Figure 11. Historical model biases in warm-season MHW properties vs. the projected future changes over the global cells of kelp under SSP2-4.5 by the end of 21st century for the a.-c. the continuous warm-season MHW (Dc, day), d.-f. the accumulated heat stress over the continuous period (Ac, °C·day).
Over the coral and kelp cells, the pattern of future changes in the five thermal properties are similar to the changes globally, with increases across most cells in Dc, HSpeak and Ac as well as many decreases in HRc and Dp, and higher magnitude of changes under the higher emission scenarios (Figure 8, 10; Table S1-6). The mean changes over the kelp cells are close to the global levels, while those over the coral cells are greater than the global averages in all three models (Table S1-6), in part because warm-water corals exist in the tropics and subtropics where the limited seasonality cause greater chance of continuous heat stress occurrence, and that contribute to higher accumulation of heat stress. For example, the mean level of Dc and Ac under SSP2-4.5 in GFDL-ESM4 is 178 day and 281 °C·day, respectively, while that across global ocean is 143 day and 246 °C·day (Figure 8; Table S1). For HSpeak, HRc and Dp, however, the differences between the projected changes over the coral cells and the whole ocean varies among the models, ranging from negative to positive (Table S1-3). While it is unknown to what extent corals and kelp could adapt to recent thermal history in the future, the projected changes might be smaller if the thermal threshold increases due to acclimation and or adaptation. For example, the global mean increase in Dc by 2100 in the coral and kelp cells is 152 and 101 day shorter, respectively, using a thermal threshold based on the previous 60 years (2041-2100) rather than the historical period (Figure 8, 10).
Similar to the global patterns, the regional changes in these thermal properties also tend to be larger in CESM2-WACCM and MRI-ESM2 than those in GFDL-ESM4, which could be related to processes driving larger biases in those models (Figure 8, 10, S11-14). The projected increases of Dc and Ac across the coral and kelp cells show statistically significant positive relationships with the present-day model biases in each of those cells (Figure 9, 11, Table S7-12). The positive relationships, particularly for CESM2-WACCM and MRI-ESM2, indicates that coral and kelp cells with large positive biases are also exhibiting large projected changes in those same variables. This suggests that every 1 day increase in the model biases, for example in Dc, the projected values increase by 0.5, 1.2, 1.8 days in GFDL-ESM4, CESM2-WACCM and MRI-ESM2, respectively. However, the relationships between the projected changes and model biases for HSpeak, HRc and Dp vary among the models with some of them not statistically significant (Table S7-12).
5. Discussion and conclusions
In this study, we evaluate the model biases in simulating five thermal properties of warm-season MHWs from three CMIP6 models, and analyze their potential role in the projected future changes with specific focus over the global regions of warm-water corals and kelp forests. By the end of 21st century, the duration, accumulated heat stress and peak intensity show systematic increases under all the future emission scenarios considered. Conversely, heating rate and the priming period display systematic decreases in the tropics and subtropics. The projected changes in warm-season MHW properties are broadly consistent in global patterns among the three models. However, there are regional disagreements on future MHW properties among the models as well as between present-day model simulations and the observations. Understanding the drivers of biases in the models is important in interpreting MHW projections and responsibly employing MHW projections in the studies of their ecological impacts. In the following section, we discuss the potential drivers of present-day model biases and the implications for future warm-season MHW projections.
The large model biases during the historical period and inter-model spread in the spatial pattern of the future projections are likely caused by different model representations of atmospheric and oceanic processes including 1) cloud formation; 2) deep convection, precipitation and storms, 3) surface winds and associated oceanic heat transport, and 4) ENSO dynamics. First, cloud representation in the tropics and subtropics help determine the peak intensity of MHWs by inducing anomalous radiation balance and surface heat flux that can cause anomalously warm SSTs. For example, the negative low cloud biases off the coast of California and Peru and in the Benguela current system lead to large warm bias in HSpeak observed in GFDL-ESM4 output (Dunne et al., 2020).
Second, heavy precipitation and storms may affect MHW duration, as associated strong winds and anomalous surface heat flux can cool the sea surface and terminate a MHW. Overestimated precipitation in parts of the tropical Pacific in GFDL-ESM4 due to Double Intertropical Convergence Zone (ITCZ) problem could contribute to a shorter MHW duration than in the other two models (Danabasoglu et al., 2020; Dunne et al., 2020; Yukimoto et al., 2019). Tropical storms as well as deep convection associated with Madden Julian Oscillation (MJO) can also drive MHW dissolution; how accurately models simulate tropical storms and behaviors could therefore affect the simulation of MHW duration and the associated accumulated heat stress (Shin & Park, 2020).
