Toward operational meteotsunami warning on the Portuguese coast

Meteotsunamis are tsunami-like sea-level oscillations initiated by atmospheric disturbances and amplified by resonance. Using a Portugal meteotsunami catalogue for 2010-2020 (39 events: 14 good-weather and 25 bad-weather), we propose an operational decision workflow that couples physically interpretable diagnostics with an atmospheric machine-learning (ML) trigger. We first test whether a compact physical formulation can explain sea-level variation. For “good-weather” cases, a regression model combining a direct pressure-response term with a resonance term improves (R² ≈ 0.40) and indicates peak amplification near a pressure-jump speed of U ≈ 20 m/s. Applying the same model to the full catalogue fails, suggesting that "bad-weather” cases may involve additional forcing and/or more complex atmospheric structure.

We then develop a meteotsunami detector using atmospheric pressure observation: pressure-jump candidates (ΔP ≥ 1.0 hPa) are consolidated, and converted into fixed 12-h multi-channel windows for a Temporal Convolutional Network (TCN) for each meteorological observatory. On an independent 2020 test set, the coastwide ensemble achieves event-level recall = 1.0 at τ = 0.30 (precision = 0.50; F1 = 0.67), but with substantial false alarms. To mitigate these limitations, we propose a two-stage warning strategy: an ML-driven atmospheric watch/advisory followed by tide-gauge (and future Distributed Acoustic Sensing, DAS) screening that first flags sea-level anomalies and then confirms meteotsunami-consistent signatures. This structure is designed to reduce false alarms while capturing events that may be weakly observed by the meteorological network.