Introduction

Species distribution models (SDMs) have become increasingly important as the marine environment changes rapidly. Distributions of species in space and/or time are modelled as a function of a range of physical, environmental, and ecological variables. SDMs have been implemented with a diverse range of statistical tools, including classic regression models (e.g., generalized linear models- GLMs, generalized additive models- GAMs, and generalized linear mixed models GLMM) [1], machine learning algorithms (e.g., random forest and artificial network analysis) [2], climatic envelope methods (e.g., dynamic bioclimatic envelope models) [3, 4] and decision tree methods (e.g., boosted regression trees) [5]. To forecast climate change impacts on species distributions, environmental outputs from Earth system models (ESMs) and regional ocean modelling systems (ROMS) are commonly used as SDM inputs [4, 6]. ESMs simulate the Earth’s entire climate system, focusing on global-scale interactions between the atmosphere, oceans, land surfaces, and ice. They provide insights into long-term climate trends and variability [7, 8]. Variation in ESM projections can arise due to differences in forcing, process configuration, feedbacks, and horizontal and vertical resolutions [9, 10]. In contrast, ROMS concentrate on specific regions or domains of the ocean, offering higher spatial resolution to study fine-scale processes such as coastal dynamics, boundary currents, and mesoscale eddies [11, 12]. The low spatial resolution of ESMs can make it challenging to adequately represent regional processes, as they are primarily designed to emphasize global-scale dynamics [7]. ESMs also do not appropriately represent depths less than 50 m [7]. However, in many cases these models are the only available information to project species distribution under climate change scenarios [9].

Newfoundland and Labrador, located on the east coast of Canada, is a dynamic region that undergoes significant climatic variations, oscillating between warm and cold periods on decadal scales [13]. Within it, three major geographical subunits characterized by distinct productivity and a reasonably well defined major marine community (ecosystem production units), have been identified: Newfoundland Shelf (North Atlantic Fisheries Organization (NAFO) divisions 2J3K), Grand Banks (3LNO) and Flemish Cap (3M) [14, 15]. A key feature of the Newfoundland system is the cold intermediate layer (CIL), consisting of a water layer below 0°C. It forms during winter as a cold surface layer and remains as an intermediate layer separate from the surface (roughly 0–50 m) when waters warm in spring [13]. The presence of the CIL heavily impacts the Grand Banks because of the shallow depth of the oceanic plateau. This creates distinct temperature patterns within the Banks. Moreover, recent observations suggest a trend of warming ocean temperatures on the Grand Banks [16]. This warming may be influenced not only by warmer air temperatures that increase the heat transfer to the upper layers of the oceans, but also by a slowing down of the shelfbreak jet, which reduces the supply of the fresh, cold and nutrient-rich waters of the Labrador Current to the Grand Banks slopes [17–19].

In the Newfoundland and Labrador region, a shift from a warm to cold phase in the 1990s, coupled with intensive fishing, triggered alterations in ecosystem productivity, leading to a substantial decline in most groundfish stocks and a surge in invertebrate biomass on the Newfoundland Shelf [20] and the Grand Banks [21]. Responding to the groundfish collapse, several fishing moratoria were imposed in 1992 and 1994 on the Newfoundland Shelf and the Grand Banks respectively, to facilitate the recovery of groundfish stocks (e.g., Atlantic cod, yellowtail flounder, haddock, among others). The impact of the collapse was profound, affecting both harvesters and plant workers in the province, with over 35,000 individuals losing their source of livelihood [22]. In the present study, we focused on three key fisheries species of the Grand Banks that exhibited distinct responses to the collapse: yellowtail flounder (Limanda ferruginea), Atlantic cod (Gadus morhua) and snow crab (Chionoecetes opilio).

The yellowtail flounder fishery in the 3LNO division was closed from 1994 to 1997 and rebounded, currently operating as a Marine Stewardship Council (MSC) certified fishery [23, 24]. Yellowtail flounder inhabits the east coast of North America, from Newfoundland to the Chesapeake Bay [23], with biomass peaking on the Grand Banks at approximately 3°C, tolerating a temperature wide range [25]. In contrast, Atlantic cod stocks (3NO and 2J3KL) remain at low levels compared to historical baselines, with only a modest increase in biomass observed in the 2J3KL stock, primarily driven by the 2J3K divisions [26]. This situation resulted in the persistent closure of the directed commercial fishery for both stocks over 30 years, although a stewardship fishery has been allowed in the 2J3KL division [26, 27]. In 2024, the 2J3KL Atlantic cod commercial fishery opened again. Atlantic cod is found across the North Atlantic Ocean, avoids temperatures exceeding 12°C [28], and shows higher biomass at 5–6°C on the Grand Banks [15].

