Physics-Informed Machine Learning for Safe Learning-Based Control in Space Applications,

The growing interest in learning-based controllers presents new opportunities for the control of autonomous space systems, positioning them as viable alternatives to conventional optimal control techniques. Neural Networks (NNs), at the core of learning-based controllers, can approximate complex nonlinear functions, adapt to diverse operating conditions, and execute efficiently on low-power hardware, making them especially suited to the constraints of space systems.
Despite their potential, learning-based controllers still lack the formal guarantees of stability and interpretability that characterize classical control techniques, and their opaque internal representations limit analysis and validation. This raises concerns about safety, a critical requirement in space missions. In this context, emerging approaches that merge machine learning with physical modeling provide promising avenues toward reliable and certifiable neural controllers. These approaches offer ways to embed physics laws and control theory principles directly into the learning process, ensuring that trained controllers remain consistent with theoretical foundations. To this end, Physics-Informed Neural Networks (PINNs) allow the integration of physical and stability constraints into the training stage, guiding learning towards desired behaviors.
In parallel, novel NN architectures, such as Kolmogorov–Arnold Networks (KANs), enable the representation of complex nonlinear functions while preserving an interpretable analytical form, bridging the gap between black-box neural controllers and classical control laws. These components, combined with data-driven optimization, define a principled process for developing trustworthy and certifiable neural controllers. The proposed methodology will be developed and validated in realistic simulation environments, laying the foundation for the safe deployment of neural controllers in operational spacecraft.