Volume 129, Issue 2 e2023JC020330
Research Article
Open Access

Understanding Biases in Indian Ocean Seasonal SST in CMIP6 Models

Sebastian McKenna

Corresponding Author

Sebastian McKenna

Climate Change Research Centre and ARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, NSW, Australia

Correspondence to:

S. McKenna,

[email protected]

Contribution: Conceptualization, Methodology, Formal analysis, Writing - original draft, Writing - review & editing, Visualization

Search for more papers by this author
Agus Santoso

Agus Santoso

Climate Change Research Centre and ARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, NSW, Australia

Center for Southern Hemisphere Oceans Research (CSHOR), CSIRO Oceans and Atmosphere, Hobart, TAS, Australia

Contribution: Conceptualization, Methodology, Writing - review & editing, Supervision

Search for more papers by this author
Alex Sen Gupta

Alex Sen Gupta

Climate Change Research Centre and ARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, NSW, Australia

ARC Australian Centre for Excellence in Antarctic Science, University of Tasmania, Hobart, TAS, Australia

Contribution: Conceptualization, Methodology, Writing - review & editing, Supervision

Search for more papers by this author
Andréa S. Taschetto

Andréa S. Taschetto

Climate Change Research Centre and ARC Centre of Excellence for Climate Extremes, University of New South Wales, Sydney, NSW, Australia

Contribution: Conceptualization, Methodology, Writing - review & editing, Supervision

Search for more papers by this author
First published: 20 February 2024


The latest generation of climate models continue to exhibit biases in their representation of climatological sea surface temperature (SST), affecting their ability to simulate climate variability including the Indian Ocean Dipole which is typically too strong. Here, we analyze the surface layer heat budget of the Indian Ocean to diagnose the processes leading to biases in climatological SST biases in 20 Coupled Model Intercomparison Project Phase 6 (CMIP6) models, in comparison to a suite of observational and reanalysis products. In the western tropical Indian Ocean, we find that weaker than observed winds reduce the strength of surface currents leading to warm SST bias. In the south-eastern tropical Indian Ocean, overly strong southeasterly winds are associated with overestimated coastal upwelling that increases cooling across CMIP6 models. In the Arabian Sea, overly strong surface winds increase latent heat loss and leads to cool SST biases. We also analyze other regions like the Bay of Bengal, where a persistent cool bias cannot be explained by seasonal heat budgets, and southern Indian Ocean, where overly strong surface winds are responsible for cool SST biases. These biases in surface processes are supported by intermodel relationships between relevant variables, thus explaining differences in simulating the climatological SSTs across the model ensemble. Our analysis suggests that biases in atmospheric processes in particular surface winds are a primary cause of biases in Indian Ocean SST. Reducing these biases would improve the simulation of climate variability in the Indian Ocean toward more reliable climate projections and predictions.

Key Points

  • Climate models show biases in Indian Ocean (IO) sea surface temperatures (SST) affecting representation of climate variability

  • Weaker winds lead to warm SST biases in western tropical IO while stronger winds cause cool biases in Arabian Sea and southeast tropical IO

  • The seasonally varying wind driven mechanism leading to SST bias is different in each region analyzed

Plain Language Summary

Climate models still have difficulties in accurately representing sea surface temperatures (SST) and climate patterns in the Indian Ocean. By examining the surface-layer heat balance of the Indian Ocean in 20 CMIP6 climate models, a suite of observational products, and inter-model relationships, we investigate why the models simulate climatological SST that is too high or too low in different regions at different times of the year. We discovered that in the western tropical Indian Ocean, the models underestimate the strength of surface currents due to weaker winds, which leads to warmer SST than observed in some months. In the southeastern tropical Indian Ocean, the models overestimate the strength of southeasterly winds, causing stronger coastal upwelling and increased cooling, resulting in cold biases. In the Arabian Sea, the models simulate excessively strong surface winds, leading to more heat loss and cooler SST. These findings highlight the relationship between atmospheric and surface ocean processes in setting SST biases in the Indian Ocean.

1 Introduction

Sea surface temperature (SST) is an important variable for the interaction between the ocean and atmosphere, and underpins climate variability on various time scales. Correctly simulating the evolution of SST in climate models is crucial for realistic climate simulations and future projections. However, climate models exhibit biases in SST. In the Indian Ocean, which is the focus of this present study, this affects their ability to accurately simulate many local and large-scale aspects of the Indian Ocean climate and its variability which depend on coupled ocean-atmosphere processes (e.g., Bellenger et al., 2014; Li et al., 2015b; McKenna et al., 2020; Tao et al., 2016; Wang et al., 2014). The Indian Ocean undergoes unique seasonal changes in surface winds and ocean currents, thus presenting several interacting processes that could give rise to biases in SST. In this study, we focus on processes leading to biases in the seasonal cycle of the Indian Ocean SST across multiple climate models, by investigating the surface layer heat budget of the tropical Indian Ocean.

The Indian Ocean has strong interannual variability (Saji et al., 1999) and has been warming at a faster rate than the other ocean basins (e.g., Dhame et al., 2020). Changes in SST are strongly coupled to the evolution of the south Asian monsoon, Madden Julian Oscillation (MJO), Indian Ocean Dipole (IOD), and Indian Ocean basin wide warming (IOB), which have pronounced impacts on climate and weather patterns over Indian Ocean-rim countries and beyond (Saji et al., 1999; Schott & McCreary, 2001; Schott et al., 2009; Webster et al., 1999; Zhang, 2005). Under increased greenhouse warming, climate models project an increased frequency of extreme positive IOD events (Cai et al., 2014, 2021). However, climate models exhibit significant biases in their simulation of interannual variability linked to the mean state (McKenna et al., 2020) which undermine these projections (e.g., Wang et al., 2021). Here we explore the mechanisms behind biases in mean-state SST and its seasonal evolution in the sixth phase of the Coupled Model Intercomparison Project (CMIP6) models, utilizing inter-model diversity and a suite of observational and reanalysis products.

Persistent SST biases in CMIP-class climate models have been suggested to be related to incorrect representation of surface heat fluxes and surface winds over certain regions like the Arabian Sea and southern Indian Ocean (Chowdary et al., 2016; Fathrio et al., 2017; Feng et al., 2023; Li, Du, et al., 2015, Li et al., 2015a; Long et al., 2020). Many CMIP5 models exhibit a too cold Arabian Sea during boreal winter and too warm western tropical IO in boreal summer (Fathrio et al., 2017). The cold Arabian Sea bias has been linked to anomalously strong winds and air sea heat loss over the region (Fathrio et al., 2017; Levine et al., 2013; Sandeep & Ajayamohan, 2014; Sayantani et al., 2016). The tropical IO, particularly the IOD-related western pole region, also presents strong biases in CMIP5 and CMIP6 (Cai et al., 2014; Fathrio et al., 2017; Long et al., 2020). There is a systematic easterly wind bias along the equator that has persisted with minimal improvement from CMIP5 to CMIP6 (Lee et al., 2013; Long et al., 2020). The annual mean pattern of SST bias in the tropical Indian Ocean is “positive IOD-like,” where the west IO tends to be warmer than observed and the east cooler than observed along with an easterly wind and associated precipitation bias (Long et al., 2020). Many of the SST biases in the Indian Ocean described above have persisted in CMIP6 (e.g., Bhattacharya et al., 2022).

