Frameworks quantifying how forcings and feedbacks contribute to AA can be classified into the following: energy budget-based diagnostics, mechanism denial experiments, latitudinally-constrained or otherwise idealized forcing, and sea ice forcing experiments.

Energy budget decompositions have been widely used to diagnose climate feedback contributions to surface warming. Individual feedback contributions are evaluated as climate feedback parameters that quantify the global mean TOA energy flux perturbation per unit of global mean surface warming (Wetherald and Manabe 1988; Soden and Held 2006; Shell et al., 2008; Huang et al., 2017a; Pendergrass et al., 2018). Although this method assumes that feedbacks are linear and additive, neural networks can account for nonlinearity (Zhu et al., 2019; Huang et al., 2021a). The energy budget decomposition method can also quantify the influence of regional feedbacks on the warming pattern alongside the radiative forcing, atmospheric energy transport, and ocean heat uptake (e.g., Crook et al., 2011; Taylor et al., 2011a,b; Feldl and Roe 2013; Armour et al., 2013; Pithan and Mauritsen 2014; Hu et al., 2020).

A complementary approach uses the SEB, which is important in the Arctic where the physical validity of the TOA framework is questioned (Pithan and Mauritsen 2014; Payne et al., 2015; Goosse et al., 2018; Henry et al., 2021). Like TOA, the contributions of individual SEB terms to surface temperature change can be diagnosed (Lu and Cai 2009a; Pithan and Mauritsen 2014; Sejas et al., 2014; Laîné et al., 2016; Sejas and Cai 2016; Boeke and Taylor 2018). A SEB decomposition includes additional non-radiative terms (surface turbulent fluxes and ocean heat storage) that are especially important when considering the surface temperature response seasonality.

An expansion of the SEB approach is the coupled atmosphere surface climate feedback response analysis method (CFRAM)—a vertically-resolved version of the energy budget decomposition method (Lu and Cai 2009b; Cai and Lu 2009; Taylor et al., 2013). CFRAM provides a three-dimensional analysis of feedback contributions to the surface and atmospheric temperature response from radiative and non-radiative processes (convection, condensational heating, surface turbulent fluxes, and horizontal heat transport) (Song et al., 2014; Yoshimori et al., 2014). CFRAM does not include a lapse rate feedback and provides a clearer diagnosis of the process contributions to the vertical warming structure. However, the CFRAM is computationally expensive and computes heat transports as a residual; explicitly calculated heat transport terms are straightforward to include in CFRAM however these terms are not routine model outputs. A disadvantage of all energy budget decompositions is that they do not provide clear insights into how different feedbacks are coupled. For example, the radiative sensitivity to albedo changes varies by a factor of two across climate models in the Arctic and Southern Ocean due to inter-model differences in mean-state cloudiness (Donohoe et al., 2020).

Mechanism denial experiments—model simulations where a physical process is “turned off” or locked—also provide insights into the role of various feedbacks (e.g., Wetherald and Manabe 1988; Ingram et al., 1989; Rind et al., 1995; Hall 2004; Vavrus 2004; Graversen and Wang 2009). These studies analyze differences between climate model simulations with a specific process “turned off” and experiments with the process “turned on,” such as sea ice albedo locking (e.g., Graversen et al., 2014), cloud locking (Vavrus 2004; Middlemas et al., 2020), and atmospheric heat transport divergence locking experiments (Graversen and Langen 2019). This approach highlights the coupling between processes that energy budget decomposition approaches cannot (Merlis 2014). The disadvantages of mechanism denial experiments are that they can modify the reference climate, introduce compensating effects, are challenging to apply to comprehensive climate models, and the results are difficult to compare with observations.

Lastly, different modelling protocols have been designed to understand the local and remote mechanisms to AA. Regionally applied greenhouse gas forcing experiments (Section 5e) are one such protocol designed to separate these contributions to Arctic warming (Alexeev et al., 2005; Chung and Räisänen 2011; Yoshimori et al., 2017; Shaw and Tan 2018; Stuecker et al., 2018). Another protocol isolates local and remote mechanisms by prescribing local and remote changes in sea surface temperature and sea ice concentration. Using this approach, Screen et al. (2012) attribute near-surface Arctic warming to local feedbacks and upper tropospheric warming to remote processes. Recent years have seen a proliferation of modeling experiments in which the sea ice component of a coupled ocean-atmosphere model is perturbed, including albedo reduction (e.g., Blackport and Kushner 2016; Liu and Fedorov, 2019), LW emissivity manipulation (e.g., Liu et al., 2019); sea-ice ghost forcing (e.g., Deser et al., 2015), ocean heat flux adjustment (Oudar et al., 2017), and sea ice nudging (McCusker et al., 2017; Smith et al., 2017). Although they all produce a consistent atmospheric circulation response (Screen et al., 2018), the various protocols make different and confounding assumptions regarding conservation of energy and melt water.

Each diagnostic method has strengths and weaknesses (Table 1) associated with technical aspects and underlying assumptions. These differences confound the ability to clearly assess the process contributions to AA. The community needs to address this issue to advance AA science.

