In satellite space missions, the ability to extract information from sensor data is crucial to monitor the operational status of the satellite, to predict future trends and anticipate potential failures. State-of-the-art machine learning approaches offer nowadays computationally efficient models to address these problems but typically lack robustness and reliability in prediction in the case of limited training data.
In the context of a joint project between the MOX Laboratory of the Math Department of Politecnico di Milano and the company Thales Alenia Space Italia, we are facing the open challenges for the realization of digital twins of space mission systems. The project, supported by an Innovation Grant of the National Center for HPC, Big Data and Quantum Computing, in the framework of the activity of the Spoke 6 “Multiscale Modelling & Engineering Applications”, aims at developing innovative scientific machine learning (SciML) approaches capable of leveraging physical knowledge of the phenomena to cope with the limited information provided by sensors. The final goal is to create interpretable and robust digital twins of satellite’s sub-systems, while maintaining the efficiency of ML models.
One relevant example that illustrates the potential and challenges of digital twinning in space missions is the thermal subsystem. Thermal management is mission-critical in satellites due to dynamic and extreme conditions from orbital motion, solar exposure, and limited heat dissipation in vacuum. Existing control strategies rely on fixed thresholds and reactive mechanisms, resulting in excessive power use, hardware wear, and limited adaptability. The main challenge is not just temperature prediction, but real-time, anticipatory control to prevent thermal excursions and ensure mission integrity.