The biases in both SST and winds in CMIP models have previously been linked to biases in the monsoons. CMIP5 models, which tend to have an overly weak southwest monsoon over the Arabian Sea, tend to have warm equatorial SST biases in the west IO, overly strong easterly equatorial winds, and a strong zonal SST gradient bias which also reduces southeast Asia precipitation (Annamalai et al., 2017; Li et al., 2015a). Overestimated easterly winds over the equatorial Indian Ocean increase upwelling, leading to overly cool SST (Chowdary et al., 2016). SST and surface wind biases associated with the South Asian monsoon typically weaken rainfall in CMIP3 and CMIP5 models. For example, the east African short rains are reduced in October to December in CMIP models (Hirons & Turner, 2018).

The biases in SST and winds are exacerbated by coupling, leading to further SST, thermocline depth, and precipitation biases (Annamalai et al., 2017). A comparison of CMIP6 and Atmospheric Model Intercomparison Project (AMIP) simulations found biases in SST primarily stem from wind and monsoon biases in the atmospheric simulation and to a lesser extent coupling errors (Long et al., 2020). Coupling between the ocean and atmosphere in models have been shown to lead to larger SST bias in the Indian Ocean compared to ocean models forced with observed surface fluxes (e.g., Coordinated Ocean-ice Reference Experiment phase II—CORE II simulations, Rahaman et al., 2020). The CORE II simulations also exhibit overly deep mixed layers leading to cool SST biases (Rahaman et al., 2020), indicating that some of the SST biases are also related to oceanic processes, such as ocean mixing.

Utilizing the AMIP and Ocean Model Intercomparison Project (OMIP) components of CMIP6 models, and targeted ocean-model experiments, Feng et al. (2023) found that temperature biases in the mixed layer of the CMIP6 coupled models mainly originate from the atmospheric component, with a small portion from coupling, consistent with the conclusion of Long et al. (2020) who examined IOD-like biases in a larger ensemble of CMIP models.

Here we take a different approach and examine in detail the heat flux components that give rise to SST biases (using a surface heat budget). We also investigate the factors that lead to intermodel differences to establish mechanistic links to the origin of the biases across a large ensemble of CMIP models. Using intermodel relationships between heat budget variables and physical processes to attribute heating or cooling to SST biases has not been conducted previously, and would allow a systematic evaluation of the mechanisms that give rise to SST biases. We find that SST biases in the Indian Ocean are primarily caused by biases in surface winds and discuss the processes by which they impact the surface layer. Section 2 covers details of the models and reanalysis data, and the implemented approach. Section 3 presents the main results, followed by a discussion and conclusion in Section 8. By analyzing an ensemble of CMIP6 models, this study provides insights into the consistency of biases in SST and processes across models, which can help the modeling community improve aspects of the models.

2 Methods and Data

Twenty CMIP6 climate models and seven observational/reanalysis products are used to understand processes underpinning sea surface temperature seasonal evolution in the tropical Indian Ocean and the associated model biases. One realization (r1i1p1f1) from the historical simulation (1950–2018) of each CMIP6 model is analyzed (see Table 1). These CMIP6 models have been chosen based on availability of the required fields used for heat budget calculations. Variables for each model have been horizontally re-gridded onto a 1° × 1° uniform grid using a bilinear interpolation.

Table 1. List of Models Used in This Study
Model Atmosphere (°lon × °lat, #vertical level) Ocean (°lon × °lat, #vertical level) Reference
ACCESS-CM2 N96 (192° × 144°, 85) Tripolar (360° × 300°, 50) Bi et al. (2019)
ACCESS-ESM1.5 N96 (192° × 145°, 38) Tripolar (360° × 300°, 50) Ziehn et al. (2019)
CAMS-CSM1-0 T106 (320° × 160°, 31) Tripolar (360° × 200°, 50) Rong (2019)
CanESM5 T63 (128° × 64°, 49) Tripolar (361° × 290°, 45) Swart et al. (2019)
CESM2 (288° × 192°, 32) (320° × 384°, 60) Danabasoglu (2019a)
CESM2-FV2 (144° × 96°, 32) (320° × 384°, 60) Danabasoglu (2019b)
CESM2-WACM (288° × 192°, 70) (320° × 384°, 60) Danabasoglu (2019c)
CESM2-WACM-FV2 (144° × 96°, 70) (320° × 384°, 60) Danabasoglu (2019d)
CMCC-CM2-HR4 (288° × 192°, 26) (1442° × 1051°, 50) Scoccimarro et al. (2020)
EC-Earth3 (512° × 256°, 91) (362° × 292°, 75) EC-Earth (2019a)
EC-Earth3-CC (512° × 256°, 91) (362° × 292°, 75) EC-Earth (2021)
EC-Earth3-Veg (512° × 256°, 91) (362° × 292°, 75) EC-Earth (2019b)
EC-Earth3-Veg-LR (320° × 160°, 62) (362° × 292°, 75) EC-Earth (2020a)
EC-Earth3-AerChem (512° × 256°, 91) (362° × 292°, 75) EC-Earth (2020b)
FGOALS-f3-L C96 (360° × 180°, 32) Tripolar (360° × 218°, 30) Yu (2019)
FGOALS-g3 (180° × 80°, 26) Tripolar (360° × 218°, 30) Li (2019)
MPI-ESM-1-2-HAM T63 (192° × 96°, 47) (256° × 220°, 40) Neubauer et al. (2019)
MPI-ESM1-2-HR T127 (384° × 192°, 95) (802° × 404°, 40) Jungclaus et al. (2019)
MPI-ESM1-2-LR T63 (192° × 96°, 47) (256° × 220°, 40) Wieners et al. (2019)
MRI-ESM2-0 TL159 (320° × 160°, 80) (360° × 364°, 61) Yukimoto et al. (2019)

We examine the heat budget of the top 50 m (T50) in each model and reanalysis/observations where available. The heat budget terms have been explicitly calculated at each grid point and model level down to 50 m, then depth averaged weighted by each vertical layer. This approach has been used in previous studies to analyze SST, model biases, and associated processes in the upper ocean (e.g., de Boyer Montégut et al., 2007; Fathrio et al., 2017; Foltz, 2003; Foltz et al., 2010; Ng et al., 2015; Santoso et al., 2010). SST and temperature at 50 m are closely related and show similar bias when compared to corresponding observations (Figure S1 in Supporting Information S1), making the upper 50 m layer a good proxy for understanding the processes that affect SST bias. We note that this approach is less physically motivated than a time and space varying mixed layer depth, but it facilitates model intercomparison for identifying biases in the surface processes across models.