Sea ice and snow cover changes via the positive surface albedo feedback are a principal driver of AA (Arrhenius 1896; Budyko 1969; MW75; Hall 2004). The surface albedo feedback operates when (high albedo) sea ice and snow cover melts and reduces surface albedo by uncovering the (low albedo) ocean and land surfaces underneath. Reducing surface albedo causes greater absorption of solar radiation that warms the surface and drives additional sea ice and snow melt. Studies estimate that the sea ice-snow albedo feedback is responsible for 30–60% of the total CO₂-induced Arctic warming (Dickinson et al., 1987; Hall 2004; Taylor et al., 2013; Boeke and Taylor 2018; Duan et al., 2019) and is the largest local Arctic feedback (Taylor et al., 2013; Yoshimori et al., 2014; Goosse et al., 2018). Multi-centennial climate simulations show that Arctic warming slows after most of the sea ice melts, further highlighting the importance of sea ice (Bintanja and van der Linden 2013; Dai et al., 2019).

The surface albedo feedback has substantially contributed to the observed Arctic warming. Observations of a reduced snowpack (Warren et al., 1999; Brown and Robinson 2011; Webster et al., 2014) and significant declines in sea ice extent, thickness, and age since 1979 indicate a reduced Arctic surface albedo (Nghiem et al., 2007; Maslanik et al., 2011; Parkinson and DiGirolamo, 2016; Kwok, 2018). Additionally, the albedo of multi-year sea ice has decreased (Riihelä et al., 2013). Perovich et al. (2007) computed that reduced surface albedo has increased the solar energy deposited into the Arctic Ocean by 89% from 1979–2005. CERES data indicate a −0.025 ± 0.004 decade⁻¹ Arctic average albedo decline and a +1.2–1.3 Wm⁻² decade⁻¹ increase in absorbed TOA solar radiation between 2000 and 2018 (Duncan et al., 2020).

The surface albedo feedback has contributed substantially to the inter-model spread in Arctic warming across multiple generations of intercomparisons (Cess et al., 1991; Holland and Bitz, 2003; Hu et al., 2020). This uncertainty results from the complexities of modeling the continuously evolving sea ice and snow coverage, thickness, and optical properties (Zhang et al., 2000; Laxon et al., 2003)—processes for which available data is insufficient. Furthermore, the rapidly evolving factors that govern surface albedo (e.g., snow and sea ice thickness distribution, topography, drift, melt pond and floe size distribution) occur at small scales making parameterization difficult (Holland et al., 2010, 2012; Schweiger et al., 2011; Jahn et al., 2012; Stroeve et al., 2014).

Sea ice and snow also modulate surface turbulent energy fluxes giving rise to the sea ice insulation feedback. This feedback operates when changes in sea ice concentration and snow and ice thickness alter the non-radiative surface fluxes (sea ice conductance and surface turbulent fluxes; Burt et al., 2016). Sea ice loss exposes a larger area of the Arctic Ocean to the atmosphere and allows for a freer exchange of water vapor, aerosol particles, energy, and momentum with the atmosphere. The sea ice insulation feedback is strongest where there are large surface and near surface air temperature differences collocated with reduced sea ice cover (Serreze et al., 2009; Screen and Simmonds 2010a; Screen and Simmonds, 2010b; Serreze and Barry 2011; Boisvert et al., 2015; Boisvert and Stroeve, 2015; Boeke and Taylor 2018; Taylor et al., 2018). In addition, thinner and less snow-covered sea ice promotes greater heat conduction through sea ice (MS80; Rind et al., 1995; Persson et al., 2017). Through these mechanisms, the ice insulation feedback warms and moistens the lower Arctic atmosphere promoting additional warming via an enhanced greenhouse effect (Kim et al., 2016; Boeke and Taylor 2018; Kim et al., 2019; Feldl et al., 2020; Chung et al., 2021).

Sea ice cover influences the Arctic SEB differently during polar day and night and in both cases strongly impacts surface temperature. Less sea ice cover during polar day decreases the surface albedo and increases SW absorption (Figures 8A,B). Less sea ice cover also promotes larger ocean waves due to longer fetches that have the potential to mechanically break-up sea ice (Rogers et al., 2016). The greater effective heat capacity of the ocean relative to sea ice suppresses warming caused by the surface energy gain during polar day, leading to ocean heat storage and a delayed sea ice freeze up (Dwyer et al., 2012). During polar night, less sea ice cover corresponds to a warmer surface temperature, weaker static stability, and larger upwards surface turbulent fluxes (Figures 8C,D); climatologically cold temperatures over sea ice enable large surface warming when the ice thins and transitions to open ocean (Feldl and Merlis 2021). Atmospheric temperature tends to be warmer in regions with less sea ice in part due to the warming and moistening of the lower atmosphere by increased surface turbulent fluxes, increasing downwelling LW (DLW) radiation (Figure 8D). The greater ocean effective heat capacity also changes the relationship between DLW and upwelling LW (ULW); over sea ice surface DLW anomalies do not lead to strong net LW flux imbalances because sea ice temperature quickly warms in response (Persson et al., 2017; Hegyi and Taylor 2018). These differences in the SEB response to a sea ice change during polar day and night are key components of our conceptual model (Section 6).