Following the collapse, Newfoundland and Labrador’s fishing industry diversified its target species, focusing on invertebrate like snow crab. Snow crab has become the province’s most valuable species, with a value of $761 million CAD in 2022, representing 58% of the total landed value of all fisheries resources in NL [22, 29]. A recent study found that the species is moving from subArctic to Arctic environments [30]. This suggests that with future warming, the species may experience biomass losses on the Grand Banks, where most of the Newfoundland and Labrador quota (59%) is currently allocated (https://www.dfo-mpo.gc.ca/fisheries-peches/decisions/fm-2023-gp/atl-14-eng.html).

The climatological changes mentioned above may present a pressing issue for Newfoundland and Labrador’s fishing sector, which in 2022 contributed $1.4 billion CAD to the local economy and currently employs over 17,000 people across 400 communities [29]. Thus, the objectives of this study were: i) to produce and compare biomass projections for three contrasting species on the Grand Banks of Newfoundland under two climate change emissions scenarios; and ii) to evaluate the importance of climate model choice in these types of projections. In particular, we compared species biomass projections using the Coupled Model Intercomparison Project (CMIP) 6 earth system models IPSL-CM6A-LR [31] and the GFDL-ESM4 [32] under low (SSP1-2.6) and high (SSP3-7.0) emissions scenarios, and the ROMS Atlantic Canada model (from now on referred as ACM) [19, 33] as climate forcings. While these projections should be interpreted with caution, they provide valuable guidance for fisheries management strategies by detecting changes in habitat suitability and anticipating future biomass trends. In assessing variation in biomass projections, we specifically analyzed the influence of climate model and emissions scenario choice, including variations in bottom temperature under low (SSP1-2.6) and high (SSP3-7.0 and SSP4-6.0) emissions pathways.

Materials and methods

Oceanographic and biological survey data

Fisheries and Oceans Canada (DFO) has been conducting annual random stratified multispecies trawl surveys in Newfoundland and Labrador since 1971, with these surveys experiencing important modifications in survey design, coverage, species recorded, and gear over time [34]. One major change in these surveys was the transition from the use of the Engels trawl to the Campelen trawl gear (i.e., 1800 shrimp trawl) in 1995–1996 [34]. This change improved catches of small-sized fishes and marked the start of the time series for commercial shellfish species being recorded. Georeferenced data on presence/absence and biomass for the focal species of this study as well as water temperature and depth comes from DFO spring surveys on the Grand Banks between 1996 and 2019. To predict biomass, we created grids with a 5 x 5 km spatial resolution which included gridded bathymetry information with a 15 arc second spatial resolution from the general bathymetric chart of the oceans project (GEBCO) (https://www.gebco.net). Spring bottom temperature data were interpolated over the Grand Banks with a spatial resolution of 0.1° x 0.1° [16] (Fig 1). Bathymetry data represented depths from 35 to 750 m since those were the depths sampled in the RV trawl survey.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Map of the average spring bottom temperature during the historical period (1996–2019) interpolated over the Grand Banks.

Isobaths indicated in light grey; Northwest Atlantic Fisheries Organization NAFO divisions boundaries are indicated with grey rectangles. Black arrows show the main currents of the region. Base map layer from Natural Earth (https://www.naturalearthdata.com/about/).

