The research, led by Alejandro González and Jorge Macías of the University of Málaga together with Cléa Denamiel of the Ruđer Bošković Institute in Croatia, evaluates the performance of the Meteo-HySEA model against the state-of-the-art AdriSC-ADCIRC modeling suite. Meteo-HySEA, developed by the EDANYA research group in Málaga, belongs to the HySEA family of codes, which also includes the widely used Tsunami-HySEA and Landslide-HySEA models recognized as European flagship tools for simulating seismically and landslide-triggered tsunamis. What sets Meteo-HySEA apart is its combination of high-resolution finite volume numerical schemes with a native multi-GPU framework, allowing it to incorporate time-dependent atmospheric pressure fields as forcing while running orders of magnitude faster than conventional CPU-based models. Crucially, it can also simulate the onshore inundation that follows, extending the modeling chain from offshore wave generation all the way to flooded harbor quays.

The benchmark for the new model was the AdriSC system, a CPU-based coupled atmosphere-ocean modeling framework that has become the reference tool for meteotsunami research in the Adriatic. AdriSC couples the Weather Research and Forecasting model, downscaled to 1.5-kilometer resolution over the Adriatic, with the two-dimensional depth-integrated ADCIRC ocean model running on an unstructured mesh refined to spatial resolutions of up to 10 meters in vulnerable zones. This minute-scale atmospheric forcing is essential to capture the speed and amplitude of tsunamigenic pressure disturbances that excite Proudman resonance across the basin and the fundamental oscillation modes of the bays, which typically fall in the 10-to-40-minute period band. However, the computational cost of such CPU-based frameworks creates serious bottlenecks for ensemble forecasting, probabilistic hazard assessment, and real-time early warning, where reducing simulation times from hours to minutes is essential.

The team focused on three well-documented events that struck the meteotsunami-prone harbors of Vela Luka on Korčula Island and Stari Grad and Vrboska on Hvar Island. The first, on 25–26 June 2014, was triggered by a train of atmospheric gravity waves propagating from the Tyrrhenian Sea across the Adriatic, producing rapid pressure perturbations of up to 2.4 hectopascals in five minutes and maximum sea levels reaching 1.5 meters inside the harbors. The second, from late June to early July 2017, was associated with a synoptic cyclone over the central Mediterranean and upper-level jet stream winds exceeding 55 meters per second; oscillations lasted nearly 24 hours, with amplitudes up to 0.69 meters at tide gauges and exceeding one meter as captured in videos from Vrboska. The third, a remarkable multi-day sequence between 11 and 19 May 2020, produced repeated waves of 0.6 to 0.8 meters driven by recurring high-frequency pressure disturbances of 2 to 4 hectopascals, flooding harbors and leaving boats ashore on multiple days.

A distinctive feature of the study is its deliberately strict evaluation protocol. No correction or tuning of the atmospheric forcing was applied, ensuring that both modeling systems were tested under realistic operational conditions rather than with the benefit of hindsight. This choice matters because retrospective adjustments of pressure fields could improve agreement with observations, but such approaches rely on prior knowledge of the event and are therefore useless in a real-time forecasting context. The evaluation drew on high-frequency observations from the MESSI observational network, which includes microbarographs measuring atmospheric pressure at one-minute intervals with a precision of plus or minus 0.01 hectopascals and radar tide gauges recording sea level with millimeter accuracy, supplemented for the earlier events by five-minute pressure records from the Crometeo network of amateur weather stations across the Croatian coast.

The results reveal a fundamental constraint that applies to any meteotsunami modeling system: simulation accuracy is ultimately limited by the quality of the atmospheric forcing. Comparing weather model simulations driven by the ERA-Interim and ERA5 reanalyses, the researchers found that the ERA-Interim forcing systematically generated more intense and widespread pressure anomalies, sometimes exceeding observed amplitudes and producing spurious fluctuations, while ERA5-driven simulations yielded more localized and lower-amplitude disturbances that aligned better with observations in timing and structure but often missed the strongest pressure jumps. Most strikingly, the ERA5-driven simulation completely failed to reproduce the sharp 2.5-hectopascal pressure jump recorded on 11 May 2020, the most intense disturbance of the entire study period. Because observational coverage over the central Adriatic remains sparse, the authors note, fully assessing the skill of atmospheric models in this region is itself a difficult task.

