Unstructured grid modelling can accurately describe complex coastal morphologies and transitional environments such as estuaries, lagoons and salt marshes, where key biogeochemical processes often occur (Bonamano et al., 2024; Cravo et al., 2024; Newton et al., 2014). This improves the representation of coastal ecosystems by enabling these environments to be included in the computational domain. The flexibility of the mesh enables changes in resolution in areas with larger spatial scales (e.g. offshore areas). Moreover, the specific parallel implementation (Micaletto et al., 2022) enables the inclusion of all 51 BFM state variables at a grid spacing of approximately 300 metres. The nested approach of ShyBFM offers flexibility in calibration of model parameters to the local areas, especially for the biogeochemical rates.
However, if not handled carefully, the limitations of unstructured grid modelling can potentially override its advantages. A discussion on the results obtained is then needed to understand the capability and limitations of SB.
Model performance
The analysis of chl a indicates the capability of SB to reproduce the phytoplankton biomass cycle, which shows bloom events between late autumn and early spring (Figs. 8a and 9a), a common feature of the northern Adriatic Sea due to the balance between light and nutrient availability and water column stratification (Vichi et al., 2003). Compared to NB chl a is well reproduced with a 45 % lower RMSE and an IQR closer to climatology (Table 2). However, there is still an overall underestimation compared to climatology, except in autumn (Fig. 9a), which shows a slight overestimation. In general, the timing of the blooming periods is consistent with the 1D modelling analysis of Vichi et al. (2003) for the area in front of the Po River. Even though the spatial variability of chl a is well reproduced, some limitations remain, particularly in the upper layers of the winter vertical profile (Fig. 6a) where the upper part of the EMODnet distribution is not well captured by the models. This discrepancy may be due both to model limitations in reproducing strong algal blooms and to inconsistencies in the data comparison, as the analysis compared a ten-year simulation (2000–2009) of SB, with a climatology spanning more than 30 years (1981–2017). It is worth noting that SB was able to reproduce two relevant phytoplankton blooms that can be noticed in 2001 and especially in 2002 (Fig. 8), during which other nutrients also exhibit consistent peaks. These events are related to the exceptionally high discharge of the Po and other rivers (Aragão et al., 2024; Verri et al., 2018), which resulted in higher nutrient loads in 2002 than the 2000–2009 mean, leading to significant mucilage events (Precali et al., 2005; Russo et al., 2005). The surface comparison of chl a with EMODnet (Fig. 5) shows similar patterns for spring, summer and autumn with a positive bias around the Po mouth and a negative bias along the coast. This pattern can be observed also in other variables of Fig. 5 and may result from uncertainties in the river nutrient loads. Other discrepancies arise from the influence of the parent model that may be strong, particularly in the northern boundary where fluxes enter the domain following the main circulation and reducing the effectiveness of domain-specific parameterizations.
The effect of nutrient fluxes entering through the northern boundary is evident in Fig. 5, especially for DIN, but also for the other variables, where larger overestimations occur in the northern part of the domain and can be attributed to the parent model (NB) due to the southward main circulation (see Fig. SI2.5 for a comparison of NB against EMODnet climatology and Fig. SI2.6 for a circulation comparison). Overall, DIN performance is good, with a 47 % reduction in RMSE compared to NB. In the final parameterization, nitrate uptake has been increased, resulting in improved agreement with climatology. However, this adjustment led to larger discrepancies with EMODnet along the coast, particularly in winter and autumn.
The phosphate performance for SB is comparable to NB. However, the IQR of SB phosphate (Fig. 9c and Table 2) is much closer to the climatology compared to NB. It is important to stress that phosphate plays a key role as a limiting nutrient in the northern Adriatic Sea, and a reliable representation of its distribution can significantly improve phytoplankton biomass equilibrium (Knjaz et al., 2024). The coupled model of Scroccaro et al. (2022) reported a substantial overestimation of phosphate in the ER coastal area, where chl a overestimation was also found, attributing this discrepancy to high uncertainty in land-borne river nutrient loads. Consistently, the comparison with EMODnet in Fig. 5 shows that the largest phosphate discrepancies occur around the Po mouth and along the coast, where several Apennine rivers discharge, supporting the role of uncertainty in river nutrient inputs. The available data enabled the computation of mean annual river load concentrations; however, actual concentrations may vary seasonally, thereby introducing additional uncertainty into the modelling framework.
Silicate results are generally satisfactory, with a 10 % reduction in RMSE (Table 2) compared to NB; however, vertical profiles (Fig. 6) show large discrepancies. The climatology indicates an increase in silicate with depth that is not reproduced by the modelling frameworks. Moreover, the IQR is underestimated by almost 50 %. The simple benthic return parameterization applied in the current model may be insufficient to represent the benthic-pelagic fluxes associated with early diagenetic processes that are particularly relevant in coastal areas (Paraska et al., 2014; Vichi et al., 2003). Silicate plays a significant role in the chlorophyll budget, as it is a fundamental component of diatom growth (Krause et al., 2019). However, silicate loads are highly uncertain due to a lack of continuous measurements. Furthermore, riverine silicate inputs are often estimated using indirect methods (Ludwig et al., 2009), which are less reliable than direct observations.