https://doi.org/10.1371/journal.pclm.0000520.g001

Climate models and emissions scenarios

We considered both low and high emissions scenarios to assess their impacts on snow crab, yellowtail flounder, and Atlantic cod distributions on the Grand Banks. We used three models to characterize these scenarios. The ACM uses ROMS version 3.5, a terrain-following, free-surface, primitive equation ocean model [33, 35]. It was configured with 30 vertical levels (layers are thinner in shallower water and thicker in deeper water), with a minimum water depth of 10 m, and an approximate horizontal resolution of 10 km (240×120 horizontal grid cells). The model encompasses the Gulf of Maine, Scotian Shelf, East Newfoundland Shelf, Grand Banks, and the Gulf of St. Lawrence, and has been demonstrated to accurately capture regional circulation patterns [36], and represent biogeochemical properties well [33]. The ACM was calibrated to observed values [33] and forced by anomalies from the GFDL high-resolution coupled model, CM2.6, with a rapid increase in CO₂ that resembles the SSP4-6.0 scenario [18]. In comparison, IPSL-CM6A-LR [31] and GFDL-ESM4 [32] are ESMs with a coarser spatial resolution of approximately 100 km. These models produce a much poorer agreement with observed temperatures, salinity, nitrate, and chlorophyll observation in the Grand Banks than the ACM [33]. ESMs also do not appropriately represent depths less than 50 m [7].

For biomass projections, we used annual average sea bottom temperature (potential temperature on bottom–thetao_bot) sourced from the Coupled Model Intercomparison Project (CMIP) 6 ESMs IPSL-CM6A-LR [31], GFDL-ESM4 [32] and the ROMS ACM [19, 33]. We used the shared socioeconomic pathway (SSP) and representative concentration pathway (RCP) scenarios SSP3-7.0 for IPSL and GFDL, and SSP4-6.0 for ACM as the high emissions scenarios due to availability, and the low emissions scenario SSP1-2.6 [37]. Unfortunately, the low emissions scenario was not available for ACM (Fig 2A and 2B). Note that while the specifics of the high emissions scenarios do not align perfectly among models (SSP3-7.0 and SSP4-6.0), they are similar enough in their warming trends, allowing for general comparisons. Additionally, the ACM outputs have been smoothed [18], whereas the ESMs have not, explaining the differences in the variability of the temperature field. Finally, we bias-corrected annual bottom temperature projections of the climate models to the local spring observations since they were consistently higher. To align our projections with observed baseline conditions, we calculated the mean temperature from the climate model and the mean observed temperature values from 2015 to 2019 (we selected these years because the projections and observations overlapped). The bias was then determined as the difference between these means and used to adjust the model output (Figs I and J in S1 Text). This adjustment aligns the model temperatures with observed conditions.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Maps and time series of average annual bottom temperature projection on the Grand Banks.

a, Time series of average annual bottom temperature for the model domain. The historical period is represented by a black line, while the IPSL-CM6A-LR, GFDL-ESM4, and ACM models are indicated by yellow, purple and green lines, respectively. Dashed lines represent the low emissions scenario (SSP1-2.6), and solid lines depict the high emissions scenario (SSP4-6.0 for ACM, and SSP3-7.0 for GFDL and IPSL). b, Maps of mean annual bottom temperature projections by climate model (GFDL, IPSL and ACM) and RCP scenarios at the end of the century (period 2071–2100). Base map layer from Natural Earth (https://www.naturalearthdata.com/about/).

https://doi.org/10.1371/journal.pclm.0000520.g002

Biomass uncertainty evaluation

We assessed point-wise prediction uncertainty of species distribution models during the baseline period (1996–2019) by conducting 100 simulations based on the joint precision matrix of our model [47]. The precision matrix, often referred to as the inverse covariance matrix, characterizes the relationships between variables assuming a multivariate normal distribution [38]. The variability in these simulations, and hence the level of prediction uncertainty, is directly influenced by the precision matrix. High variability in biomass at each grid point reflects greater uncertainty in the model’s predictions. We repeated this approach to assess biomass projections from each climate model under the high emissions scenarios at the end of the century (period 2077–2100). Finally, we partitioned uncertainty in the biomass estimates among climate models and RCPs by fitting a linear model, with annual biomass estimates as a response variable and climate models (GFDL, IPSL and ACM) and scenarios (low and high emissions) as covariates, allocating residual error to parameter uncertainty (e_i) as in formula 3.

Biomass_i represents the annual biomass estimate for the ith observation; β₀ is the intercept term; β₁ and β₂ are coefficients corresponding to the predictor variables X₁ (climate model) and X₂ (emissions scenario), respectively. Finally, e_i represents the residual error term, capturing unexplained variation in the biomass estimates.

Dominance analysis is a statistical technique used to assess the relative importance of predictor variables in explaining variance in a dependent variable, and was applied to evaluate the relative importance of each component [48, 49]. We smoothed the results by computing 10-year averages to mitigate noise and highlight the underlying trends. We acknowledge that having three climate models and only two scenarios–only one in the case of the ACM model–may induce bias in assessing predictor importance.

