Data-driven dynamic substructuring for spacecraft applications

Spacecraft are commonly treated as built-up structures which are assembled to obtain a system-level response. This process may need to be repeated on a case-by-case basis due to payload configuration differences between spacecraft sharing a common bus design or even variability among nominally identical units. This study proposes a data-driven dynamic substructuring (DDDS) framework that replaces selected physical substructures with machine-learning (ML) meta-models, referred to as ML substructures, to accelerate system-level analysis. Each ML substructure’s dynamics is condensed to its interface, after which a meta-model is trained to reproduce that interface behavior. Once trained, these ML substructures couple seamlessly with FE or experimentally derived substructures to form a hybrid model. All substructures are combined through a dual-assembly formulation that preserves the mathematical independence of each substructure within the hybrid model. The ML-based meta-model is specified by a feed-forward neural network (FNN) with supervised learning, which can predict the interface parameters based on various inputs, such as material, geometric and damping properties of the respective part. Consequently, the meta-model effectively captures the subsystem’s interface dynamics pertinent to a pre-selected parameter space, rather than a unique configuration. Clearly, the principle applies to the hybrid model as well. The proposed method is validated for a dummy satellite comprising a primary structure and interface ring, as well as three payload decks. One of the payload decks is represented by an ML meta model, while the remaining structure retains its conventional FE description. The network is trained on variations in panel add on mass, core and face thicknesses, and damping properties, and is then embedded within an FE–ML hybrid assembly. The results of the hybrid model confirm that DDDS delivers accurate system-level predictions without incurring the prohibitive costs of repeated full system analyses.