Against this backdrop, Meteo-HySEA performed encouragingly. The model successfully reproduced the timing and spatial variability of sea-level oscillations across all three events and generally yielded higher amplitudes than AdriSC-ADCIRC under the same forcing. During the June 2014 event, for example, Meteo-HySEA produced peak amplitudes of roughly one meter in Vela Luka compared with 0.6 meters from AdriSC-ADCIRC, suggesting it may better capture the potential extremes of meteotsunami events. However, a systematic shortcoming also emerged: Meteo-HySEA consistently overestimated the dominant wave periods, particularly in semi-enclosed basins. In Stari Grad, its median simulated wave periods exceeded observed values by up to 50 percent. The authors attribute this to the wet-dry technique used to simulate inundation, which causes the effective geomorphology of harbors to evolve over time, and to limited coastal bathymetric coverage. Both models, it should be noted, also overestimated observed periods, indicating that harbor resonance characteristics remain difficult to represent accurately in general.

To probe the differences between the two systems more deeply, the team ran controlled numerical experiments with synthetic pressure disturbances based on the analytical formulation developed for the AdriSC meteotsunami surrogate model. These idealized disturbances, defined by parameters such as origin, propagation direction, amplitude, translation speed, period, and spatial width, were chosen to produce the strongest plausible meteotsunamis in each harbor. The comparison exposed systematic differences in how the two models trap and dissipate energy within semi-enclosed basins. In Vela Luka, Meteo-HySEA predicted extreme elevations exceeding five meters under the most severe synthetic forcing, compared with peaks closer to four meters for AdriSC-ADCIRC, along with a slower decay of oscillations. The analysis also uncovered a genuine physical phenomenon: a massive, localized meteotsunami setup of up to 1.8 meters at the tip of Vela Luka harbor, generated by resonant mass transport, on which the high-frequency seiches ride. Both components, the authors stress, must be retained together to assess total coastal flooding hazard.

The study also candidly documents technical limitations. Spurious hotspots appeared in maximum sea-level maps for Stari Grad and Vrboska, driven by sparse raw bathymetric data near coastlines and numerical artifacts of projecting complex coastal boundaries onto high-resolution structured grids, where abrupt bathymetric steps cause the flow to behave as if hitting a vertical wall. In Vrboska, the narrow channel characteristics could not be resolved even at seven-meter grid spacing, shifting the modeled coastline relative to satellite imagery. The authors emphasize that numerical interpolation cannot artificially reconstruct missing physical topography, making high-quality, high-resolution topobathymetric surveys a critical requirement for nearshore areas if the inundation capabilities of the new model are to be fully exploited.

Nevertheless, the overall verdict is that GPU-based solvers like Meteo-HySEA represent a promising pathway toward next-generation meteotsunami forecasting and hazard assessment. Because the model runs orders of magnitude faster than CPU-based alternatives, it opens the door to real-time operational use, ensemble forecasting, and explicit treatment of atmospheric and boundary-condition uncertainties, all of which are essential given that the lead time between disturbance detection and coastal impact in the Adriatic is typically less than three hours. The authors outline future work on three fronts: implementing wind stress, tidal, and general circulation forcing in the next release of the code; systematic validation with dense observational networks to determine whether the longer persistence of oscillations in Meteo-HySEA reflects more realistic harbor seiches or whether ADCIRC’s stronger damping better represents physical energy dissipation; and testing the model’s operational potential through integration with real-time atmospheric forecasts and early warning protocols in collaboration with civil protection agencies. With atmospheric forcing remaining the dominant source of forecast uncertainty, advances in convection-permitting ensembles and machine-learning-based nowcasting are expected to further improve fidelity, strengthening coastal resilience in the Adriatic and potentially at meteotsunami hotspots worldwide.

Subject of Research: Evaluation of the GPU-based Meteo-HySEA model for simulating atmospherically driven meteotsunami events and harbor flooding in the Adriatic Sea

Article Title: Assessing Meteo-HySEA performance for Adriatic meteotsunami events

Article References: Assessing Meteo-HySEA performance for Adriatic meteotsunami events. (n.d.). https://doi.org/10.1007/s11069-026-08351-y

Image Credits: AI Generated

DOI: 10.1007/s11069-026-08351-y

Keywords: meteotsunami, Adriatic Sea, Meteo-HySEA, GPU modeling, harbor resonance, Proudman resonance, coastal flooding, early warning systems, atmospheric pressure disturbances, AdriSC-ADCIRC, sea-level oscillations, numerical simulation