Results

For all species, models that best explained spatial patterns of biomass distribution included depth and temperature as fixed effects and the spatial random field (Tables A-C in S1 Text). The spatial random field represents biomass deviations in space that are not accounted for by covariates. Its inclusion substantially improved model performance, especially for yellowtail flounder and Atlantic cod. Model estimates are available in Table D in S1 Text.

We used the model fit to predict biomass estimates of the focal species on the Grand Banks of Newfoundland during 1996–2019. Then, we forecasted to 2100 under the low and high emissions scenarios. The predictive capacity of the models varies among species, with yellowtail flounder having the highest predictability (R = 0.74; AUC = 0.97), followed by snow crab (R = 0.55, AUC = 0.89) and Atlantic cod (R = 0.55, AUC = 0.82) (Fig G in S1 Text). We also assessed the out-of-sample predictability, training our model with data from 1996 to 2016 and testing it for the period 2017:2019. The predictability was (R = 0.7; AUC = 0.97) for yellowtail flounder, (R = 0.41, AUC = 0.89) for snow crab and (R = 0.47, AUC = 0.79) for Atlantic cod (Fig H in S1 Text).

Our findings underscored a long-term decline in the projected biomass of snow crab for all climate models and scenarios except for GFDL under the low emissions scenario (Fig 3). If we focus on the values at the end of the century (2077–2100), greater losses were observed for the high emissions scenario, especially for IPSL model -47% [± 1.67% SD], followed by ACM -10.4% [± 1.59% SD] and GFDL -14.68% [± 1.59% SD]. Under the low emissions scenario, biomass trends fluctuated around the historical average. Values at the end of the century (2077–2100) for IPSL projected biomass losses of -3.8% [± 1.6% SD], whereas GFDL projected biomass gains of 3.7% [± 1.57% SD].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Projections of snow crab biomass by climate model and emissions scenario.

a, Projections under low emissions (SSP1-2.6, left) and high emissions scenarios (SSP4-6.0 for ACM and SSP3-7.0 for IPSL-CM6A-LR and GFDL-ESM4 scenario, right). b, Projections by climate models for low and high emissions scenarios. Biomass changes are relative to the predicted values of the reference period (1996–2019), indicated by shaded grey area. Solid colored lines depict average projected biomass, while shaded areas indicate standard deviation. Zero change is represented by a horizontal dashed line.

https://doi.org/10.1371/journal.pclm.0000520.g003

We also observed a long-term decline in yellowtail flounder biomass for the IPSL model under the low and high emissions scenarios -8.4% [± 0.38% SD] and -29% [± 0.36% SD], respectively (Fig 4). Biomass projections at the end of the century (2077–2100) for the GFDL model under both emissions scenarios indicated biomass values close to the historical average, -1.4% [± 0.38% SD] for the low emissions and 0.39% [± 0.38% SD] for the high emissions scenario. Finally, the ACM model suggests a small increase in biomass at the end of the century (period 2077–2100) of 4.45% [± 0.34% SD].

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Projections of yellowtail flounder biomass by climate model and emissions scenario.

a, Projections under low emissions (SSP1-2.6, left) and high emissions scenarios (SSP4-6.0 for ACM and SSP3-7.0 for IPSL-CM6A-LR and GFDL-ESM4 scenario, right). b, Projections by climate models for low and high emissions scenarios. Biomass changes are relative to the predicted values of the reference period (1996–2019), indicated by shaded grey area. Solid colored lines depict average projected biomass, while shaded areas indicate standard deviation. Zero change is represented by a horizontal dashed line.

https://doi.org/10.1371/journal.pclm.0000520.g004

SDMs projected a long-term decline in Atlantic cod biomass for the GFDL-low emissions scenario of -11.5% [± 0.53% SD] at the end of the century (2077–2100), and an increase in biomass for the rest of model and scenarios (Fig 5). The biomass projection for IPSL-low emission scenario indicates a gain of almost 9.8% [± 0.53% SD] at the end of the century. Under the high emissions scenario, GFDL oscillates around the historical average but showed an increase of 3.12% [± 0.53% SD] at the end of the century (period 2077–2100). IPSL and ACM both showed a similar trend of biomass gain at the end of the century, reaching values of 15.21% [± 0.55% SD] and 19.2% [± 0.52% SD], respectively.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Projections of Atlantic cod biomass by climate model and emissions scenarios.