The temperature tendency of the surface layer is examined and broken down into the net surface heat flux, heat fluxes due to horizontal and vertical advection, and other unresolved processes that are included in a residual term, according to the following equation
T t = Q net ρ c h u T x + v T y + w T z + Residual , $\frac{\partial T}{\partial t}=\frac{\mathrm{Q}\text{net}}{\rho ch}-\left(u\frac{\partial T}{\partial x}+v\frac{\partial T}{\partial y}+w\frac{\partial T}{\partial z}\right)+\text{Residual},$ (1)
where T denotes potential temperature, u is zonal current, v is meridional current, and w is vertical current. The heat budget terms are calculated at each grid cell and then averaged over the top 50 m. Where w is not available as an archived output, it has been calculated using mass continuity (only ACCESS-CM2 and ACCESS-ESM1.5). Qnet is the net heat flux term at the surface and has been calculated as: net shortwave + net longwave − latent heat − sensible heat. Some CMIP6 models archive additional heat fluxes to the ocean associated with runoff, rainfall, and evaporation, however these terms are generally small compared to other fluxes and are not included in Qnet calculations. The budget depth h is fixed at 50 m, ocean density, ρ, is treated as constant (1,026 kg/m3), and specific heat capacity of seawater, c, is set to 3986 J kg−1 K−1. Where model vertical grid cell boundaries are not at exactly 50 m, the values have been linearly interpolated to 50 m. The heat budget terms have been calculated for each month of the historical timeseries, after which the long-term mean monthly climatology is calculated.

The residual term in Equation 1 includes processes that cannot be explicitly calculated using the monthly output, including horizontal/vertical mixing, entrainment, sub-monthly transient eddy heat fluxes, and shortwave that exits from the bottom of the 50 m layer. This contributes to a residual term that is relatively large and of similar order of magnitude to advection terms.

Monthly SST bias in CMIP6 models is estimated in comparison to NOAA OISSTv2 (Reynolds et al., 2002). Gridded ARGO data (Roemmich & Gilson, 2009) are also used to evaluate subsurface temperature. Simulated Qnet and component fluxes (shortwave, longwave, sensible heat, latent heat) are compared to ERA5 reanalysis. Previous work suggests that ERA5 provides the best estimate of Qnet in the tropical Indian Ocean compared to other gridded products (Hersbach et al., 2020; Pokhrel et al., 2020). Specifically, ERA5 heat fluxes were more similar to observations from mooring arrays and buoys in 3 different locations around the IO compared to other reanalysis and blended observations. For completeness and closure of ocean reanalysis heat budgets, we present Qnet from ocean reanalysis products, and OAFlux as a gridded observational estimate, derived using a synthesis of satellite and Numerical Weather Prediction model output (Yu & Jin, 2014). Fields down to 50 m from three ocean reanalysis products (GODAS, SODA3.4.2, ORAS5) are analyzed for heat budget components like heat advection (Behringer et al., 1998; Carton et al., 2018; Zuo et al., 2019). Lack of subsurface circulation observations make it hard to produce reliable observational estimates of advection driven heat fluxes. We therefore compare CMIP6 model heat advection with ocean reanalysis and examine the differences and inter-model relationships. An important finding from our evaluation is that there is often a large spread in reanalysis estimates, sometimes exceeding the spread across CMIP models, severely hampering the estimate of subsurface temperature and heat flux biases in the models.

We examine intermodel relationships between subsurface heat flux terms and various surface and subsurface state variables to gain insight into the causes of intermodel differences and model biases. Our model inter-comparison approach differs from some of the previous studies (e.g., Fathrio et al., 2017), as we compare the heat budget of models against observational or reanalysis data sets and analyze intermodel relationships in relation to the observations and reanalysis. We can expect that when the intermodel relationships/correlations between surface variables and heat flux terms are strong and statistically significant, a bias in the surface variable potentially contributes to generating SST biases. Table 2 summarizes the observational and reanalysis products used.

Table 2. Observational and Reanalysis Products Used
Name Variables Reference
NOAA OISSTv2 SST Reynolds et al. (2002)
ARGO SST, θ Roemmich and Gilson (2009)
ERA5 Qnet, shortwave, longwave, sensible, latent, τu, τv Hersbach et al. (2020)
OAFlux Qnet Yu and Jin (2014)
GODAS θ, U, V, W, τu, τv, Qnet Behringer et al. (1998)
SODA3.4.2 θ, U, V, W, τu, τv, Qnet Carton et al. (2018)
ORAS5 θ, U, V, W, τu, τv, Qnet Zuo et al. (2019)

3 Results

There are large systematic regional IO SST biases across the CMIP6 ensemble that vary strongly with season (Figure 1). Similar to CMIP5 models, CMIP6 exhibits a pattern of SST bias, such as the Northern Indian Ocean's prominent cold bias from February to April and the Arabian Sea's cool bias peaking from January to June and reemerging by December (Fathrio et al., 2017; McKenna et al., 2020; C. Wang et al., 2014) The western tropical Indian Ocean has a distinct warm bias present from June to October. The southeast tropical Indian Ocean has a weak but significant warm bias off the coast of Java in December. From August to November there is a cool bias in the region extending into the south Indian Ocean. There is also a consistent cool bias in the Bay of Bengal all year round, which is an area of importance to the South Asian Monsoon.

Details are in the caption following the image

Multi-model mean SST bias of CMIP6 ensemble compared to NOAA OISSTv2 by month. Stippling is where at least 15/20 models agree on sign (significant at 95% in binomial test). Regions of interest are outlined in boxes on the top left panel.

While CMIP6 models exhibit SST biases, they effectively replicate the seasonal east-west SST gradient variation (Figure 2). CMIP6 also shows some systematic biases in surface currents, which advect warm and cool water around the Indian Ocean and lead to temperature biases. Coastal surface currents like the Somali Current show particularly large biases in CMIP6; in JJA the northward Somali current is too strong (Figure 2). Additionally, near equator surface flow is typically weaker than reanalysis for most of the year, especially in the western Indian Ocean. These near equator biases in surface current could be related to biases in surface wind.

Details are in the caption following the image

CMIP6 multimodel median SST and surface currents by season (left column), and (right column) CMIP6 surface current bias relative to reanalysis median (SODA3.4.2, ORAS5, GODAS). Surface currents are depth averaged down to 50 m. Colormap on the right hand column indicates the relative surface current speed bias.

Our analysis focuses on four key regions with significant SST and surface current biases—WTIO, SETIO, Arabian Sea, and Bay of Bengal—as these areas, particularly critical for the IOD, consistently show larger SST anomalies (Figure 1); particularly given that the simulation of the IOD is often too strong in CMIP5 and CMIP6 models (Cai & Cowan, 2013; McKenna et al., 2020). The Arabian Sea is also an important area affecting the formation of the IOD (Sayantani et al., 2014) and the South Asia monsoon. Large temperature variations in this region associated with upwelling variability have broader climate impacts (Izumo et al., 2008). This region is known to exhibit large cold SST bias in climate models (Annamalai et al., 2017; Fathrio et al., 2017). We also briefly explore the Bay of Bengal region, as it shows systematic cold SST bias that can affect monsoon simulation, the cool bias in the south Indian Ocean, and a sensitivity analysis off the Somali Coast. In the following sections we examine seasonal variations in the surface layer heat budget in the CMIP6 models compared to observations and reanalysis for the four regions. Inter-model relationships are examined to identify processes that give rise to inter-model spread, and to further understand surface temperature biases in the CMIP6 models.