The SB model successfully reproduces the seasonal oxygen cycle (Figs. 8d and 9d), but it is affected by a systematic positive bias of 10 %–15 % (Fig. 9d). Although surface performance is good (Fig. 7), the lack of light attenuation with depth due to ISM prevents the representation of realistic vertical profiles (Zang et al., 2020). The CDOM could be also responsible for a significant fraction of light absorption (Álvarez et al., 2023; Campanelli et al., 2017) in coastal areas, but it is not considered in the configuration used here. This limitation is particularly relevant for the considered testcase, which is characterised by potentially high ISM and CDOM loads from the Po River, affecting phytoplankton photosynthetic processes (Bignami et al., 2007; Penna et al., 2004). Oxygen spatial IQR is also largely underestimated compared to climatology in both SB and NB with values that are less than one third of the climatological range (Fig. 9d and Table 2). Vichi et al. (2003) highlighted the importance of ISM in regulating phytoplankton biomass and nutrient uptake, showing that higher ISM concentrations were associated with lower chlorophyll a and higher phosphate concentrations. Both ISM and CDOM omissions likely contribute to the vertical profile discrepancies reported here. The inclusion of optical formulation accounting for CDOM (Álvarez et al., 2023) is identified as a future development. Moreover, the benthic processes may account for a large fraction of oxygen depletion in the bottom water (Seitaj et al., 2017) emphasizing the need to include benthic-pelagic coupling in coastal marine models. However, the use of more complex models should be supported by measurements of sediment variables that are lacking in the area, otherwise leading to high uncertainties in the benthic variables and their feedback to the water column. This is the main reason for the choice of a simpler benthic return closure in this first implementation of SB.
It is worth noting that in SB tides were added, which are absent in NB. Tides can reach 1 m amplitude in the northern Adriatic Sea and they can significantly enhance the mixing in water column, potentially explaining differences in stratification between SB and NB (Fig. 6).
Temperature influences phytoplankton and zooplankton growth through the metabolic (Q10) dependence in BFM, yet the temperature differences in Fig. 3 do not map onto the chl a differences in Fig. 5: the winter underestimation of chl a coincides with a slight warm bias (∼1 °C; Fig. 3c), while a much larger spring cold bias (∼-3 °C; Fig. 3g) leaves the chl a difference (Fig. 5e) essentially unchanged from the other seasons. This suggests that phytoplankton biomass here is controlled less by temperature than by river inputs and open-boundary conditions. Discrepancies in salinity (Fig. 3) can be attributed primarily to uncertainty in the prescribed river discharge, which sets the freshwater input and hence the strength and extent of the low-salinity plume. In addition, biases in the simulated coastal circulation, which advects the plume southward along the shelf, may displace the freshwater signal relative to observations; however, in the absence of current measurements this contribution cannot be quantified directly. SB and NB show similar temperature and salinity patterns (Fig. SI2.3) and vertical profiles (Fig. 7), indicating that these features are largely inherited from the shared forcing and boundary conditions rather than introduced by the differences in the modelling framework (grid, coupling, numerics, etc.).
The northern Adriatic coastal zone is an advection-dominated system, in which wind and river forcing govern the along-shelf transport and spread of phytoplankton biomass (Polimene et al., 2006; Zavatarelli et al., 2000). In such a regime the first-order upwind scheme adopted for horizontal advection, although robust, monotone and positivity-preserving, properties that are desirable for biogeochemical tracers, is only first-order accurate, and its diffusive truncation error tends to smooth the sharp cross-shelf and plume gradients that characterise the region, although no sensitivity experiments were carried out to isolate it from the effects of the boundary conditions, river loads or the specific choice of biogeochemical parameters. Nevertheless, a higher-order advection scheme would be expected to improve the simulated spatial patterns and is left for future work.
Pelagic remineralization rates driven by bacterial activity also play a crucial role in coastal waters (Nixon, 1981) and were the focus of part of the calibration phase, with significant impacts in oxygen and inorganic nutrient concentrations. However, no reference data were available for direct comparison with modelled bacterial biomass.
Overall, the results demonstrate the ability of SB to reproduce the physical and biogeochemical cycles of the testcase area with promising results when implemented accurately with specific parameter calibration and output validation. SB outputs are comparable to those of NB, which serves as the benchmark model, and for several variables, SB shows improved performance relative to the climatology.
In terms of computational cost, SB required around 695 core-hours per simulated year with 36 MPI processes, while NB used around 1728 core-hours per simulated year with 288 MPI processes. This means a wall clock of around 19 h for SB and 6 h for NB per simulated year. NB was almost 3 times faster but used around 2.5 times more core-hours compared to SB.