The wave forecasting system is based on a regional configuration of WAVEWATCH III™ v5.16 WW3, implemented on two domains. The first domain covers the Estuary and Gulf of St. Lawrence (EGSL) between 45.5 and 51.9° N latitude and 70 and 55° W longitude. The computational grid consists in a uniform rectilinear grid of 0.015 × 0.01° spatial resolution (∼ 1 km). The other domain covering the North Atlantic (NA) between 45.5 and 51.9° N latitude and 70 and 55° W longitude with 0.5° resolution is used to compute wave boundary conditions for the WW3-EGSL domain at Cabot Strait and Belle-Isle Strait (Fig. ). The model bathymetry is obtained from the General Bathymetric Chart of the Oceans dataset (GEBCO v2019, ), developed through the Nippon Foundation-GEBCO Seabed 2030 Project which provides bathymetric data on a global coverage on a 15 arcsec grid. The wave spectrum is discretized into 25 frequency bins ranging from 0.05 to 0.5 Hz and 36 directions (10° resolution). Model parameters and source term specifications are listed in Table . One forecast cycle is initialized daily at 00:00 UTC and provides 48 h lead time predictions.

The Gulf of St. Lawrence is a seasonally ice-covered semi-enclosed sea connecting the St. Lawrence river watershed to the North Atlantic. The duration, maximal coverage and estimated volume of the seasonal ice cover exhibit a strong interannual variability, and are very strongly controlled by winter air temperatures . The maximum coverage vary from roughly one quarter of the entire area shown in Fig.  at the end of the mildest winters to almost a full coverage a for the severest winters. This means that except in rare occasions that are becoming rarer as climate warms, sea ice in the EGSL coexists with wind waves generated in ice-free waters. The WW3 model configuration used here considers that wave energy E attenuates exponentially in space such that E(x,σ) = E0(σ)exp⁡(-α(σ)x) where α is an empirically determined attenuation coefficient that is a function of the wave angular frequency σ. Here this coefficient has been tuned to fit observations carried out during the BicWin campaigns in the EGSL , such that 2α(σ)=5.8×10-3σ3.2. However, as sea ice is a prescribed field, the system does not represent coupled interactions that would feedback on sea ice distribution. For instance, the wave radiative push exerted by waves on sea ice when they attenuate, and modifications to the floe size and ice internal strength caused by wave-induced break-up seefor a review are not accounted for.

In this section we use data collected between 31 March and 15 January to evaluate forecasts skills during the summer or ice-free period (see Figs.  and ). Results of the statistical analysis are shown in Table .

The analysis of significant wave height forecasts for +24 to +48 h lead times indicates that the system generally performs well throughout the EGSL. Across all wave conditions, the mean absolute error (MAE) ranges from 0.13 to 0.22 m, and the root mean square error (RMSE) varies from 0.18 to 0.30 m. This corresponds to a 21 % error relative to the average wave height observed at station IML-10 and 43 % at station GAL. Results show minimal bias, with the mean bias error (MBE) ranging from -0.03 to 0.07 m, and regression slopes a1 between 0.73 and 0.93, suggesting that most errors are random rather than systematic.

When focusing on wave height values larger than the 90th percentile, the error represents only 15 % of the average wave height at IML-10 and 30 % at GAL. However, the bias increases slightly, with MBE values between 0.07 and 0.32 m, and regression slopes ranging from 0.68 to 0.89. This indicates a slight underestimation of the highest waves by the model.

As expected, the IML-10 station yields the most accurate results due to its offshore location in deep water (445 m depth), where it is directly influenced by Atlantic swell entering through the Cabot Strait. In contrast, the GAL station shows the largest discrepancies. Located in shallow water, its environment is characterized by the presence of sandbars that can move rapidly, altering the local bathymetry and influencing wave propagation, making accurate predictions more challenging.

The same statistical analysis was conducted for the mean wave period and mean wave direction. These results are presented in figures and tables of Appendix . The analysis of the mean wave periods (Table ) shows that these are generally well represented by the system with most of the data falling close to the 1:1 line. Nevertheless, predictions exhibit some degree of imprecision for all stations with a RMSE ranging from 0.69 to 1.06 s. This inaccuracy is associated with low-amplitude waves; when considering only the highest 10 % of waves, RMSE values drop between 0.59 and 1.0 s, corresponding to errors of 9 % to 21 % relative to the average period observed across stations. As with significant wave height, the IML-10 station exhibited the best performance levels.

