and time, neutral upper atmosphere. Data assimilation (DA) models fuse background models with near-real-time (nrt) observations. They have demonstrated superior performance when compared with their non-DA equivalents. DTM_nrt developed by CNES, which assimilates global mean exospheric temperature, is such a model based on AI . The model was modernized in 2025, and presently an autoencoder GRU model is used for the predictions. However, listed below in order of importance are identified shortcomings and limitations in the current version of DTM_nrt: 

1- Which altitude is optimal when assimilating exospheric temperature, and which altitudes cannot be used (e.g. too low and therefore exospheric temperature not reached yet)? This study establishes over which altitude range DTM_nrt is operable.

2- What is the uncertainty of the neutral density predictions for the nowcast and each of the forecasting horizons (i.e., 24, 48 and 72 hr)? This requires quantifying the uncertainties due to the quality of the assimilated data (e.g., less accurate during solar minimum, or at higher altitudes), of the geomagnetic activity forecast when provided and the level of activity, and of the ML model itself (e.g., with or without rebalancing of the training data). The uncertainty quantification, in absence of for example the uncertainty in the geomagnetic activity forecasts, may use an ensemble model approach. 

3- How can exospheric temperatures inferred from satellites in inclined (i.e., not near-polar) orbits be used? Since many objects are in non-polar or even low-latitude orbits, an algorithm must be developed that compensates for biases in the temperature corrections due to the non-global coverage, thereby making their exploitation also possible.

4- How can the spatial resolution of DTM_nrt be increased from a global bias correction to a model corrected also for latitudinal and diurnal variations, or for a variable temperature height profile by means of temperature gradients? This study concerns both evaluating and selecting pertinent data (satellites) as well as accommodating the new corrections in the model’s DA algorithm.

5- Is it necessary to update or re-train the model every year, or after specific phases in the solar cycle, in order to maintain its predictive performance? 

6- To what degree is the performance of the model dependent on the background model? 

The objective of the above studies is the development of an advanced prototype of a comprehensive DTM_nrt model that enables accurate thermosphere density predictions up to 72hr out, which is essential for future VLEO operations. The studies will be performed through a mix of simulations, and processing of ephemeris, GNSS and SLR data.