3.1 WTIO

Figure 3 depicts the WTIO region's seasonal surface temperature and heat budget evolution in CMIP6 models. Satellite observations from NOAA OISSTv2 show that SST in the WTIO peaks in April, with a secondary peak in November (Figure 3a), with minima in August, and to a smaller extent, in January. The biannual nature of SST is related to the seasonal cycle in incoming solar radiation in the region, localized effects of peak latent heat loss in June, July, August (JJA) and upwelling related cooling (Schott et al., 2009). In situ observations from ARGO show similar temperature evolution, as do the three ocean reanalysis products (Figure 3a). The CMIP6 models simulate a broadly similar seasonal cycle, but with important systematic biases in some months. In particular, the MMM is too cold in March-April by around 0.5oC (compared to NOAA OISSTv2 or ARGO), and too warm in June-August by around 0.5oC. Consequently, the amplitude of the SST seasonal cycle is systematically smaller in CMIP6 models compared to observations. A similar bias was noted for CMIP5, albeit for a slightly different region (Fathrio et al., 2017).

Details are in the caption following the image

Monthly climatological time series in WTIO region of (a) SST for CMIP6 multi-model median and interquartile range (blue) and NOAA OISSTv2 (purple), ARGO [surface only] (olive), (b) temperature tendency of CMIP6 median and interquartile range (blue), NOAA OISSTv2 (purple), and GODAS (orange) ORAS5 (red), SODA3.4.2 (green), ARGO [depth averaged](olive) (c) net heat flux ERA5 (brown), OAFLUX (pink), (d) zonal advection, (e) meridional advection, (f) vertical advection, and (g) the residual.

The temperature tendency (Figure 3b) describes the amount of heating or cooling occurring from 1 month to the next in the upper 50 m of the ocean. ARGO observations indicate WTIO warming from January to April, peaking in February, and subsequent cooling through May to July, with maximum cooling in June. Warming occurs again in August to October followed by cooling in November and December. This pattern is broadly seen in the reanalysis products, although some large differences occur in November to February when reanalysis products show reduced cooling and heating compared to ARGO in November/December and January/February respectively. The temperature tendency in CMIP6 MMM follows a similar seasonal cycle, but with a weaker magnitude and a lag of about 1 month in the first half of the year (Figure 3b). Overall, the CMIP6 models simulate insufficient heating in January to March (approx. 10–20 W/m2, Figure 3b), consistent with the cool SST bias over March-April (Figure 3a). The rapid cooling in May and June seen in the observational products is also notably weaker in the CMIP6 models by around 20–30 W/m2, consistent with the warm bias in JJA SST (Figures 3a and 3b). In July and August, there is more cooling in CMIP6, contributing to the reduction in SST biases. For the rest of the year, the differences between the CMIP6 models and observational products are not systematic (i.e., the observed value lies well within the model spread).

The net air-sea heat flux (Qnet) has similar overall seasonal variability to temperature tendency but with a distinct positive offset, providing a nearly year-round warming (Figure 3c). There is a large offset between Qnet and temperature tendency in reanalysis and CMIP6 (Figures 3b and 3c), compensated by a combination of ocean advection and unresolved processes. Ocean reanalysis products (GODAS, SODA, ORAS5) tend to show less Qnet heating compared to ERA5/OAflux data, with the largest differences in June to December (∼10–20 W/m2). The CMIP6 models have systematically too little Qnet heating throughout the year compared to ERA5, and an even larger bias compared to OAFlux observations (Figure 3c). Notably, CMIP6 Qnet is closer, though still weaker, compared to ocean reanalysis; however, in May to September GODAS indicates more cooling and in May to June ORAS5 shows more cooling than CMIP6. The WTIO's net air-sea flux significantly differs from ERA5 (Figure 4).

Details are in the caption following the image

Qnet and components for CMIP6 MMM in solid lines, ERA5 in dashed. Orange: shortwave radiation, Purple: Qnet, Green: latent heat flux, blue: longwave radiation, red: sensible heat flux. (a) WTIO, (b) SETIO, (c) Arabian Sea.

Decomposing Qnet into its components shows that the discrepancy between CMIP6 and ERA5 Qnet is primarily related to excessive latent heat loss in CMIP6 (Figure 4a). Latent heat is primarily controlled by wind speed and surface moisture, suggesting biases in one or both of these factors are important for understanding the errors in air-sea heat fluxes. The air temperature difference can also be important for latent heat fluxes in the Indian Ocean (Du & Xie, 2008), but wind and humidity are the dominant drivers here.

Latent heat bias between CMIP6 and reanalysis is greatest between July and April with a significant inter-model correlation between latent heat flux and Qnet (0.52, p = 0.02) (Figure 5a) indicating that models with stronger latent heat loss tend to have lower Qnet. In May and June there is still a significant relationship, but the level of bias is reduced (Figure S4a in Supporting Information S1). The bias in latent heat flux can be explained by biases in wind for December to April, where inter-model correlation between latent heat loss and wind stress (r = −0.53, p = 0.02) suggests that the stronger-than-observed wind speed in the CMIP6 models is an important factor in the overly strong latent heat loss (Figure 5b). On the other hand, a lack of inter-model relationship in July to November suggests that wind biases are unlikely to explain latent heat flux bias for the region during this earlier part of the period (Figure S4b in Supporting Information S1), suggesting that surface moisture biases may be important.

Details are in the caption following the image

Scatter plots in WTIO region of (a) Latent heat flux and Qnet in July-April average, (b) surface wind stress and latent heat flux in December to April average, (c) zonal heat advection and temperature tendency in May-June average, and (d) surface wind stress and zonal heat advection in May-June average. Regression slope, Pearson correlation coefficient and p value shown to 2 significant figures. Observations/reanalysis shown in colored squares. Observed/reanalysis data only available for one field is plotted as vertical or horizontal lines.

The shortwave radiation term in October to February is too low compared to ERA5 and contributes to the low Qnet bias in these months (Figure S4c in Supporting Information S1). Similar to the analysis on latent heat above, intermodel correlation of shortwave radiation and Qnet in October to February shows that models with reduced shortwave radiation also have lower Qnet than ERA5. There is no obvious explanation for the reduction in shortwave radiation, but biases in deep convection and clouds may be important contributors.

We further analyze Qnet to explain the negative bias in temperature tendency in January-March. The inter-model correlation between these variables is strong (r = 0.91, p < 0.01, Figure S4d in Supporting Information S1), indicating that models with reduced heating by Qnet in January to March typically have reduced overall heating. Qnet differences are linked to amplified latent heat loss and diminished shortwave radiation (Figures 5a and 5b; Figure S4e in Supporting Information S1). Excessive simulated windspeed in January to March (Figure S3 in Supporting Information S1) exacerbates the latent heat bias, while weakened shortwave radiation contributes to diminished overall heating, leading to the observed cool SST bias in February–April.

For heat advection terms, we are unable to compare to direct observations, but compare against reanalysis (GODAS, SODA3.4.2, and ORAS5). Some clear differences in the evolution of heat advection are evident. For zonal heat advection, all reanalysis products exhibit warming in January-March (JFM) peaking at around 20 W/m2 (Figure 3d), followed by cooling that peaks in June but with large inter-reanalysis spread (−25 to −50 W/m2) followed by weak cooling until November.