a, Projections under low emissions (SSP1-2.6, left) and high emissions scenarios (SSP4-6.0 for ACM and SSP3-7.0 for IPSL-CM6A-LR and GFDL-ESM4 scenario, right). b, Projections by climate models for low and high emissions scenarios. Biomass changes are relative to the predicted values of the reference period (1996–2019), indicated by shaded grey area. Solid colored lines depict average projected biomass, while shaded areas indicate standard deviation. Zero change is represented by a horizontal dashed line.

https://doi.org/10.1371/journal.pclm.0000520.g005

The analysis revealed a consensus among earth system models (IPSL and GFDL) regarding the locations with the most pronounced losses in snow crab biomass, specifically at the north and south of the Grand Banks (Fig 6). However, ACM depicted a less pronounced decline (Fig 6A). The majority of yellowtail flounder biomass changes were observed on the southeast shoal of the Grand Banks. GFDL and ACM mostly agreed for projected spatial biomass changes. In contrast, IPSL suggested a stronger decline in the southern part of the southeast shoal (Fig 6B). Spatial biomass patterns for Atlantic cod showed the biggest disagreement among climate models (Fig 6C). GFDL predicted losses in the north and southern edge, with some gains on the southern shoal. IPSL indicated biomass losses on the southeastern shoal and gains on the rest of the Banks. Finally, ACM projected gains mostly in the south of the Banks.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 6.

Spatial patterns of species biomass changes (in kg/25 km2) for a, snow crab; b, yellowtail flounder and c, Atlantic cod on the Grand Banks of Newfoundland by climate model (GFDL-ESM4, IPSL-CM6A-LR and ACM) under the high emissions scenarios (SSP4-6.0 for ACM and SSP3-7.0 for IPSL and GFDL) during the 2077–2100 period relative to the historical period (1996–2019). The Avalon Peninsula is the southern piece of land (in grey). Base map layer from Natural Earth (https://www.naturalearthdata.com/about/).

https://doi.org/10.1371/journal.pclm.0000520.g006

Uncertainty

We observed marked spatial variability in SDM biomass projections during the historical period (1996–2019), with standard deviations of biomass ranging from 0 to 4 kg/25 km² for snow crab, yellowtail flounder and Atlantic cod (Fig 7-SDM). Areas of high variability were associated with areas of biomass absence, likely driven by the gamma component of the delta-gamma model (Figs B, D and F in S1 Text). For snow crab, the periphery and the southeast shoal of the Grand Banks showed high variability (Fig 7A). For yellowtail flounder, variability was higher in the northern part of the Banks and on the periphery (Fig 7B). In the case of Atlantic cod, variability was pronounced around the Avalon peninsula and on the periphery of the Banks (Fig 7C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 7.

Spatial uncertainty of SDM biomass estimates (historical period; 1996–2019) and climate model projections (GFDL, IPSL and ACM) at the end of the century (2077–2100) measured as the standard deviation from 100 simulation draws for (a) snow crab, (b) yellowtail flounder and (c) Atlantic cod projected biomass. Base map layer from Natural Earth (https://www.naturalearthdata.com/about/).

https://doi.org/10.1371/journal.pclm.0000520.g007

Variability in biomass projections amplified between the historical period to the end of the century (2077–2100), where the standard deviation of the biomass increased, especially for yellowtail flounder (30 kg/25 km²) (Fig 7).

We partitioned uncertainty to assess the relative contribution of each component: climate model, SSP-RCP scenario and SDM parameters (Fig 8). For all three species—snow crab, yellowtail flounder, and Atlantic cod—climate model uncertainty was the dominant factor contributing to biomass uncertainty across all years, consistently showing the highest influence (above 50%). However, the contribution of SDM parameters and scenarios varied over time. In the three cases, the influence of SDM parameters and scenarios increased slightly after 2050, indicating that uncertainty associated to SDMs parameters and SSP-RCP scenarios become more important as projections extend further into the future.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 8. Relative uncertainty in biomass projections for a, snow crab; b, yellowtail flounder and c, Atlantic cod, partitioned across climate models (i.e., IPSL, GFDL and ACM), emissions scenarios (low-SSP1-2.6 and high-SSP4-6.0 & SSP3-7.0) and SDM parametrization.

https://doi.org/10.1371/journal.pclm.0000520.g008

Discussion

Depth and temperature are known to be key factors that determine distribution and survival of many species, including snow crab [50], yellowtail flounder [51, 52] and Atlantic cod [28, 53]. As expected, both depth and temperature were relevant variables to explain spatial distribution. The best fitting models also included the spatial random field that explains variability not captured by the covariates.

