Neural network based heat load for the SPIRAL2 superconducting LINAC

SPIRAL2 is a superconducting linear accelerator (LINAC) operated at GANIL (Caen, France), consisting of 26 bulk niobium cavities cooled to 4 K. These cavities enable the acceleration of protons and heavy ions up to 33 MeV/u with beam currents reaching 5 mA. Continuous monitoring of the thermal load dissipated by the cavities during operation provides a sensitive indicator of their performance and stability. The datasets presented here capture these dynamic heat-load variations and serve as training material for the development of neural-network–based virtual observers. Several deep learning architectures—including dense neural networks (DNN), long short-term memory (LSTM) networks with and without attention mechanisms, convolutional neural networks (CNN), and hybrid stacked models—are explored to infer cavity behavior and detect anomalies in real time. These data and models contribute to advancing intelligent diagnostics and predictive maintenance for superconducting accelerator infrastructures.