The CMIP6 models typically underestimate the magnitude of cooling by zonal heat advection compared to reanalysis (Figure 3d). There is a large model spread in the amount of zonal advective heating (January to April) and cooling (June to August) (Figure 3d), and typically a weaker seasonal range, with the peak cooling delayed to July rather than June in the reanalysis (Figure 3d). The reduction in cooling from zonal heat advection in May-June in CMIP6 relative to reanalysis products is responsible for much of the overall cooling bias in CMIP6 in May-June. The inter-model correlation of zonal heat advection and temperature tendency in May-June is 0.83 (p < 0.05) (Figure 5c). This shows that models with overly weak zonal heat advection tend to have overly weak cooling.

CMIP6's weak zonal heat advection in May and June is linked to weaker-than-observed southwesterly winds in the WTIO (Figures S2 and 3 in Supporting Information S1), associated with an overly weak South Asian monsoon simulation (Li et al., 2015a; Long et al., 2020). This leads to delayed and reduced zonal cooling in CMIP6 from June to August compared to reanalysis, where heat typically exits the WTIO via eastward currents in May-June. The multi-model median reflects weaker winds and zonal heat advection than ERA5 and ocean reanalysis (Figure 5d), with a significant positive correlation (r = 0.68, p < 0.01) between wind stress and zonal flow (Figure S4e in Supporting Information S1). Reduced eastward surface transport in the WTIO during May-June, a result of the systematic wind bias, leads to insufficient zonal advective cooling, thereby causing a warm SST bias.

The meridional heat advection is positive throughout the year, with the largest warming effect from June to September in reanalysis (Figure 3e). The JJA peak in meridional warming may seem counter-intuitive at first, as there are broadscale southwards currents in the region which would lead to cooling north of the equator. However, south of the equator southwards currents lead to warming as they transport warm equatorial water south (Figure 2). In addition, the northward flow of the Somali Current is very strong and transports warm equatorial water north. Within the boxed region (Figure 1), the spatial averaging is outweighed by the meridional advection that warms the region.

CMIP6 simulates a broadly comparable seasonal cycle in meridional heat advection, with a slight delay in the peak (in August vs. July with an inter-quartile range of 25–60 W/m2; Figure 3e). However, the stronger and delayed August peak in meridional heat advection is unlikely to play an important role for the underestimated cooling in CMIP6 (Figures 3b and e), as the temperature tendency bias occurs in May-June.

Vertical heat advection is generally negative in the WTIO across the reanalysis, with two major cooling peaks in the year: first in April/May peaking at around −20 W/m2 in reanalysis, then again in November peaking at −30 W/m2 (Figure 3f). During the monsoon periods (January/February, and JJA), the vertical heat advection is small. The vertical heat advection term in CMIP6 MMM tends to be broadly similar to the reanalysis, although model spread is substantial (Figure 3f).

The residual term (Figure 3g) shows cooling throughout the year in most reanalysis except for a weak warming in JJA in GODAS. The residual appears to be the most important term for moderating Qnet which is positive most of the year. As Qnet decreases in JJA, so does the magnitude of the residual. In CMIP6, the consistently negative residual year-round implies ongoing cooling processes, due to penetrative shortwave radiation at the base of the surface layer, entrainment, and vertical mixing of colder water when the mixed layer depth exceeds 50 m.

The residual term in CMIP6 MMM shows large differences from reanalysis in the October to March period, where the residual is typically smaller than reanalysis products. The differences in the residual likely stem from the biases that were previously mentioned—zonal advective cooling and Qnet biases. The bias in the residual may also be due to vertical temperature gradient or wind speed biases that impact the amount of mixing.


The SETIO region is important for the genesis of IOD events and exchange with the Pacific Ocean. The region has a cool SST bias from July to November (Figure 1) in the CMIP6 MMM, and a southeasterly wind bias from June to December (Figure 6a; Figure S2 in Supporting Information S1). Observational and reanalysis data show the SETIO region at its warmest in April/May and coolest in September, a pattern also seen in CMIP6 MMM, though exaggerated by lower temperatures from July to November (Figure 6a).

Details are in the caption following the image

Monthly climatological time series in SETIO region of (a) SST for CMIP6 multimodel median and interquartile range (blue) and NOAA OISSTv2 (purple), ARGO [surface only] (olive), (b) temperature tendency of CMIP6 median and interquartile range (blue), NOAA OISSTv2 (purple), and GODAS (orange) ORAS5 (red), SODA3.4.2 (green), ARGO [depth averaged](olive) (c) net heat flux ERA5 (brown), OAFLUX (pink), (d) zonal advection, (e) meridional advection, (f) vertical advection, and (g) the residual.

Despite the qualitative similarity in temperature tendencies across all models and observations, CMIP6 MMM shows an exaggerated seasonal cycle, marked by excessive cooling in July-August and warming in November-December, aligning with its SST bias (Figure 6b). Changes in temperature tendency generally mirror Qnet variations, but Qnet's positive offset leads to a net annual warming from surface fluxes. Unlike the temperature tendency, there is a large spread in Qnet across observational/reanalysis products (Figure 6c). Most reanalysis/observations show that there is heating between September and May, with modest cooling in JJA. The CMIP6 MMM/inter-quartile range of Qnet is lower than both ERA5 and OAflux for every month except November when it is higher than ERA5 (Figure 6c). CMIP6 models show a systematic negative bias in Qnet across most of the year compared to ERA5 and OAflux with MMM biases of more than −20 W/m2 during JJA.

Decomposing Qnet we find that latent heat loss tends to be overly strong throughout the year and shortwave radiation appears to be overly strong in October to March, but overly weak in May to August compared to ERA5 data (Figure 4b). Excess latent heat loss in CMIP6 is an important factor in explaining the weak Qnet compared to ERA5 for most of the year except February, and July. There is a significant inter-model correlation between latent heat flux and Qnet for each month except February and July—the average correlation over all months is 0.56 (p < 0.05; Figure S5a in Supporting Information S1). The overly strong CMIP6 southeasterly wind bias seen in June-July can help explain the excessive latent heat loss across models. Latent heat loss can be strongly modulated by surface wind strength. In the SETIO, differences in wind strength across models are primarily related to meridional winds in June-July indicated by a strong intermodel correlation between surface wind stress and meridional wind stress (r = 0.92, p < 0.01, Figure S5b in Supporting Information S1). Indeed, we find a moderate intermodel relationship between meridional wind and latent heat flux in June–July (r = −0.42, p = 0.07, Figure S5c in Supporting Information S1), that is, stronger southerly winds typically lead to stronger latent heat loss. However, over the rest of the year this inter-model relationship, and the relationship with zonal wind is not significant indicating that moisture/humidity may be a more important factor in latent heat loss.

Despite a systematic bias in CMIP6 Qnet, driven by overly strong winds compared to ERA5 in June-July, overall cooling in June-September across models cannot be explained by this term. The inter-model correlation between Qnet and temperature tendency in June to September (r = −0.40, p = 0.08, Figure S5d in Supporting Information S1) is negative and indicates that models with overly strong cooling from air sea heat fluxes tend to correspond with more realistic (lower absolute) cooling tendency. This negative correlation suggests that in SETIO, Qnet does not drive inter-model variations in the cooling tendency, suggesting that ocean processes play an important role.