Projected changes in species biomass to 2100 under low and high emissions scenarios showed an overall decline in snow crab and yellowtail flounder biomass compared to the historical period (1996–2019), whereas Atlantic cod is expected to gain biomass across the 21^st century under the high emissions scenario. However, in the GFDL simulation, biomass declines through the mid-century before increasing. While there are differences in the magnitude of biomass projections across climate models, a consensus in the direction of changes was evident in all models except for GFDL-low emissions scenarios. This divergence can be attributed to the GFDL-low emissions scenario forecasting a decline in bottom temperature by the end of the century, in contrast to temperature increases projected by the other models and scenarios. The IPSL model projected the biggest warming on the Grand Banks.

Changes in species geographic distributions can be analyzed by focusing on their range boundaries. At leading edges, species expand into new territories as environmental conditions improve, while at trailing edges, population extirpations lead to range contractions [54]. Marine ectotherm species tend to fully use their potential latitudinal ranges in relation to their thermal tolerance limits [55], and distribution shifts are thought to be higher at the leading edge than at the trailing edge [56], although this is case specific [57]. Snow crab are commonly found in subpolar and Arctic regions, favoring cold water environments below 5°C [58]. On the Grand Banks of Newfoundland, higher biomass has been observed at temperatures close to 0°C [25]. Being stenothermic, snow crabs are believed to be highly responsive to changes in bottom temperature, particularly during their juvenile period [59]. Snow crab on the Grand Banks of Newfoundland are at the trailing edge of their distribution. This is because they are near their upper thermal threshold, making them particularly susceptible to ocean warming. A recent study has detected a general shift of snow crab distribution from the Newfoundland region into the Arctic [30]. Here, we found that snow crab biomass on the Grand Banks is expected to decline throughout the 21^st century for all models under the high emission scenarios, while remaining close to the historical period under the low emissions scenario.

On the Grand Banks, yellowtail flounder thrive in sea bottom temperatures ranging from 1–6°C, with biomass peaking at around 3°C [25]. Despite the species’ ability to tolerate a broad range of temperatures, there is evidence that stock productivity declines significantly at both extremes of the temperature range [23]. The case of yellowtail flounder is interesting as uncertainty widely amplifies over time, which was not the case for snow crab and Atlantic cod. Yellowtail flounder distribution is centered around the southeast shoal of the Grand Banks due to the presence of a mild current system, facilitating the retention of eggs and larvae [23]. This region is considered as a nursery ground for yellowtail flounder [51], and is the area in which higher warming is expected on the Grand Banks (Fig 2B). We found that yellowtail flounder biomass projections are expected to remain close to the historical average for GFDL and slightly higher for ACM, but to be negatively impacted for the IPSL model under high and low emissions scenarios.

Atlantic cod tolerates higher temperatures than the other two species, with a peak in biomass found at 5°C on the Grand Banks [25]. Our analysis indicates that by the end of the century, most climate models (excluding the GFDL-low emissions scenario) predict an increase in Atlantic cod biomass on the Grand Banks. In particular, the ACM showed the greatest biomass gains. These results contrast with findings in the Gulf of Maine, where biomass decrease was noted with ocean warming [60, 61], but are in agreement with forecasted thermal habitat for cod on the Newfoundland and Labrador Shelf [62]. The difference can be attributed to the distinct thermal environments of these regions. The Gulf of Maine has an average bottom temperature of 7.1°C [60], while the Grand Banks experiences significantly cooler temperatures, averaging 1.34°C during our historical period (1996–2019). This suggests that Atlantic cod could potentially benefit from moderate warming in the Grand Banks, while those in the Gulf of Maine may be experiencing temperatures beyond their physiological limits, leading to declines in biomass [54]. Gains in Atlantic cod biomass are expected to occur mostly in the southern Grand Banks (3NO stock).