Temperature feedbacks are major contributors to AA and contribute substantially to the inter-model differences in CMIP5 (Pithan and Mauritsen 2014). Temperature feedbacks are related to the efficiency of radiative cooling to space and are decomposed into contributions from a vertically-uniform temperature change (the Planck feedback) and the effect of the deviation from a vertically-uniform warming (the lapse rate feedback).

The Planck feedback contribution to AA originates from the nonlinearity of blackbody radiation with temperature, such that at colder temperatures, a larger increase in temperature is required to increase outgoing LW radiation (OLR) by 1 W m⁻². The Planck feedback is negative at all latitudes and contributes to AA because it is more negative at low latitudes. However, this nonlinearity effect may be small. Henry and Merlis (2019) replace the nonlinear temperature dependence of blackbody radiation with a linearized version in an idealized moist GCM and find that it does not modify the surface temperature change pattern, as energy transport and lapse rate changes compensate.

The lapse rate feedback contribution to AA originates from the meridional gradient of the feedback sign, negative at low latitudes and positive at high latitudes. In the tropics, convection pins the atmospheric temperature profile to the moist adiabat leading to a larger warming in the upper troposphere than at the surface. This “top-heavy” vertical warming structure leads to a larger increase in OLR per unit increase in surface temperature—a negative lapse rate feedback. By contrast, the Arctic lapse rate feedback is positive because stable stratification promotes bottom-heavy warming. At high-latitudes, the atmosphere is close to radiative-advective equilibrium causing the lapse rate feedback to depend on the type of perturbation: a change in greenhouse forcing, for example, has a more bottom-heavy temperature response than a change in atmospheric heat transport (Payne et al., 2015; Cronin and Jansen, 2016).

This dependence on perturbation type presents a challenge in determining the relative importance of radiative, surface-based, and advective controls on the lapse rate feedback. In the absence of a surface albedo feedback, Henry et al. (2021) finds that the increase in CO₂ and water vapor alone cause a surface-enhanced warming, consistent with analytic column model results (Cronin and Jansen 2016). Song et al. (2014) argue that the water vapor and albedo feedbacks cause the positive Arctic lapse rate feedback. Mechanism denial experiments reveal that the surface albedo feedback enhances the high-latitude lapse rate feedback (Graversen et al., 2014; Feldl et al., 2017a), or equivalently surface-amplified warming is found in targeted sea ice loss experiments (e.g., Screen et al., 2018). Further, the Arctic lapse rate feedback is strongly correlated across models with summer sea ice loss and cold-season increases in surface turbulent heat fluxes (Feldl et al., 2020; Boeke et al., 2021). Atmospheric energy transport changes tend to reduce the Arctic lapse rate feedback (Feldl et al., 2020) via increases in moist energy transport and decreases in dry energy transport that warm the mid-troposphere and cool the near-surface atmosphere (Henry et al., 2021). Moreover, the decrease in dry transport is strongly controlled by the surface albedo feedback strength (Feldl et al., 2017b; Henry et al., 2021). The high latitude lapse rate feedback results from the sum of these different processes with strong evidence for the importance of surface processes (Cai and Lu 2009; Boeke et al., 2021).

From the surface perspective, the temperature feedback manifests as increased DLW radiation due to atmospheric warming, warming the surface and increasing ULW radiation. The coupling between increased DLW and ULW via the greenhouse effect constitutes a positive feedback loop amplifying surface and atmospheric warming (Sejas and Cai 2016; Zeppetello et al., 2019). Previous studies argue that this feedback accounts for most of the Arctic surface warming (Pithan and Mauritsen 2014; Laîné et al., 2016; Sejas and Cai 2016). Additional studies point to the importance of increased clear-sky DLW on the fall/winter Arctic warming maximum (Lu and Cai 2009a; Boeke and Taylor 2018). Though important to AA, the surface perspective of the temperature feedback does not provide clear insight into the processes that trigger it.

Cloud processes modulate the radiative fluxes and thermodynamic structure of the Arctic atmosphere. The TOA Arctic cloud feedback in CMIP5 is generally negative (Zelinka et al., 2012) and is positive from the surface perspective (Taylor et al., 2013; Boeke and Taylor 2018), indicating that cloud feedback both increases TOA reflected SW and increases surface DLW (Taylor et al., 2011b; Taylor et al., 2013; Pithan and Mauritsen 2014). The magnitude and large inter-model spread of the Arctic cloud feedback comes from model discrepancies in the projected changes in cloud fraction, particularly at low-levels, and optical depth (Vavrus 2004, Vavrus et al., 2008, Vavrus et al., 2011; Liu et al., 2012; Morrison et al., 2019; English et al., 2014, Vignesh et al., 2020). Multiple interacting processes, illustrated in Figure 9, contribute to inter-model differences in the Arctic cloud feedback: surface-atmosphere coupling, cloud microphysics and precipitation, and interactions with large-scale atmospheric dynamics and thermodynamics (e.g., Curry et al., 1996).