Heat advection in the SETIO region has a unique seasonal cycle. Zonal advection of heat in reanalysis is generally weakly negative with an average cooling of around −10 W/m2 in reanalysis (Figure 6d). There is slightly greater cooling from July to September which then reduces in October. From June to October there are southeasterly winds in the region likely driving some offshore transport. However, in the models this southeasterly wind is stronger than observed (Figures S2 and S3 in Supporting Information S1), which likely leads to an enhanced offshore transport. In CMIP6, zonal advective cooling has a similar seasonality to reanalysis (Figure 6d) but with a large inter-model range.

Meridional advection warms the SETIO from March to December but has little effect in boreal winter (Figure 6e). CMIP6 meridional advection has comparable magnitude and seasonal cycle to reanalysis (Figure 6e). In June however, when the meridional advection peaks in CMIP6, it is stronger relative to reanalysis (GODAS and ORAS5). This may be offsetting the large CMIP6 Qnet bias in June (Figure 6b, c,e).

Vertical advection in the SETIO region generally has a cooling effect throughout the year, although it has little effect or slight warming in April–May in reanalysis (Figure 6f). GODAS exhibits large cooling in July–September (∼−20 W/m2) relative to the other two products. The seasonal cycle of vertical heat advection in CMIP6 is similar to GODAS, indicating an overly strong cooling in July–October compared to the reanalysis ensemble average. This excessive cooling by vertical heat advection is an important contributor to the cooling tendency in July–October, as well as in explaining inter-model variations. The inter-model correlation for July–October (r = 0.48, p < 0.05, Figure 7a) indicates that models with stronger cooling by vertical heat advection tend to have stronger overall cooling. The inter-model correlation is similar when excluding parts of the Indonesian seas that are included in the SETIO region indicating that the upwelling is concentrated to the west of Indonesia and offshore (Figure S6 in Supporting Information S1).

Details are in the caption following the image

Scatter plots in SETIO region of (a) vertical heat advection and temperature tendency in July to October average, (b) zonal current and vertical advection in July to October, (c) zonal current and zonal wind stress in July to October, and (d) zonal wind and vertical heat advection in July to October. Regression slope, Pearson correlation coefficient and p value shown to 2 significant figures. Observations/reanalysis shown in colored squares. Observed/reanalysis data only available for one field is plotted as vertical or horizontal lines.

These differences in vertical heat advection compared to reanalysis products may relate to differences in offshore transports and upwelling associated with biases in surface wind stress. In June to December the SETIO region has prevailing southeasterly winds (Figure S3 in Supporting Information S1). The CMIP6 MMM and GODAS, surface wind stress in July to October is stronger than observed (Figure S2 and S3 in Supporting Information S1). The stronger than observed southeasterly winds off the Indonesian coast enhance zonal transports away from the coast and coastal upwelling. This is indicated by CMIP6 offshore westward current bias in JJASON (Figure 2). A significant inter-model relationship (r = 0.60, p < 0.01, Figure 7b) between zonal current and vertical heat advection in July to October shows that stronger westward flow in models leads to more intense cooling by vertical heat advection, also indicating increased upwelling associated with offshore transports. The enhanced surface zonal water transports are related to biases in surface wind stress across models. Given the SETIO region encompasses area close to the equator, the contribution of Ekman driven flow is likely small for much of the box, so we instead analyze how the surface wind stress affects currents in the same direction.

The regional inter-model relationship between zonal wind and zonal current in July to September is strong (r = 0.82, p < 0.01, Figure 7c), such that models with stronger easterly winds tend to have stronger westward currents in CMIP6 models. Combining these relationships, there is also a relationship between zonal wind strength and vertical heat advection (r = 0.50, p < 0.05 Figure 7d) in July to September. This suggests that the excessive cooling associated with overly strong vertical heat advection, results from an enhanced zonal transport, away from Indonesia that intensifies coastal upwelling.

The residual in the SETIO region is generally negative and varies considerably across the reanalysis products (Figure 6g). This may be due to differences in vertical mixing across products, with GODAS indicating a positive offset in July-September lining up with the anomalous peak in vertical heat advection.

3.3 Arabian Sea

Exhibiting a biannual SST cycle, the Arabian Sea reaches peak temperatures in May and October, with the lowest in February and a secondary minimum in August (Figure 8a). However, CMIP6 models show a systematic underestimation of temperatures, particularly from January to May, with biases up to 1.5°C (Figure 8a). The biennial seasonality is again largely related to strong seasonal changes in insolation, and processes related to the monsoons (Behera et al., 2000). The CMIP6 models are systematically too cold over much of the year with the largest cold bias from January to May (MMM too cold by up to 1.5oC, Figure 8a). In June to October, CMIP6 SSTs closely follow observations but begin to diverge in November with weak cold biases to the end of the year. Similar biases in the Arabian Sea have been reported in CMIP5 and related to the simulation of the southwest monsoon (Annamalai et al., 2017; Fathrio et al., 2017).

Details are in the caption following the image

Monthly climatological time series in Arabian Sea region of (a) SST for CMIP6 multimodel median and interquartile range (blue) and NOAA OISSTv2 (purple), ARGO [surface only] (olive), (b) temperature tendency of CMIP6 median and interquartile range (blue), NOAA OISSTv2 (purple), and GODAS (orange) ORAS5 (red), SODA3.4.2 (green), ARGO [depth averaged](olive) (c) net heat flux ERA5 (brown), OAFLUX (pink), (d) zonal advection, (e) meridional advection, (f) vertical advection, and (g) the residual.

In the Arabian Sea, ARGO and reanalysis data show two heating peaks in April and October (around 90 W/m2) and cooling peaks in December and July (−70 to −100 W/m2). CMIP6 models, however, underrepresent this tendency from October to April (by about 10 W/m2) and overcompensate from May to July (by 20 W/m2), leading to small SST bias in the subsequent months (Figures 8a and 8b).

ERA5 data show that Qnet primarily drives temperature changes from September to May in the Arabian Sea (Figure 8c), but its higher values in JJA suggest additional oceanic cooling mechanisms. CMIP6 models exhibit clear Qnet biases throughout the year (Figure 8c). Between November and March, CMIP6 MMM Qnet is up to 50 W/m2 smaller than ERA5, indicating one or more components of Qnet are misrepresented in these months (Figure 8c). In April-June, CMIP6 MMM Qnet has a warm bias around 20 W/m2 larger than ERA5. In November to March Qnet is an important variable in driving differences in the strength of temperature tendency, shown by the intermodel correlation of r = 0.90 (Figure 9a). This indicates that models with more negative Qnet tend to have overly strong cooling in November to March.

Details are in the caption following the image

Scatter plots in Arabian Sea for (a) Qnet and temperature tendency in November to March, (b) latent heat flux and Qnet in November to March, (c) surface wind stress and latent heat flux in November to March, (d) shortwave radiation and Qnet in April to June. Regression slope, Pearson correlation coefficient and p value shown to 2 significant figures. Observations/reanalysis shown in colored squares. Observed/reanalysis data only available for one field is plotted as vertical or horizontal lines.

An analysis of net heat flux components reveals excessive latent heat loss in CMIP6 from November to March compared to ERA5, aligning with Qnet biases. The importance of latent heat is supported by a stong intermodel correlation between Qnet and latent heat terms (r = 0.78, p < 0.05, Figure 9b). The overestimated latent heat loss is related to overly strong northeasterly winds in November to March given the strong negative correlation between these variables (r = −0.68, p < 0.05, Figure 9c). In the Arabian Sea the CMIP6 MMM has an overly northeasterly wind bias from November to March compared to ERA5. This would result in excessive latent heat loss.

