Upper-level troughs and ridges are fundamental drivers of mid-latitude weather, organizing cyclogenesis, precipitation, and temperature extremes. Despite their importance, automatic detection remains difficult: existing algorithms rely on rigid rules that struggle to capture the high geometric variability of synoptic features. We introduce a physics-informed deep learning framework for automated axis detection that incorporates physical knowledge into the learning process, by transforming a curvature-based detection algorithm into continuous, differentiable operators embedded within an attention-based network, so geometric criteria are learned from data and axis detection draws on context from the surrounding flow. The network takes ECMWF Reanalysis v5 (ERA5) 500hPa geopotential height and horizontal wind fields as input, and is trained and evaluated against a new benchmark of expert-labelled Mediterranean trough and ridge scenes. The network output is a confidence map, converted into precise axis lines based on a cyclonic vorticity advection rule. The model significantly outperforms classical baselines (F1 rises from 0.64 to 0.84 for troughs and from 0.54 to 0.75 for ridges). The framework demonstrates generalization capabilities beyond its training data: it qualitatively transfers well to global mid-latitudes without retraining, and the same architecture, initially trained only on troughs, reaches competitive ridge-detection accuracy after fine-tuning on just ten additional labeled ridge scenes. Applying the detector to historical reanalysis, we construct the first expert-calibrated, deep-learning-based climatology of upper-level trough and ridge frequency for the Mediterranean basin, providing a powerful and easily portable tool for investigating upper-level circulation.</p> <p>Project page: <a href="https://sela-omer.github.io/upper-level-trough-ridge-detection">https://sela-omer.github.io/upper-level-trough-ridge-detection</a>