The Arctic optical depth feedback is shaped by changes in cloud thermodynamic phase. In response to warming, cloud ice transitions to water increasing cloud albedo and causing a negative feedback (Figures 9C,D; Mitchell et al., 1989; Li and Le Treut 1992). This feedback is sensitive to cloud ice in the control climate, by determining the amount of ice available to transition. The cloud phase feedback magnitude is likely biased negative in most contemporary climate models due to excessive cloud ice and too little supercooled liquid under present-day conditions, yielding unrealistically large increases in mixed-phase cloud optical thickness with warming (Tsushima et al., 2006; Klein et al., 2009; Komurcu et al., 2014; McCoy et al., 2016; Tan et al., 2016). This cloud optical depth feedback bias may have broader implications to AA by enhancing the Arctic lapse rate feedback (Tan and Storelvmo 2019). Recent model experiments revealed that while global cloud feedbacks warm the Arctic, the local feedback contributes negligibly to Arctic warming (Middlemas et al., 2020) suggesting a potential remote influence (Section 5e). However, the model exhibits a low mixed-phase supercooled liquid bias and likely an optical depth feedback that is too negative.

The stability of the lower troposphere affects cloud processes and constitutes a cloud feedback mechanism. Arctic cloud fraction and optical thickness tend to increase with reduced lower tropospheric stability (LTS; Barton et al., 2012; Solomon et al., 2014; Taylor et al., 2015; Yu et al., 2019). In response to increased CO₂, LTS is expected to decrease, promoting increased cloud fraction and optical depth with a seasonally varying character (Boeke et al., 2021). CMIP5 models show substantial cloud-induced warming in fall and winter coincident with large reductions in LTS (Figures 9C,D; Boeke and Taylor 2018). These reductions in LTS are in part due to the large reductions in sea ice (Pavelsky et al., 2011). Thus, cloud changes induced by the LTS mechanism are coincident with cloud-surface turbulent coupling (Kay and Gettelman 2009; Shupe et al., 2013; Solomon et al., 2014; Taylor et al., 2015; Yu et al., 2019).

Cloud-surface coupling represents an important mechanism through which sea ice influences cloud feedback. Sea ice loss tends to increase cloud fraction and optical depth through increased surface evaporation (Curry et al., 1996; Taylor et al., 2015; Abe et al., 2015; Huang et al., 2017b; Morrison et al., 2019; Huang et al., 2021b). However, the sensitivity of clouds to sea ice loss depends on the cloud-surface coupling state and the air-surface temperature gradient. This condition-dependent behavior is responsible for the seasonality of the cloud response to sea ice loss (Figures 9A,B); observational studies find that more liquid clouds result from reduced sea ice in all seasons except summer (Kay and Gettelman 2009; Boisvert et al., 2015; Taylor et al., 2015; Morrison et al., 2018; Huang et al., 2019a; Huang et al., 2019b). Weak air-surface temperature gradients and decoupled cloud layers are typical in Arctic summer conditions (Shupe et al., 2013). Recent research suggests that LH and SH flux increases may elicit different cloud responses (Li et al., 2020). The evidence suggests that the cloud-sea ice feedback promotes surface warming in non-summer months.

Cloud masking influences AA by modifying the strength of other feedbacks. Cloud masking is defined as the difference in the cloud feedback determined using the cloud radiative effect and that using the partial radiative perturbation approach. The cloud masking effect operates by altering the TOA radiative perturbation from a feedback relative to clear-sky and is sensitive to present-day cloud properties. Most pertinent for the Arctic, cloud masking damps the magnitude of the surface albedo feedback by reducing the TOA radiative perturbation from a surface albedo change (Soden et al., 2004). Several studies indicate that the cloud masking effect reduced the TOA radiative impact of the observed surface albedo decline by ∼50% (Sledd and L’Ecuyer 2019; He et al., 2019; Alkama et al., 2020; Stapf et al., 2020). While not a feedback, the cloud masking effect highlights a mechanism through which present-day cloud properties influence Arctic climate change.

Lastly, microphysical processes influence the evolution of cloud radiative properties and modulate cloud feedback (Figure 9). Cloud microphysical processes represent sources and sinks of mixed-phase cloud liquid and ice and modulate the water amount, phase partitioning, and the number and size of hydrometeors (Curry et al., 1996; Beesley and Moritz 1999; Klein et al., 2009; Yang et al., 2015; Tan and Storelvmo 2016; Barrett et al., 2017; Furtado & Field 2017; Wang et al., 2018). However, cloud microphysical processes and their interactions with aerosols are poorly represented in climate models. Ice nucleation mechanisms and ice-nucleating particle (INP) properties and sources are either poorly constrained or not represented in models (Xie et al., 2013; English et al., 2014; Komurcu et al., 2014; Schmale et al., 2021). Mixed-phase cloud INP recycling (Fan et al., 2016; Solomon et al., 2018), secondary ice production (Rangno and Hobbs 2001; Lawson et al., 2017; Sotiropoulou et al., 2020; Zhao et al., 2021) and biological INP-sea ice interactions (Wilson et al., 2015; Irish et al., 2017; Hartmann et al., 2019; Creamean et al., 2020) remain unresolved or unrepresented. In addition, the efficiency of the Wegener-Bergeron-Findeisen process (Tan and Storelvmo 2016), updraft velocity, and ice crystal fall speeds (Ervens et al., 2011; Tan and Storelvmo 2019) are also poorly constrained. These gaps in our understanding of cloud microphysical processes preclude a more quantitative assessment of the Arctic cloud feedback and its influence on AA. Observational constraints that statistically characterize the range of Arctic cloud types are needed to improve parameterized processes and reduce cloud-related uncertainty.