From April to June, the Qnet bias reverses and there is overly strong heating from air-sea fluxes in the models leading to a too strong temperature tendency (Figure S7a in Supporting Information S1). During this period, both latent heat and shortwave radiation are exaggerated across models (Figure 9d). In April to June, models with less latent heat loss have a stronger net heat flux compared to ERA5 (Figure S7b in Supporting Information S1). There is no significant relationship between latent heat and surface wind in April-June. Shortwave radiation in May and June is too strong compared to ERA5. There is an intermodel correlation between shortwave and Qnet (r = 0.81, p < 0.01, Figure S7c in Supporting Information S1) that indicates that models with higher than observed shortwave tend to have higher than observed Qnet. The combined effect of stronger shortwave radiation and weaker latent heat flux in May and June leads to stronger Qnet, which reduces the cool bias in boreal summer. While in principle the weaker latent heat loss could lead to less cloud cover, there is no significant inter-model relationship between latent heat and shortwave radiation (r = 0.16, p = 0.51), suggesting that other factors are important in driving the bias in shortwave radiation in May to June. Changes in local convection are also unlikely to explain the reduced cloud cover as SST is biased too warm during this period, suggesting that large scale circulation biases may be important.

Between June and October, zonal and vertical heat advection have their largest contributions in observations and reanalysis (Figure 8). Throughout the year zonal heat advection is negative, indicating cooling through east/west flow in this region. From June to September, the Arabian Sea experiences cooling, which relates to the summer monsoon and an increase in eastward flow associated with southerly wind-driven Ekman transport. The CMIP6 MMM exhibits a cooling from zonal heat advection that is consistent with the reanalysis, but there is a large inter-model spread in the strength of the cooling (Figure 8d). Our model heat budget indicates that horizontal heat advection is not a large contributor to heating bias.

Vertical heat advection in the Arabian Sea peaks in JJA and is close to zero otherwise. The summer monsoon cooling relates to a seasonal increase in upwelling induced by southwesterly winds. CMIP6 MMM shows a cooling from vertical advection broadly similar to reanalysis products, but there is a large model range (Figure 8f). Unlike other regions, the residual term is not always negative suggesting that processes other than vertical mixing must be contributing (Figure 8g).

3.4 Other Regions

Our analysis extends to other Indian Ocean regions with notable biases, including the Somali Coast, Bay of Bengal, and the southern subtropical Indian Ocean. In the Bay of Bengal, CMIP6 models accurately capture the biannual SST cycle but consistently show mean temperatures about 0.5°C cooler than observed, with surface heat flux as the predominant driver of seasonal changes (Figure S8 in Supporting Information S1). Horizontal heat advection has little impact on the domain-averaged temperature, while vertical heat advection contributes to cooling from June to August. The seasonal cycle and magnitude of all these heat budget components is well replicated in CMIP6 models, meaning that the cold SST bias does not apparently stem from seasonal differences between models and observed fluxes. Instead, the mean bias appears to be related to a difference in mean state, as annual averaged heat budget terms are also similar to reanalysis. Thus, there is no overall substantial bias in the rate of temperature change throughout the year in this region, despite a systematic shift of the seasonal temperature cycle compared to observations.

The subtropical Indian Ocean around 80°−100°E, 10°–15°S has substantial temperature bias in CMIP6 from around May to December (Figure S9 in Supporting Information S1). The observed SST peaks in February and is lowest in August/September, following the southern hemisphere seasonal cycle. In CMIP6, the region cools too much from May to November, leading to a cool bias in the following months. Heat budget analysis shows there is a cooling bias in Qnet relative to ERA5 almost all year round. An intermodel correlation between Qnet and temperature tendency in May to August, when the cooling bias is largest, indicates generally in these months that the increased cooling from Qnet is largely responsible for increased cooling overall (Figure S10a in Supporting Information S1). Qnet shows too much cooling, due excessive latent heat loss as demonstrated by a significant intermodel correlation (Figure S10b in Supporting Information S1). The overly strong latent heat loss is likely related to moisture biases, given that there is no significant intermodel relationship between latent heat flux and surface winds (Figure 10c in Supporting Information S1). Other heat budget terms are generally similar to the reanalysis products.

Above we examined the drivers of the WTIO region (Section 3). Given that the size of the region may alias some drivers (in particular coastal and open water drivers may be different), we examine a smaller region (40–60E, 10S–10 N) closer to the Somali coast. This analysis shows that the seasonal cycle of SST is very similar between the smaller and larger regions, and the emergence of warm SST biases in May to October is still present. The seasonal evolution of the heat budget for the region is also very similar to WTIO (Figure S11 in Supporting Information S1). There is a bias in the strength and timing of the zonal heat advection term in May to July, which likely leads to the warm SST bias in the region. With a very similar heat budget to WTIO, we analyzed the intermodel relationships that demonstrate how surface wind biases impact on zonal heat advection in WTIO. These mechanisms are robust for the Somali Coast region, with the weaker surface winds leading to weaker zonal flow, and in turn weaker zonal heat advection (Figure S12 in Supporting Information S1). We note that CMIP6 meridional heat advection is stronger in July—September than reanalysis products. This strong positive heat advection is likely due to the overly strong Somali Current as shown in Figure 2.

4 Discussion and Conclusions

Our surface layer heat budget analysis identifies key systematic biases in CMIP6 SSTs, linked to heat budget term variations across the Indian Ocean. These heat budget biases are found to be closely related to systematic biases in surface winds. We summarize the mechanisms at play in each of the regions analyzed in a schematic (Figure 10). In the WTIO, overly warm SSTs from June to August result from weaker southwesterly winds in May–June, that reduce eastward flow north of the equator (Figure 10a). The weak eastward flow means that there is less advection of relatively cool water eastward, leading to a warm SST bias that persists until August. Weak monsoon winds lead to weaker surface flow and warm biases in the western Indian Ocean (Fathrio et al., 2017). In the Arabian Sea, CMIP6 models show a consistent cold bias during the winter monsoon, exacerbated by overly strong northeasterly winds and resulting in increased latent heat loss from November to March (Figure 10b). In SETIO, CMIP6 simulates overly strong wind speed in July to September due to excessive zonal transport off the coast of Indonesia, in turn driving enhanced upwelling (Figure 10c). In SETIO there is also excessive latent heat loss, but inter-model analysis suggests that latent heat flux is not driving cooling biases across models, indicating that other factors such as surface currents and wind-driven upwelling are leading to the cool surface temperature biases. There is a systematic cold bias in the Bay of Bengal throughout the year, despite that the monthly heat budget variables are similar between CMIP6 MMM and reanalysis products. Additionally, annual average heat budget terms do not indicate any process is especially biased through the year, suggesting a mean state change.

Details are in the caption following the image

Schematic of mechanisms driving SST bias in CMIP6 models, showing annual mean SST bias (relative to NOAA OISSTv2) and surface wind stress bias (relative to ERA5) the seasonality of bias. (a) WTIO region seasonality of SST bias and mechanism driving June to August warm bias. (b) Arabian Sea region seasonality of SST bias and mechanism driving November to May cold bias. (c) SETIO region seasonality of SST bias and mechanism driving August to November cold bias in CMIP6 models.