The comparison of mean wave direction (Table ) show a very good agreement between observations and forecasts. Results closely align with the 1:1 line with mean absolute error ranging between 18 and 27° for all waves. This error narrows to 15–19° for the 10 % of the highest waves indicating improved accuracy for more energetic sea states. The evaluation of system performances over the forecast lead time (Fig. ) indicates that model accuracy remains consistent up to +48 h, with no significant degradation in the quality of the results over time. Although our configuration provided reasonably good results during the summer period, further improvements in forecast skills could be achieved with some adjustments. In particular, increasing the spectral frequency resolution may improve the representation of wave growth in the high-frequency range of the spectrum, especially during the development of wind seas following calm conditions . Moreover, the use of alternative source term formulations for wind input and wave dissipation, such as ST4 or ST6, could potentially further enhance model performances .

Observations at stations LPM, CDR and GAL between 15 January and 31 March are used for assessing model skills during the winter period when the EGSL is partially covered by sea ice. Comparisons of significant wave height (Table , Fig. ) show that the forecast skills are considerably reduced compared to the summer period. The model strongly underestimates the wave height when sea ice is present. The scatter index (SI, see Appendix ) exhibits errors of 88 % to 142 % of the average observed wave height. In CDR and GAL particularly, the model largely fails to represent the observed sea state, as shown by the slope of the regression line near 0 in GAL, and did not capture the most extreme events recorded during this period, associated with values exceeding 3 m. This poses a potential challenge for coastal hazard predictions in a seasonally ice-covered environment like the EGSL and emphasizes the critical need to improve wave-ice interactions for such application.

Multiple factors can hinder the model's ability to accurately predict the sea state in the presence of sea ice. Obviously, errors associated to the wave attenuation parameters, which is known to be highly heterogeneous is one important potential cause of error. There is a strong dependence of wave propagation on the ice mechanical properties (elasticity parameters and flexural failure thresholds) that depends on sea ice history (salt content, fractures, refreezing), which can be highly heterogeneous. As a result, the decay rate is a complex function of these properties, including thickness and floe size, rather than being solely based on the wave frequency. However, these parameters are rarely constrained concurrently with observations of wave attenuation rates and are still absent from empirical formulations.

But more essentially, our results underscores how critical the accuracy of the sea ice predictions is for such application. Indeed, the rate of wave energy decay in sea ice observed in the EGSL is O(10-3-10-2) m⁻¹. This means that – even with some degree of uncertainty on this value – waves with periods of 3–10 s (typical for the EGSL) will lose 75 % of their energy within just a few kilometers from the ice edge when the ice concentration is 1. That distance is in the order – if not smaller – of the spatial resolution of the sea ice model used in our system, as well as that of typical regional sea ice models. Consequently, even minor uncertainties over one or a few grid cells in the position of the ice edge or in the ice fraction can significantly increase the error at a given location and time , regardless of how well wave propagation in sea ice is represented in the wave model.

In Fig. , a snapshot of the predicted ice concentration is compared with satellite-derived observations from Sentinel-2 multispectral images (level 2A, accessed from Copernicus Services on 9 March 2021 at 15:16:49 UTC). Here, ice extent has been retrieved with a simple threshold-based algorithm using RGB (red-green-blue) and SWIR (Short-Wave Infrared) bands. The cloud coverage was first filtered by applying SWIR > 0.3 (SWIR being the average of bands b11 and b12) and pixels were flagged as sea ice when RGB > 0.3 (RGB defined as the average of bands b2, b3 and b4). Note that this procedure has not been developed and tested to be quantitatively accurate in a broad range of situations or used for automatic detection of sea ice, but only to analyse and interpret results obtained here. Sea ice concentration was then computed by binning the ice pixels on the same grid as the RDPS-CGSL model (5 km spatial resolution) for comparison. At the scale of the EGSL, spatial fields of satellite-derived and model ice concentration look (at least visually) quite similar with sea ice being mainly concentrated south of the GSL. At a finer scale however, the observed coverage is much more heterogeneous, the ice being stretched into streaks of dense concentration surrounded by open water. Figure h shows the evolution of wave energy along a transect through the observed and predicted ice field (with a constant attenuation coefficient of α = 10⁻³ m⁻¹). Results shows that all wave energy decay within the first 50 km when propagating across the model ice field while significant wave energy is still present after 100 km for the observed field. This simple analysis illustrates how the presence of structures at the submesoscale in the ice field affects significantly the wave propagation in the EGSL. Despite their significance, the physical mechanisms driving the emergence of these structures are still very poorly represented in continuum sea ice models. Specifically, the extent to which characteristics of the fragmented sea ice – such as the floe size distribution (FSD) – influence its mechanical behavior and its deformation in response to external stresses like wind, waves, and oceanic currents remains unclear .

