A Graph-based approach towards a Hybrid Digital Twin for Satellite Thermal Management
Please login to view abstract download link
In space technology, predictive maintenance of satellites plays a pivotal role in ensuring mission efficacy and maximizing operational longevity, particularly concerning thermal management systems. This presentation introduces an innovative scientific machine learning (SciML) framework to digital twinning that combines physics-based and data-driven methodologies to develop a model of a satellite's thermal subsystem. The main challenge addressed in this research is the limited availability of detailed geometric data and material properties, coupled with the inherent structural complexity of satellites. These constraints preclude the use of high-fidelity numerical methods typically employed during satellite design phases. Consequently, for real-time thermal operation management, fault detection, and predictive maintenance, data-driven models become indispensable. In this work, we approximate the satellite's thermal behaviour using a Lumped Parameter Thermal Model (LPTM), which represents thermal dynamics through a graph-based structure. This approach leverages the sparsity of heat transmission paths connecting satellite subsystems for which thermal telemetry's data are available. The LPTM parameters, corresponding to both nodal and connective thermal properties, are learned from data collected from onboard sparse thermal sensors. To enhance the model's fidelity, we integrate physical constraints such as incoming heat flux from solar radiation and heat dissipation by the satellite's electrical systems and active heat lines. This integration accounts for both mission-specific and platform-dependent inputs and operational conditions. Our hybrid approach, combining data-driven learning with physics-based modelling, offers a computationally efficient solution for real-time thermal prediction and monitoring. The resulting digital twinning technology supports predictive maintenance without necessitating high-resolution structural information, making it a practical tool for operational decision-making processes aimed at extending satellite lifespans. This work has been financially supported by ICSC—Centro Nazionale di Ricerca in High Performance Computing, Big Data, and Quantum Computing funded by European Union—NextGenerationEU through an Innovation Grant.