The diversity of Arctic surface types (e.g., sea ice, ocean, and land) dictates features of the spatial structure and seasonality of AA. Surface-type dependent characteristics and processes such as albedo, surface turbulent fluxes, vertical and horizontal heat transport, and heat capacity control the impact of each surface type. Understanding how specific surface types influence the spatial distribution and seasonality of AA may help reduce the inter-model spread.

Explanations of regional variations in AA must consider the underlying surface. Observed temperature changes indicate that regions with the largest sea ice loss are warming most rapidly (Robock 1983; Screen and Simmonds, 2010a; Bekryaev et al., 2010; Boisvert and Stroeve 2015; Figure 2). Moreover, the regional characteristics of warming within a climate model is driven by differences in surface properties and feedbacks (Laîné et al., 2016). Figure 10 illustrates CMIP6 model projections showing that the magnitude and seasonality of warming is a function surface type: namely, sea ice-retreat, sea ice-covered, ice-free ocean, and land (Figure 10; definitions in caption).

Several processes conspire to cause the largest Arctic warming in sea ice-retreat and sea ice-covered regions (Figure 10). Surface albedo and sea ice insulation feedbacks strongly enhance surface warming (MS80; Screen and Simmonds, 2010a; Taylor et al., 2013; Pistone et al., 2014; Boeke and Taylor 2018). Cloud feedbacks are also positive in these regions, especially in fall/winter (Section 5c). Strong LTS, seasonal ocean energy transfer drive the release of stored ocean heat via SH and LH fluxes (Figures 11C,F), and changes in surface thermal inertia contribute to the maximum winter warming in these regions (Robock 1983; Sejas et al., 2014; Laîné et al., 2016; Sejas and Cai 2016; Boeke and Taylor 2018; Feldl et al., 2020).

The characteristics of the warming response in ice-free ocean regions differ from sea ice regions. Ice-free regions have a weaker and almost seasonally uniform warming (Figure 10) resulting from the large ocean heat capacity (Dwyer et al., 2012) and weaker positive feedbacks (especially the surface albedo feedback; Boeke and Taylor 2018). Thus, the SEB response is smaller than in sea ice regions and shows an opposite net flux change during winter from differing SH flux responses (Figures 11B,E). Additionally, changes in ocean heat transport also influence the warming (Section 5f), however it is unclear if these changes affect these regions differently.

While warming in land regions has a similar seasonal structure as sea ice, different surface characteristics indicate that a different set of processes cause this signal. Seasonal differences in the surface albedo feedback occur due to the earlier spring peak in land snowmelt compared to sea ice melt (Taylor et al., 2011b). Additionally, the surface albedo feedback is weaker (smaller increases in surface absorbed SW; Figure 11D) over snow-covered land than over sea ice because of smaller albedo differences with the underlying surface, despite being at a lower latitude (Taylor et al., 2011a). Surface turbulent flux changes cool the land during summer as opposed to during winter as in sea ice regions (Figures 11A,D; Laîné et al., 2016; Letterly et al., 2018); the summer warming minimum over land results from increased cooling and earlier snowmelt rather than increased heat storage as in sea ice regions (Boeke et al., 2021). The small heat capacity of land combined with the nonlinearity of the temperature dependence of LW surface cooling (Henry and Vallis 2021) and increased local atmospheric heat transport from sea ice loss to land regions (Deser et al., 2010; Burt et al., 2016; Boeke and Taylor 2018) also contribute to the winter amplification over land.

Despite considerable efforts particularly over the last decade, the role of remote influences on AA is still debated. Here, we define remote impacts on Arctic warming as any warming that occurs due to non-Arctic changes (equatorward of 60°N). Thereby, remote effects are not merely associated with changes in meridional heat transports but include the local feedbacks they initiate or mediate (e.g., water vapor and cloud feedbacks). Understanding the partitioning between local and remotely-induced warming (Figure 12) is crucial for reducing uncertainty in the impacts of non-well mixed climate forcings (e.g., aerosols; Chung and Räisänen, 2011). Further, simulated Arctic warming and variability may depend on the models’ representation of tropical Pacific variability (e.g., Baxter et al., 2019; Ding et al., 2019) and improving Arctic projections may require improved modeling of teleconnections.

Early EBM studies identified the strong impact of meridional heat transports on polar temperatures (Budyko 1969; Sellers, 1969, North, 1975), and, in GCMs, the opposing responses of dry static energy (DSE) and LH transports due to reductions in the meridional temperature gradient and increases in the moisture gradient (MW80). Flannery (1984) extended the dry EBM approach to include the separate effect of increased LH transport with warming; EBMs continue to be used to study polar warming (e.g., Hwang et al., 2011; Rose et al., 2014; Roe et al., 2015; Merlis and Henry 2018; Armour et al., 2019; Feldl and Merlis 2021).

