The detection of MeV-GeV neutrinos from astronomical sources is a long-lasting challenge for
neutrino experiments. The low flux predicted for transient sources, such as solar flares, and their
low-energy signature, requires a detector with both a large instrumented volume as well as a high
density of photomultiplier tubes (PMTs). We discuss how KM3NeT can play a key role in the
search for these neutrinos. KM3NeT is a Cherenkov neutrino telescope currently under deployment, located at the bottom of the Mediterranean Sea. It consists of two arrays of Digital Optical
Modules (DOMs): KM3NeT/ORCA and KM3NeT/ARCA, which are optimised for the detection
of GeV neutrinos for oscillation studies, and higher-energy astronomical neutrinos respectively.
We exploit the multi-PMT configuration of KM3NeT’s DOMs to develop the techniques that allow
the disentangling of the MeV-GeV neutrino signature from the atmospheric and environmental
background. Comparing data with neutrino simulations we identify the variables with discriminating power, and by applying hard cuts we are able to reject a large fraction of background. We
present a graph neural network approach to classify signal from background. To further improve
the sensitivities compared to previous studies, we will make use of the Hierarchical Graph Pooling
with Structure Learning algorithm and will use graph-structured data to reproduce the hit geometry on the DOM. This will allow for stronger constraints on the hits and reduce the fraction of
background that survives the selection.