While temperature is a key factor influencing species biomass and distribution, other factors may also play a significant role. In the Grand Banks, the community structure has shifted since the collapse in the early 1990s [63], altering predator-prey interactions and productivity dynamics. Concerns have been raised that fishery-induced changes in these interactions could significantly hinder or even prevent the recovery of depleted populations like Atlantic cod [64]. Moreover, documented changes in life history traits—such as reduced age and size at maturity—and increased natural mortality may complicate recovery, even when thermal habitat conditions are optimal [26, 65, 66]. Continuing fishing and bycatch in low productivity conditions could further delay recovery [67].

Variability in the predictive capacities of the SDMs directly impacts the reliability of biomass projections. For instance, our leave future out cross-validation showed that yellowtail flounder model exhibited the highest predictability (R = 0.7; AUC = 0.97), indicating more confidence in biomass projections. Conversely, the Atlantic cod model had the lowest predictability (R = 0.47; AUC = 0.79). Snow crab falls in between with moderate model predictability (R = 0.41; AUC = 0.89). However, the RV trawl survey has low catchability for this species [68]. Hence, our model may be underestimating the impact of ocean warming on snow crab biomass.

The low spatial uncertainty of the SDMs indicated consistency in the biomass predictions during the historical period (1996–2019). However, projecting to 2100 under the high emissions scenario (SSP3-7.0 for IPSL and GFDL and SSP4-6.0 for ACM) showed increased uncertainty, especially for yellowtail flounder. This finding aligns with other species distribution studies [47, 69], reflecting challenges in extrapolating predictions to novel conditions and/or the model’s capacity to capture the underlying mechanisms governing species distributions [49]. In all species, there is a notable overlap between regions with biomass absence and areas of high uncertainty (Figs A, D and F in S1 Text). Furthermore, observation uncertainty, linked to bias in the sampling coverage, can lead to an incomplete depiction of a species’ entire environmental niche [70]. While the Grand Banks region has relatively comprehensive sampling coverage, we anticipate observation uncertainty to be more pronounced near the Avalon peninsula, since in this area independent inshore sampling is performed, and at the periphery of the Grand Banks, corresponding to deeper waters that are less sampled [71]. Depth sampling limitation may elucidate the higher uncertainty observed in the biomass predictions for the three species at the periphery of the Grand Banks. Another possible explanation is related to the spatial resolution of the climate models, increasing bias near the slope of the Banks.

Variation in spatial warming among climate models is crucial to understand uncertainty surrounding species distribution projections. In the present analysis, we found greater variation among climate models than among scenarios. We also found differences in the variability of the temperature projections, with ACM temperature increasing steadily, while the ESMs, GFDL and IPSL showed high variability. This is likely due to low spatial resolution of ESMs, leading to incorrect circulation patterns due to inaccuracies in the bathymetry (Figs K and L in S1 Text). In a dynamic coastal region like the Grand Banks, changes in circulation patterns, particularly the strength of the shelf break current, play a crucial role in determining sea bottom temperature [18, 19]. These small-scale circulation features are poorly captured by ESMs, and can even vary among ROMSs [19]. In the same way, the cold intermediate layer is a key feature of the Banks that also affects the temperature pattern distribution, with colder waters prevailing in the northern Grand Banks due to the advection of winter-origin waters from the Labrador Shelf, and warmer temperatures observed in the southern Grand Banks, including the shallow southeast shoal [13]. The CIL falls within a specific depth range (~50–200 m); thus models with high bathymetry bias would likely misrepresent the CIL.

Supporting information

Acknowledgments

We would like to thank those who assisted with the dataset: Andrew Cuff, Andrew Roberts and Darrel Mullowney. We would also thank the DFO technicians and scientists involved in the research surveys. Advanced computing resources are provided by the Digital Research Alliance of Canada, the organization responsible for digital research infrastructure in Canada, and ACENET, the regional partner in Atlantic Canada. ACENET is funded by Industry Science & Economic Development, the provinces of New Brunswick, Newfoundland & Labrador, Nova Scotia and Prince Edward Island, as well as the Atlantic Canada Opportunities Agency. The present study has been proofread using ChatGPT [version GPT-3.5] to enhance clarity and readability for effective communication. After using this tool, R.R.D. reviewed and edited the content as needed and takes full responsibility for the content of the publication.