Despite different meridional shapes of the forcing due to CO₂ and solar constant changes, Wetherald and Manabe (1975) and MW80 found that the meridional shape of the response was similar. Langen and Alexeev (2007) identified a preferred polar amplified response mode whose shape is determined by the strength of the TOA radiative restoring feedback and the DSE and LH transports (also see Merlis and Henry (2018)). The concept of a preferred mode is strengthened by the linearity between Arctic and global mean temperature change inferred from the paleoclimate record (Miller et al., 2010) and CMIP5 models (Yoshimori et al., 2017).

GCM experiments have been performed to gauge the remote impact on Arctic warming. Some used a direct extra energy term added to the SEB (“ghost forcing”, Alexeev et al., 2005; Park et al., 2018), some used latitudinally confined CO₂ increases (Chung and Räisänen, 2011; Shaw and Tan, 2018; Stuecker et al., 2018; Semmler et al., 2020) while others specified SST increases at lower latitudes (Yoshimori et al., 2017). Common to these approaches is that any Arctic warming that occurs, does so due to the indirect effects of the remote warming. Chung and Räisänen (2011) attribute 60–85% of Arctic warming to non-local drivers, Yoshimori et al. (2017) find 60–70%, Park et al. (2018) ∼50%, Shaw and Tan (2018) ∼60%, and Stuecker et al. (2018) ∼50%. These studies indicate that non-Arctic forcing increases non-Arctic temperatures, which in turn increase Arctic temperatures. Local-Arctic feedbacks then amplify this remotely induced Arctic warming (Figure 12) to produce a final remotely induced warming that accounts for half or more of the full Arctic warming.

AA therefore arises in part due to an asymmetry between low-to-high and high-to-low latitude impacts: low-latitude warming is efficiently communicated poleward while high-latitude warming is less efficiently communicated equatorward (Alexeev et al., 2005; Chung and Räisänen, 2011; Park et al., 2018; Shaw and Tan, 2018; Stuecker et al., 2018; Semmler et al., 2020). Non-Arctic warming tends to produce a rather uniform meridional warming pattern and thereby does not itself cause AA (Park et al., 2018; Stuecker et al., 2018). Nevertheless, the fact that non-Arctic warming does not stay localized, as opposed to local-Arctic induced warming, implies that remote effects contribute significantly to Arctic warming. Similarly, moist EBMs and idealized GCMs produce polar amplification in the absence of a surface albedo feedback due to the down-gradient transport of moist static energy (Alexeev et al., 2005; Langen and Alexeev 2007; Roe et al., 2015; Armour et al., 2019; Russotto and Biasutti 2020).

Tropical impacts on Arctic warming (e.g., Rodgers et al., 2003) have been elaborated in the “tropically excited Arctic warming mechanism” (TEAM, Lee et al., 2011; Lee 2012; 2014). Enhanced convection in the Pacific warm pool leads to strengthened or more frequent excitement of poleward propagating Rossby waves. Through dynamic heating and increased moisture transport into the Arctic, the wave dynamics increase the DLW radiation and lead to warming. The role of tropical Pacific Rossby wave-driven teleconnections to the Arctic has been highlighted for observed warming over northeastern Canada and Greenland (Ding et al., 2014) and Arctic sea ice trends and variability (Ding et al., 2017; Baxter et al., 2019; Ding et al., 2019; Topál et al., 2020). Planetary waves dominate the transport of heat and moisture into the Arctic and can drive temperature increases (Graversen and Burtu 2016; Baggett and Lee 2017). Synoptic waves also transport heat and moisture to the Arctic, but in smaller amounts and only in conjunction with a background of amplified planetary waves (Baggett and Lee 2017).

Several studies have concluded that atmospheric heat transport changes play a small or negligible role in AA, finding a negative correlation between polar amplification and atmospheric heat transport changes (Hwang et al., 2011; Kay et al., 2012; Boeke and Taylor 2018). However, this conclusion may result from neglecting the amplified warming effect of latent heat transport discussed below. Due to the nearly canceling effects of increased LH transport and decreasing DSE transport, models with high AA tend to simulate only small or even negative net heat transport changes. Similar conclusions of a subsidiary role for atmospheric heat transport were drawn by Pithan and Mauritsen (2014), Stuecker et al. (2018) and Feldl et al. (2020) using a TOA kernel-based approach and Taylor et al. (2013) using CFRAM. The discrepancy between these studies and those showing the importance of low-latitude impacts and LH transports is likely due to 1) the effect of transport-driven increases in LH is amplified by accompanying changes in specific humidity and clouds (i.e., a “water vapor triple effect” Figure 12; Cai 2006; Graversen and Burtu, 2016; Baggett and Lee, 2017; Lee et al., 2017; Yoshimori et al., 2017; Graversen and Langen, 2019), 2) differing attribution of warming to local and remote processes, and 3) a focus on vertically-integrated energy transport, which does not account for a disproportionate effect of lower versus upper tropospheric transport on surface temperature (Cardinale et al., 2021).

