SpeckleNet: a large-scale PSF subtraction deep learning model for exoplanetdetection and characterization in high contrast imaging

We introduce SpeckleNet, a next-generation deep learning framework for direct exoplanet detection and characterization in high-contrast imaging (HCI), developed in the context of the VLT/SPHERE instrument and designed to adapt to its forthcoming upgrade, SPHERE+. SpeckleNet integrates Variational Autoencoders (VAEs), Latent Diffusion Models (LDMs), and conditional learning to model and reconstruct the stellar Point Spread Function (PSF) in a powerful generative framework. Unlike previous approaches such as ConStruct, SpeckleNet explicitly incorporates the temporal correlations of speckle noise during training, enabling a more accurate and dynamic separation of stellar noise from planetary signals. A key limitation of traditional algorithms like Principal Component Analysis (PCA) is their tendency to introduce planet self-subtraction, significantly reducing sensitivity to faint sources. SpeckleNet is specifically designed to overcome this issue by learning a more faithful, non-destructive reconstruction of the PSF that preserves potential exoplanet signals. Trained on a uniquely extensive and diverse SPHERE dataset, SpeckleNet sets a new benchmark for robustness and sensitivity in variable observing conditions. Through transfer learning, SpeckleNet can leverage its learned knowledge from SPHERE data and adapt to the improved capabilities of SPHERE+, ensuring performance as AO instrumentation evolves. Together, these innovations establish SpeckleNet as a powerful new tool for the search for exoplanets.