Keras Reconstructs Cosmic-Ray Showers End-to-End

The breakthrough is not directly identifying a black hole or supernova; it is turning an existing detector into a much more capable scientific instrument.

• –Shared LSTMs learn waveform features while hexagonal convolutions encode the array’s geometry and rotational symmetry.
• –Jointly predicting energy, mass, shower maximum, core, and direction improves generalization over isolated task models.
• –The method preserves information discarded by traditional charge-and-arrival-time features.
• –Results indicate increasingly heavier cosmic rays at the highest energies, with composition changes aligned to major spectrum features.
• –Simulation-to-real-world mismatch and uncertainty calibration remain critical before these models can support definitive source attribution.