Third, surface winds over the ocean influence SSTs and trigger anomalous heat stress at regional scales through affecting air-sea heat flux, wind-driven anomalous zonal advection and turbulent mixing (Bond et al., 2015; Sen Gupta et al., 2020; Holbrook et al., 2019; Oliver et al., 2017). In CESM2-WACCM, underestimated upwelling due to damped wind stress in the eastern boundary current regions could contribute to the positive biases in MHW duration off the west coasts of California, South Africa and South America (Danabasoglu et al., 2020). In GFDL-ESM4, the positive biases in the equatorial cyclonic wind stress and the negative biases in the zonal surface winds off the equator could enhance the surface heat loss at the equator and weaken surface heat loss off the equator by affecting vertical mixing of the warm surface later with cooler waters at depth (Dunne et al., 2020). This may partially explain the pattern of negative biases in MHW duration in the equatorial Pacific and more poleward positive biases. Shorter durations in the tropics in GFDL-ESM4 compared with the other models could also be driven by the shallower mixed layer in GFDL-ESM4, which would allow sea surface to warm or cool faster and stronger through air-sea heat flux. Faster and stronger warming due to the shoaled representation of the mixed layer depth might contribute to the higher peak temperatures (HSpeak) and larger rate of anomalous heat stress development (HRc) in the tropics, relative to the other models.
Fourth, the inter-model variability in the thermal properties over the tropics are also related to model disagreements on the magnitude, location and timing of ENSO-driven SST anomalies. Though the simulations of ENSO dynamics have been improved with the latest generation of GCMs included in CMIP6, many uncertainties that can affect ENSO-driven heat stress remain (Beobide-Arsuaga et al., 2021; Brown et al., 2020; Jiang et al., 2021). Large warm biases for HSpeak in the eastern tropical Pacific by GFDL-ESM4 may be related to the underestimated convection in the western equatorial Pacific and stronger thermal stratification in the Pacific cold tongue region that could drive overestimated SSTs during ENSO events (Dunne et al., 2020). In contrast, there is no warm bias across the tropical Pacific in CESM2-WACCM and MRI-ESM2, which have improved representation of the stratocumulus clouds, and ocean mixing and stratification (Danabasoglu et al., 2020; Yukimoto et al., 2019).
Another fundamental source of model bias in simulating warm-season MHWs is the spatial resolution of the atmosphere and ocean components of the models. The spatial resolution of the atmospheric module in a GCM can affect many processes, notably cloud formation. Though this factor cannot explain the differences among the models we examined, as the models employed the same resolution the atmosphere (Danabasoglu et al., 2020; Dunne et al., 2020; Yukimoto et al., 2019), it could influence the models’ performance relative to observations. In addition, the spatial resolution of the ocean module in most GCMs is not fine enough to resolve small-scale processes, like boundary currents and mesoscale eddies, which may drive underestimates of heat stress which arise from variations in these oceanic processes. For example, in coastal and boundary current regions, underestimated magnitude of mesoscale eddies could lead to a negative heat stress bias due to its effects on heat transport (Guo et al., 2022; Hayashida et al., 2020; Oliver et al., 2019; Pilo et al., 2019), as shown in the negative biases of HSpeak in these regions across all models.
Coarse-spatial resolution can also cause unrealistically smooth SST time series due to high serial autocorrelation (Oliver et al., 2019), thereby leading to overestimated duration of continuous heat stress and underestimated duration of priming. The spatial resolution of GFDL-ESM4 is finer than that of CESM2-WACCM and MRI-ESM2 (Danabasoglu et al., 2020; Dunne et al., 2020; Yukimoto et al., 2019), which may partially explain the lower magnitude of model biases in GFDL-ESM4 and smaller coefficient of the relationship between historical model bias and projected change. While the resolution in the tropics and subtropics is similar among the models, the C-grid employed by GFDL-ESM4 could represent more realistic boundary features than the B-grid employed by the other two models; for example, the equatorial undercurrent could be represented up to twice as accurately using the C-grid as opposed to the B-grid at the same spatial resolution (Dunne et al., 2020). The positive biases for the duration and negative biases for peak intensity in most of the ocean by CESM2-WACCM and MRI-ESM2 might also be related to the limited spatial resolution of their ocean couplers. The systematic cold bias in the subpolar North Atlantic is known as a common error feature in GCMs due to the poor representation of mesoscale eddies (Danabasoglu et al., 2020; Dunne et al., 2020; Yukimoto et al., 2019).