The “water vapor triple effect” represents the multiple influences of water vapor on the Arctic energy budget relating to the misattribution of local and remote Arctic warming (Figure 12). Water vapor transport from mid-latitudes into the Arctic has multiple effects on the Arctic energy budget beyond the release of latent heat at condensation; before condensation, the added water vapor increases the greenhouse effect and after condensation it leads to increased cloudiness, which in Arctic winter has a warming effect. We call this the water vapor triple effect and it means that, per unit of energy transported into the Arctic, latent heat transport is more efficient than DSE transport at warming the Arctic. Graversen and Burtu (2016) show that LH transport eventually leads to an order of magnitude more Arctic warming than DSE transport per unit of energy. When looking at net heat transport changes, this amplified effect is overlooked and the change in total atmospheric heat transport is an unreliable measure of the full effect of atmospheric dynamics (Yoshimori et al., 2017). In offline feedback diagnostic approaches, the water vapor triple effect is attributed to local feedbacks (e.g., water vapor, cloud, lapse rate). Thus, many local feedbacks, as conventionally defined, are not exclusively local in nature. Attribution of total Arctic warming should be considered as four contributions: 1) the local forcing, 2) local feedbacks activated by the locally-induced warming, 3) the remote impacts/forcing and 4) local feedbacks activated by the remotely-induced warming. Impacts 1) and 2) should be counted as locally-induced warming while impacts 3) and 4) should be counted as remotely-induced.

The transport of energy by the oceanic circulation modulates Arctic temperature and sea ice and thus can influence AA. Observations show enhanced ocean heat transports into the Arctic through the Fram Strait and the Barents Sea in recent years (Karcher et al., 2003; Schauer et al., 2004; Dmitrenko et al., 2008; Skagseth et al., 2008; Spielhagen et al., 2011; Årthun et al., 2012). Climate models simulate enhanced high-latitude ocean heat transport under global warming (e.g., Holland and Bitz, 2003; Bitz et al., 2006; Hwang et al., 2011; van der Linden et al., 2019). Several studies suggest that this increased ocean heat transport contributes to Arctic warming (Holland and Bitz, 2003; Hwang et al., 2011; Mahlstein and Knutti, 2011; Singh et al., 2017); in contrast, other studies argue that changes in ocean transport are not correlated with Arctic warming (e.g., Pithan and Mauritsen, 2014; Laîné et al., 2016). This discrepancy mostly comes from the difference of the latitudes where the ocean heat transport is focused (Nummelin et al., 2017). Ocean heat transport increases poleward of 60^oN are positively correlated with AA (Holland and Bitz 2003; Hwang et al., 2011; Mahlstein and Knutti 2011).

Several mechanisms contribute to enhanced poleward ocean heat transport under anthropogenic warming. Several studies indicate that increased ocean heat transport in the subpolar North Atlantic is mainly due to warmer Atlantic water (Jungclaus et al., 2014; Koenigk and Brodeau 2014; Nummelin et al., 2017), while other studies highlight ocean circulation changes (Bitz et al., 2006; Rugenstein et al., 2013; Winton et al., 2013; Marshall et al., 2015; Oldenburg et al., 2018; van der Linden et al., 2019). In the latter mechanism, changes in the North Atlantic subpolar gyre or the Atlantic Meridional Overturning Circulation (AMOC) are argued to be important. For example, a strengthened subpolar gyre causes increased oceanic heat transport into the Barents Sea that decreases sea ice and increases oceanic heat release. An anomalous cyclonic circulation is then induced over the Barents Sea that intensifies westerly winds and further promotes oceanic heat transport and warming in the Barents Sea (Ådlandsvik and Loeng, 1991; Goosse et al., 2003; Bengtsson et al., 2004; Arzel et al., 2008; Guemas and Salas-Melia, 2008; Semenov et al., 2009).

Alternatively, the role of AMOC change in high-latitude ocean heat transport and AA is debated. In GFDL models, a stronger AMOC weakening is linked with less high-latitude warming (Rugenstein et al., 2013; Winton et al., 2013). van der Linden et al. (2019) show that changes in the North Atlantic subpolar gyre play a prominent role in modulating ocean heat transport into the Arctic, while AMOC change is a secondary factor in the EC-Earth model. Additionally, the relationship between AMOC and high-latitude ocean heat transport could be different under internal variability and anthropogenic warming (Oldenburg et al., 2018). AMOC is not a one-way forcing on Arctic climate; Arctic sea ice melt under anthropogenic warming may also slow the AMOC after multiple decades (Sévellec et al., 2017; Liu et al., 2019; Li et al., 2021). A recent study (Liu and Fedorov, 2021) reconciles the two-way interaction between Arctic sea ice and the AMOC, which modulates the changes in both variables in a warming climate.