The inherent granular and patchy nature of sea ice also raises the problem of dealing with unresolved subgrid-scale heterogeneity and nonlinear processes. Indeed, existing parameterizations of wave energy attenuation by sea ice often scale with the ice concentration and, in some cases, with ice thickness and floe size. Within a model grid cell, these properties are typically represented by bulk quantities (e.g. total ice fraction, mean thickness and mean floe size), therefore not retaining the sea ice spatial distribution and variability. From a mathematical perspective, if hice denotes the ice thickness, F the wave attenuation, which vary nonlinearly with hice, and the overbar denotes averaging, then F(hice)‾≠F(hice‾). These two terms are equal only if F is linear or weakly nonlinear, which is not the case for many wave attenuation parameterizations currently used in WW3. This is analogous to numerical weather prediction where subgrid spatial heterogeneity such as topography, land use, soil and vegetation characteristics affects significantly calculations of land-surface-atmosphere heat and momentum exchanges when highly nonlinear processes are involved . Solutions to this long-standing problem have included the use of probability density functions (PDFs) to statistically represent heterogeneity or explicitly by partitioning grid cells into sub-grid tiles of homogeneous properties . In sea ice models such as CICE, subgrid-scale variability is accounted for using the floe size and thickness distribution (FSTD), which describes the fraction of the grid cell area covered by floes of a specific size and thickness. Incorporating FSTD into wave attenuation parameterizations could potentially lead to some improvements in modeling accuracy.

Comparisons conducted in the previous section highlight the global performances of the system under normal conditions. While several energetic conditions have been captured during the observation period, no extreme storm events leading to coastal hazards occurred. In order to evaluate the system performances under stormy conditions, we performed a hindcast run during the passage of the tropical storm Dorian, which hit the Gulf of St. Lawrence in September 2019 and caused severe damages across Atlantic Canada, especially in Nova Scotia, New Brunswick, Prince Edward Island and Magdalen Islands .

The low pressure system originated off the west coast of Africa on 19 August 2019 before developing rapidly into a category 5 hurricane while hitting the Bahamas on 1 September. It further moved northeastward while weakening along the North American coast. The system then passed over the Gulf of St. Lawrence as a post-tropical storm with winds over 100 km h⁻¹ and gusts of 157 km h⁻¹ recorded at Wreckhouse Brook, Newfoundland . Water levels observed in Canada’s Maritime provinces were particularly affected. A storm surge of 1.85 m was observed at the Shediac tide gauge in New Brunswick. In Nova Scotia, the surge reached 1.22 m and resulted in water levels of 0.61 m above the higher high water large tide (HHWLT) level (the average of the water level annual maximum over a period of 19 years), approaching the historical record and causing severe floods along the coast of the Prince Edward Island .

The storm passed through the Magdalen Islands on 8 September. At that time, two Nortek AWACs were moored at two different locations along the coast of the islands, Baie de Plaisance (BDP) and Pointe aux Loups (PAL), recording hydrodynamical conditions during the storm (see Fig. ). This dataset was used to evaluate the capabilities of the system to predict wave conditions during extreme events. Flood limits and maximum run-up extent were also assessed using field debris measurements conducted with GPS-RTK immediately after the storm passage. This dataset is used in the second part of this study to evaluate the ability of the system to predict the coastal TWL.

The maximum significant wave height recorded at Baie de Plaisance during Dorian reached 5.58 m at 02:00 UTC on 8 September and 6.82 m at Pointe aux Loups 10 h later, both associated with periods of ∼ 7.5 s. Figure  shows the predicted wave field at +48 h lead time for these two particular times. Figures  and  present the comparison of the significant wave height Hs, mean period Tm02 and mean direction θ time series observed and predicted by the system for the 0–24 and 24–48 h lead time windows. Forecasts show a very good agreement with the observations for all three wave parameters. Both amplitude and the timing of the storm peak are well captured by the model. There are also very few discrepancies between 0–24 and 24–48 h forecasts. Note however that before and after the passage of the storm at Baie de Plaisance, the system demonstrates a poor ability to represent the period of very low amplitude waves (< 0.2 m). Evaluation of wave models performances under extreme conditions have been carried out many times, with mixed outcomes . Best results often involved de-biasing wind fields and calibrating parameters values for wind generation source terms Sin . In our case, we kept the default parameter values for the wind generation/dissipation source terms and used HRDPS U10 as provided by ECCC without any de-biasing or calibration process. It confirms therefore that wind forcing accuracy is a major determinant of wave prediction skills and the importance of having access to high resolution, reliable wind forecasts in order to accurately predict extreme sea states. These results are very encouraging and show that this wave model configuration provide robust and reliable wave forecasts for further use in storm hazard prediction.