Our results extend previous studies that identified sources of Indian Ocean biases (e.g., Fathrio et al., 2017; Feng et al., 2023; Long et al., 2020) through an analysis of 20 CMIP6 models. However, unlike previous studies that focused on limited regions or lacked comparisons to reanalyses and observations, our study uncovers previously unexplored coupled processes, and shows structural biases across CMIP6 models that lead to seasonal surface temperature biases. We found that many of the biases seen in CMIP5 models persist in CMIP6, such as those in the western tropical Indian Ocean which are linked to overly weak surface currents related to wind biases, and in the Arabian Sea to overly strong surface heat flux cooling (Fathrio et al., 2017).

Examination of a number of reanalysis products revealed that there are large differences across the products that depend on the region and variable considered. There is large variance across the net heat flux products used and has been a topic of discussion in the literature (e.g., Karmakar et al., 2018; Pokhrel et al., 2020; Valdivieso et al., 2017). The differences are in part due differences in bulk formular that the different products use, different assimilation techniques and differences in data that are fed into the reanalysis product. This stems from the sparsity of subsurface observations in the Indian Ocean making it difficult to validate certain model variables. In the SETIO region for example, vertical velocity varies considerably across the reanalysis products in July to October, making it difficult to establish the actual bias in models, highlighting the importance to sustain and expand observations over the Indian Ocean. As there are large discrepancies in the observational and reanalysis products, it is difficult to establish biases in subsurface terms of models. Therefore, to gain insight into the causes of systematic biases in heat fluxes, we examine intermodel relationships to link subsurface variables and surface variables. We find several significant links between different variables in CMIP6, allowing us to identify the possible underlying causes for the differences in simulated processes compared to reanalysis.

Our analysis spans large regions, where finer scale features may be averaged out, potentially affecting the nuanced understanding of regional variations. For instance, in Section 6 we analyze the Somali coast to investigate processes that could be aliased by the large WTIO region. The SETIO region extends over the Maritime Continent, which could alter signals of ocean mixing and surface wind stress with varying depths and topography. We performed some sensitivity analysis which revealed that the results from Section 3 are insensitive to small changes in the domain boundaries, and that the mechanisms summarized in Figure 10 are robust.

The causality of SST bias and surface wind stress is hard to determine in a coupled system. There have been recent studies specifically in CMIP5 and CMIP6 that help isolate which model components are responsible for biases. For example, wind biases in CMIP6 have been found to originate in the atmospheric component of the model. These wind biases in turn cause oceanic processes to be incorrectly simulated, leading to temperature biases at the surface and at depth (Feng et al., 2023). Comparisons between AMIP and CMIP such as those performed by Long et al. (2020) also provide useful insights into the role of air-sea coupling in giving rise to biases. Long et al. (2020) found that IOD-like bias in SST and precipitation exists in many CMIP6 models and that the atmospheric model component accounts for 2/3 of the IOD-like bias, with 1/3 attributed to biases in air-sea coupling. Models with overly strong surface winds tend to have overly strong SST gradients in CMIP class models compared to AMIP runs. Biases in the atmospheric model component, such as the overly strong south Asian summer monsoon is amplified upon coupling with the ocean, leading to the biases in precipitation and winds. In the Arabian Sea, SST bias in CORE II forced ocean simulations is smaller than SST bias in CMIP5 suggesting that coupled processes amplify existing errors in the ocean component of the models (Rahaman et al., 2020). Further understanding the causality of SST biases is needed, as the coupled ocean-atmosphere system has biases in both components.

Biases in the Indian Ocean mean state are linked to the simulated IOD which is typically too strong in CMIP6 models (McKenna et al., 2020). We find that there is a positive IOD like bias in the mean state of the 20 CMIP6 models that contributes to overall stronger IOD across models, consistent with Long et al. (2020). Strong negative correlation (r = −0.68, p < 0.01) between the JJA-averaged SETIO-WTIO SST difference and the SON IOD standard deviation indicates that models with positive IOD like mean state bias tend to have stronger IODs. This suggests that the bias in the west IO and east IO affects the strength of the IOD in the CMIP6 ensemble. The inter-model differences in the strength of the IOD has also been linked to different biases in mean-state SST in the Pacific Ocean as well (McKenna et al., 2020). In CMIP5, models with cool SST in the east IO and the Pacific Ocean cool tongue tend to have stronger IOD. In CMIP6 this relationship is shifted, with IOD strength now related to the Pacific Ocean warm pool temperature. This shows that biases in remote basins can affect biases in the Indian Ocean. Further research is needed to unravel the mechanisms driving IOD-like mean state biases, their strength, and their connections to the Pacific Ocean.

Many of the shortfalls in simulating interannual variability stem from climatological biases in the mean state like winds, SST, and monsoon circulation. The biases in the seasonal cycle/mean state impact simulations of climate variability in the Indian Ocean such as the IOD and IOB which are phase-locked to the seasonal cycle (Halder et al., 2021; Li et al., 2015b) and are affected by factors such as the mean stratification. In addition, models with more shallow thermocline domes on the southwest IO simulate stronger IOB (Zheng et al., 2016), showing that the accuracy of mean state ocean simulation is important for the simulation of interannual variability. Given we now have increased understanding of mechanisms leading to seasonal cycle biases in SST, wind simulation could be improved to help reduce SST biases.

Understanding SST bias is a complex problem as it is a variable affected by coupled processes in the ocean and atmosphere. Biases in different model components are interconnected and can feed back into each other. The situation is made doubly difficult by our inability to understand if certain processes are indeed biased or if they are different to reanalysis because of a lack of observations. Here we have combined multiple observational data sets, reanalysis and exploited intermodel differences in an attempt to understand important seasonal biases in the relatively understudied Indian Ocean. The intermodel comparison of seasonal heat budget components and their relationship to other surface variables show the dynamics that lead to SST bias in the CMIP6 multimodel ensemble.


The authors acknowledge ARC CE170100023 and the Australian Government's National Environmental Science Programme (NESP). A. Santoso was supported by Centre for Southern Hemisphere Oceans Research (CSHOR). This research/project was undertaken with the assistance of resources and services from the National Computational Infrastructure (NCI), which is supported by the Australian Government. NOAA Optimum Interpolation (OI) SST v2 and NCEP Global Ocean Data Assimilation System (GODAS) data provided by the NOAA PSL, Boulder, Colorado, USA, from their website at https://psl.noaa.gov. ARGO data were collected and made freely available by the International Argo Program and the national programs that contribute to it (http://www.argo.ucsd.edu, http://argo.jcommops.org). The Argo Program is part of the Global Ocean Observing System. Open access publishing facilitated by University of New South Wales, as part of the Wiley - University of New South Wales agreement via the Council of Australian University Librarians.

    Conflict of Interest

    The authors declare no conflicts of interest relevant to this study.

    Data Availability Statement

    All CMIP6 data is freely available at https://esgf-node.llnl.gov/search/cmip6/. All observational and reanalysis data used is summarized in Table 2, with citations. The scripts used to create all figures, regrid data, and calculate heat budget terms are available at https://github.com/SebastianMckenna/Heatbudget_notebooks.