Uncertainty in projected MHW properties may be larger in the tropics where SST can be more sensitive to model ability to simulate the aforementioned driving processes. In the tropics, a small positive bias in the mean state of SST could lead to large bias in the duration and accumulated heat stress due to the limited seasonality, which is reflected in the large positive biases for Dc and Ac in the tropical Pacific by CESM2-WACCM and MRI-ESM2. This is also shown in the large positive relationship between the future model projections and present-day model bias for Dc and Ac in the warm-water coral reef cells. Given this likely amplification of model bias, the future projections of the duration and accumulated heat stress, two metrics used for exploring ecological impacts of MHWs, need to be interpreted carefully in impacts modelling and research.
Future MHW studies can focus on examining the role of above driving processes using high resolution models with a large model ensemble of SST outputs, as it may advance the predictability of MHWs. Modelling experiments could be designed to examine the role of each of the physical processes which we identified might drive the model biases in simulating warm-season MHW properties. Given the essential contribution of high spatial resolution to the accuracy of MHW projections, future research characterizing MHW properties and their ecological impacts would benefit from using outputs from GCMs with finer ocean and atmospheric grids, although employing finer resolution SST output requires greater computational resources. Future work could also incorporate outputs from more GCMs and ESMs when data is available, considering the fact that this study is restricted in examining systematic biases for warm-season MHW properties due to the limited availability of daily SST model outputs. This would not only create a more robust ensemble of future projection, it would enable a more thorough analysis of the processes that systematically drive model biases in simulating MHWs. Meanwhile, including more ensemble member projections for each model may further constrain the uncertainties in terms of internal variability (e.g., ENSO) that could influence heat stress conditions.
It should be noted that the choice of thermal threshold for defining MHWs is fundamental to computing MHW projections and their ecological impacts. Most of the MHW projection studies to date used a fixed, historical thermal threshold to define MHWs (e.g., Hobday et al., 2016). As the ocean warms, marine organisms and ecosystems may adjust to a warmer baseline via physiological acclimatization, direction selection and changes in community structure, such that heat stress calculated from historical conditions is not representative. For example, there is evidence that coral reefs exposed to frequent heat stress may acquire higher thermal resistance (Hughes et al., 2018; Morikawa et al., 2019), though it may come with reduced coral diversity and structural complexity (Donner & Carilli, 2019; Magel et al., 2019). To better examine the projections of MHWs and their ecological impacts, more studies need to incorporate the role of acclimatization and adaptation into the definition of the heat stress baseline (Logan et al., 2014; McManus et al., 2020, 2021). We conducted a simple additional analysis here by quantifying MHWs relative to a rolling MMM threshold that represents theoretical adjustment to warming over time. Though the results are intuitive, the projection of less severe MHW properties assuming the rolling thermal threshold demonstrate the high sensitivity of MHW projections to the choice of threshold. This highlights the necessity of incorporating variable thermal thresholds, based on research into acclimation and adaptation in marine organisms and ecosystems (Alsuwaiyan et al., 2021), into future MHW projection and impact research.
With continued climate change and associated ocean warming, MHWs will continue to, or even more substantially, threaten marine ecosystems and the associated cultures, fisheries and incomes of local and Indigenous peoples (Cooley et al., 2022). To best understand the impact of increasing warm-season MHWs on marine organisms and ecosystems, we need to look beyond the accumulated intensity and examine the thermal properties like duration, heating rate and priming period. We also need to consider the ability of models to describe the processes driving MHW development and dissolution, as well as the extent to which organisms and ecosystems may adjust to warming. Considering these factors, and the biases they may create in model output, is important for researchers studying the impact of MHWs on ecosystems. This cautious analysis of MHW projections is necessary to better inform policymakers and marine resource managers tasked with protecting marine life, and the people who depend on marine life, from the rising threat of MHWs.