Presented at the GHRSST XXIII international science team meeting, 27 June-1 July 2022, online and in-person (Barcelona). #GHRSST23 Short abstract In the framework of the Copernicus Marine Service, the Italian National Research Council (CNR) – Institute of Marine Sciences (ISMAR) is responsible for producing and distributing operational near-real-time (NRT) Sea Surface Temperature (SST) products over the Mediterranean and Black Seas. The CNR-ISMAR SST processing chain, which includes several modules, from data extraction and preliminary quality control to cloudy pixel removal and satellite images merging, provides daily (night-time) merged multi-sensor (L3S) and optimally interpolated (L4) foundation SST fields at high (HR) and ultra-high (UHR) spatial resolution (i.e., over 1/16° and 1/100° regular latitude-longitude grids, respectively). However, the effective resolution of UHR L4 products strictly depends on the availability of high-resolution cloud-free measurements. The Optimal Interpolation algorithm makes use of HR L4 data remapped onto a 1/100° regular grid as first-guess (which means UHR SST features are already filtered out) and it is not able to reconstruct small scale features unless valid L3 observations are present within a short temporal window. For this reason, CNR is presently working to improve the MED NRT SST products’ effective resolution and SST gradients’ accuracy through the development of deep learning models. In particular, the application of Convolutional Neural Networks (CNN) in the process of reconstructing high-resolution images from low-resolution ones, the so-called single image Super Resolution (SR) technique, has demonstrated an impressive potential. Here we present preliminary results on the achievements and the limitations in applying this specific class of artificial intelligence techniques to improve the effective resolution (and SST gradients) of our MED-NRT-L4-UHR product.