全球变化知识资源中心

For global-mean temperature projections, aerosol effects and cloud response are leading sources of uncertainty in radiative forcing and climate feedback, respectively². For regional precipitation projections, we have shown that atmospheric circulation change is the major source of uncertainty (Supplementary Fig. S1). In the tropics, the circulation is coupled with patterns of SST change, whereas in the extratropics, internal variability, random but organized into large-scale spatial patterns, exacerbates the circulation uncertainty.

The problem of regional climate change projections presents a range of challenges in terms of physical understanding, the observational record, climate models and the simulations that we perform with them. For example, what are the long-term observational trends, and what are their causes? How sensitive are regional climate change patterns to forcing types with different spatial distributions (GHGs versus aerosols)? How can we predict robust patterns of circulation and precipitation change? How do systematic errors in models affect the change patterns? What are the relative roles of internal variability and forced response? These questions pose new problems of ocean–atmosphere–land interactions. Understanding these interactions will allow us to reduce circulation uncertainty and build confidence in regional climate projections.

Observations. The quality of the observational record is an inherent source of uncertainty, particularly pertaining to variability on decadal and longer timescales. Limited duration, incomplete spatial coverage and observational errors hinder our ability to characterize past changes and attribute them to anthropogenic forcing, and limit our ability to evaluate models⁶⁵.

The tropical Pacific provides an example. Observational data sets disagree on the pattern of tropical Indo-Pacific SST change^{30, 66}. Spatial variations in SST trends (0.2 °C per century) are generally smaller than the global SST increase (0.6 °C per century), approaching observational errors and/or internal variability. These spatial patterns drive atmospheric circulation changes, which in turn determine rainfall change patterns, as described above. Since all datasets are imperfect, seeking physical consistency among observations, for example between the tropical SST gradient and trade winds⁶⁷, is a way to infer regional change patterns. The assimilation of data into models seeks such consistency, and proves effective for studying variability on synoptic to decadal timescales. Reanalysis products, however, often are not appropriate for climate change studies⁶⁷, as the quality and quantity of assimilated data change over time. A new generation of reanalysis suitable for climate change research is necessary, with use of coupled assimilation to improve consistency between ocean and atmospheric data.

Knowledge of the strengths and limitations of observational data sets is imperative for understanding past climate change, evaluating models and constraining projections. Community efforts to gather such knowledge from experienced data users and developers, and to share it with the wider climate community via 'open-source' platforms (for example, https://climatedataguide.ucar.edu/) are essential⁶⁸. To facilitate multimodel assessments, open-source assessment packages for climate models can be valuable resources. For example, the Climate Variability Diagnostics Package (http://www2.cesm.ucar.edu/working-groups/cvcwg/cvdp) provides key metrics of internal climate variability across models, with comparison to observations⁶⁹. Ongoing efforts to produce a meaningful set of metrics on mean states, internal variability, and response to external forcing are integral to advancing regional-scale model evaluation (http://www-metrics-panel.llnl.gov/wiki/FrontPage). The challenge is to convert insights from model evaluation to model improvements.

Impact of model errors on projections. Despite limitations of observational records, model biases are clearly evident, reducing confidence in regional projections. A common problem is excessive summertime drying of soils in continental interiors, which may impact the land–sea warming ratio. Models simulating excessive summer Arctic sea-ice may have too weak polar amplification⁷⁰. In the tropics, convection and rainfall are organized into east–west elongated bands called the intertropical convergence zone (ITCZ). A long-standing bias is the so-called 'double' ITCZ, referring to models' failure to keep the ITCZ north of the Equator over the eastern Pacific and Atlantic. The double ITCZ bias is related to atmosphere–ocean coupling errors and is likely to affect rainfall change projections in the South Pacific Islands⁷¹ and elsewhere. The 30–60-day Madden–Julian Oscillation is another phenomenon poorly represented in many models⁷² and affecting confidence in projections of the South Asian monsoon, especially the subseasonal variability such as active/break cycles. Thus, despite a relatively robust understanding of tropical rainfall changes (see 'Mechanisms for regional climate change' above), the precise pattern in any particular model may not be credible.

Some biases persist over multiple model generations. It is important to move beyond routine model evaluation (for example, root mean square errors) and develop innovative techniques to evaluate processes impacting regional projections. The equatorial Pacific cold tongue, for example, results from interaction of trade winds and ocean upwelling (Bjerknes feedback). The cold tongue extends too far west in most models, skewing ENSO SST anomaly patterns and hence atmospheric teleconnections. The balance between the Bjerknes feedback and damping by upwelling and surface heat fluxes determines the magnitude and pattern of SST response to global warming^{16, 18}. This balance varies considerably among models. Most CMIP5 models project larger warming in the eastern than western equatorial Pacific¹⁴. But if the upwelling damping were stronger, this change in east–west gradient could reverse⁷³, altering ENSO's magnitude and spatial pattern⁴⁹. Model evaluation should quantify these ocean–atmospheric feedbacks and their role in determining the spatial pattern of SST change. Such process-based model evaluation challenges the observational record, as estimates of process-level variables may only be available from field campaigns in sparse regions and times.

A further challenge is that model processes often involve complex interactions between resolved dynamics and multiple parameterization schemes. It is not the best strategy to update parameterization schemes in isolation, as physical consistency of multiple processes is required. The 'assembly' stage of model development, often erroneously called 'model tuning', would benefit from tighter integration with process-based model evaluation. For example, long-standing tropical biases like the double ITCZ may be influenced by extratropical errors, such as Southern Ocean clouds⁷⁴ and the Atlantic meridional overturning circulation⁷⁵.

Statistical methods have been suggested to adjust regional projections based on evaluation of model errors. Bayesian techniques use large model ensembles with perturbed parameters and weight each member according to its ability to reproduce observations^{76, 77}. Such approaches take into account uncertainties from multiple sources: models, observations and physical understanding. This allows us to move beyond simple ensemble mean and standard deviation approaches common in regional assessments (Fig. 1). The concept of 'emergent constraints' derives relationships between observable quantities and future projection variables in multimodel ensembles and uses the relationship to re-weight the multimodel projections in a similar way to the Bayesian approach^{70, 78}. Emergent constraints cannot deal with errors common to all models, highlighting the need for innovative complementary approaches to improving models.

Effects of internal variability. Any individual observed or simulated climate trajectory contains contributions from internal variability and external forcing. The relative importance of these two contributions depends on temporal and spatial scale, and on the variable of interest^{3, 79, 80}. In the extratropics, internal variability plays a dominant role in multidecadal atmospheric circulation changes, shaping regional patterns of temperature and precipitation changes⁸⁰. For example, large uncertainties in North American air temperature and precipitation trends projected over the next 50 years stem mostly from internal circulation variability⁸¹. To the extent this internal variability is unpredictable, the resultant uncertainty is irreducible. This 'single realization effect' is large enough to mask the forced regional response, presenting a major challenge for understanding and communication of regional climate change^{45, 82}.

Owing to internal variability, ensemble-mean regional climate trends may be misleading^{83, 84}. The top panels of Fig. 4 provide an example of a probabilistic representation of winter SAT trends at a grid point near Vienna, Austria, based on a 30-member initial condition ensemble⁸¹. The trend distribution is broad for 1976–2005; even with the forced response of 0.2 °C per decade, there is a 20% chance that the 30-year SAT trend is negative. As trend length increases, the radiatively forced trend increases while the trend distribution narrows, indicating reduced importance of internal variability.