Ocean heat transported into the Arctic from the Atlantic influences Arctic warming and relates to the inter-model spread. Inter-model differences across 18 CMIP6 models (Figure 13) illustrate the relationship between Arctic warming and ocean heat transport across different latitudes. The correlation is positive and becomes statistically significant near 70^oN and strengthens moving poleward (Figure 13A), a result consistent with previous studies (Holland and Bitz 2003; Hwang et al., 2011; Mahlstein and Knutti 2011). At 80^oN where much of the Atlantic ocean heat enters the Arctic via the Fram Strait, the correlation between Arctic warming and ocean heat transport reaches 0.91. Thus, models with more (less) ocean heat imported into the Arctic via the Atlantic sector simulate stronger (weaker) warming during 2015–2,100 under SSP5-8.5 (Figure 13B).

Long-term climate change and mean energy budgets symbolize the accumulation of short timescale, episodic events. The nature of episodic events has implications for our understanding and projecting of AA. In the seasonal mean, the wintertime Arctic SEB and lower tropospheric temperature profiles are dominated by radiative cooling and strong stable stratification (Serreze et al., 1992). However, at any point in time and space, the Arctic winter boundary layer over sea ice or land tends to be either in a radiatively clear state with no clouds or ice clouds or a cloudy state with low-level liquid containing clouds (Stramler et al., 2011). In the radiatively clear state over sea ice, surface radiative cooling (∼-40 W m⁻²) drives surface-based temperature inversions with strengths of ∼10–15 K. In the radiatively cloudy state, the surface is in approximate radiative balance with the cloud layer and a weaker temperature inversion is elevated to or above the cloud-top (Sedlar et al., 2012; Pithan et al., 2014).

These two states occur at different stages of air-mass transformations (Pithan et al., 2018; Nygård et al., 2019). Following the intrusion of warm, moist air masses from lower latitudes, radiative cooling leads to cloud formation driving the boundary layer into the cloudy state. After several days over Arctic sea ice or land, cooling and drying of the air-mass causes the mixed-phase cloud to glaciate or decay, transitioning to the clear state. The moisture supply aloft and cloud-top radiative cooling lead to cloud top moisture inversions (e.g., increases in specific humidity with height). Given the differences in the thermodynamic profile and the SEB between these states, changes in their frequency of occurrence can impact wintertime sea ice growth, near-surface air temperature and lapse-rate, water vapor and cloud feedbacks.

Episodic variability can influence AA through multiple mechanisms. Changes in the frequency of radiatively clear and cloudy states due to a change in the magnitude or frequency of moist air mass intrusions and atmospheric rivers could alter the SEB and cloud feedback. Observational analyses suggest an increase in the number of moist intrusions has already contributed to wintertime Arctic warming and reduced sea ice growth (Woods and Caballero 2016a; Woods and Caballero 2016b; Graham et al., 2017a; Graham et al., 2017b; Hegyi and Taylor 2018). The initial properties of incoming air-masses could also change, influencing the longevity of mixed-phase clouds; warmer, moister, and potentially more aerosol laden air-masses are possible due to warming at lower latitudes. The potential impact of AA and sea ice loss on the frequency of circulation states with strong meridional advection has been intensely investigated over the past decade and continues to be debated (e.g., Cohen et al., 2020). Lastly, surface turbulent fluxes over the ice-free ocean represent another mechanism by which episodic variability can influence AA as the magnitude of SH and LH fluxes can change by ∼100 Wm^-2 depending upon whether the prevailing winds are from sea ice to ice-free ocean or vice versa (Taylor et al., 2018). A quantitative understanding of the Arctic system response to episodic heat and moisture transport events, air-mass transformation, and cloud formation is needed to reduce uncertainty in Arctic projections.

1. A sustained observing system that resolves key Arctic processes is vital. A pursuit is underway (e.g., integrated Arctic Observing Network (AON)) and this work must continue. In addition, we recommend routine Arctic field expeditions with a MOSAiC-like (https://mosaic-expedition.org) scope to provide the missing data needed to advance understanding (Shupe et al., 2020) for all seasons. Our vision is a permanent, floating Central Arctic observatory.

Polar amplification has been studied in depth for at least 50 years. While the leading explanation for amplified polar warming remains the surface albedo feedback and strong stratification at high latitudes, new details highlight the important role of atmosphere, ocean, and sea ice coupling processes. The highly coupled nature of the polar regions is a source of substantial uncertainty in regional climate projections. Our understanding of polar amplification has been wedded to computational and technological advances that have enabled more complex climate simulations with more detailed physical parameterizations. The role of observations has also evolved from a tool for model tuning to now being used for direct analysis.

While these advances have contributed to our understanding of polar amplification and must continue, an important step remains; to raise Arctic climate sensitivity on the climate modeling priority list, giving it equal priority to global climate sensitivity. Currently, state-of-the-art knowledge of Arctic processes (e.g., surface turbulent flux bulk formula) have not yet been widely implemented in climate models (Bourassa, 2013). Given the rapidly changing Arctic sea ice conditions, older parameterizations developed under thicker, multi-year sea ice conditions are likely to be less applicable in the ‘new’ Arctic with a predominantly seasonal sea ice cover. Giving Arctic climate sensitivity a high priority ensures the rapid integration of knowledge into climate models and will accelerate the reduction in Arctic climate projection uncertainty.