Will the difficulty of forecasting weather be affected by climate change? To answer this question, the first step is to recognise the chaotic nature of atmospheric dynamics1. Despite the increase in resolution and complexity of weather forecast systems, atmospheric forecasts face an insurmountable predictability limit2. This stems from the intrinsic properties of the atmospheric attractor—a high-dimensional geometric object on which all the possible atmospheric states settle—and is often referred to as the “butterfly effect”, or dependence on initial conditions3. Atmospheric turbulence injects energy at all spatial and temporal scales, generating chaos and limiting short-term predictability. Predictability further depends on space and time, so that the detection and interpretation of changes is extremely complex4. Indeed, predictability is not a constant, and can be strongly affected by the large-scale atmospheric configuration from which the forecast is initialised5. For example, mid-latitude transitions from blocked to zonal atmospheric states are less predictable than a persistent large-scale zonal flow6. By large-scale zonal flow, we refer to a strong eastward (zonal) jet with few meridional oscillations. If the proportion of predictable vs. unpredictable patterns were to change under anthropogenic forcing, the atmosphere’s intrinsic predictability would also change accordingly. The task of diagnosing changes in atmospheric predictability—and more generally in mid-latitude atmospheric dynamics—under anthropogenic forcing is therefore beset with difficulties7,8,9. Achieving this when focusing on averaged quantities, such as climate variability indices, presents serious challenges10.

Here, we diagnose predictability by relying on two objective metrics that sample the properties of the atmospheric attractor. Such metrics provide information on the number of active degrees of freedom of the system and on the typical timescales of the flow, and hence a measure of the intrinsic predictability of an atmospheric state. This is in contrast to predictability defined relative to the performance of a numerical weather prediction model. Computing such metrics was a major challenge until recently11,12. However, advances in dynamical systems theory now allow us to compute them for instantaneous atmospheric states13. Their calculation is based on analogues, or recurrences, of instantaneous (weather) patterns. Recent results have demonstrated their effectiveness in classifying weather patterns leading to climate extremes in the North Atlantic region13,14, as well as hemisphere-wide atmospheric variability15,16.

We specifically focus on evaluating changes in the intrinsic predictability of the atmospheric circulation over the North Atlantic under anthropogenic forcing, as represented by the above dynamical systems indicators. In the historical period, there is disagreement between reanalysis datasets. For the future period, most of the projections analysed here point to an increase in the atmosphere’s intrinsic predictability. The analysis of Atmospheric Model Intercomparison Project (AMIP) simulations with 4 K warmer oceans and four times atmospheric CO2 concentrations points to the prominent role of a warmer ocean in driving this increase. We name this robust signal the hammam effect. This occurs through the enhancement of zonal atmospheric patterns, which are more predictable than meridional configurations.


Data and predictability metrics

Our analysis is based on two dynamical systems metrics: the local (in phase space) dimension d and the persistence θ−1 17. In simple terms, they describe the recurrences of a system around a state ζ in phase space. In our case, ζ would be a latitude–longitude map of a given variable for a given day and dataset. Values of d and θ−1 are obtained for every time-step (i.e. every state ζ) in the dataset of interest. d provides information on how the system can reach ζ and how it can evolve from ζ, and is a proxy for the system’s active number of degrees of freedom. This information is intrinsically linked to the predictability of ζ, but is local in nature. d can change rapidly over timescales of a few days meaning that, when applied to atmospheric fields, it cannot be interpreted in the same way as the error of a medium-range weather forecast initialised from ζ. θ−1 describes the persistence of ζ in time. A very persistent state is typically highly predictable, while a very unstable state yields low persistence. In this sense, the information provided by θ−1 is more directly linked to that provided by the error of a forecast initialised from ζ18. d and θ−1 therefore provide complementary information on the intrinsic predictability of an instantaneous state of the atmosphere. Details of how these two metrics are computed are provided in the Methods. We compute the above two dynamical systems metrics for the sea-level pressure field in the North Atlantic region (22.5°N–70°N and 80°W–50°E) for several datasets. These include the three longest reanalysis ensembles available over our domain, namely: the National Ocean and Atmosphere Administration’s (NOAA) atmospheric reanalysis of the twentieth century (20CRv2c)19 with 56 members covering the period 1851–2014; the European Centre for Medium Range Weather Forecasts’ (ECMWF) atmospheric model integrations of the twentieth century (ERA20CM, as ERA20C but with no synoptic meteorological data assimilated and all observational information incorporated in the boundary conditions and forcing20), with 10 members covering the period 1900–2100; and ECMWF’s coupled climate reanalysis of the twentieth century (CERA20C)21 with 10 members covering the period 1900–2010. They are complemented by a suite of state-of-the-art climate model simulations, including Coupled Model Intercomparison Project phase 5 (CMIP5) model simulations from 1850 to 2100 (see Methods and Supplementary Table 1 for the exact periods used in this study). The choice of region is motivated by the greater abundance of observations in the North Atlantic during the historical period than in other parts of the globe. The choice of SLP is similarly motivated by the fact that it is well constrained in the long reanalyses, as well as being representative of the main large-scale variability modes and atmospheric features of the North Atlantic region22. The versatility of our metrics allows for their use in transient simulations with different horizontal resolutions, without the need for regridding or detrending. In order to corroborate our findings, we further analyse AMIP simulations with atmospheric CO2 concentrations increased by a factor four and ocean temperatures increased by 4 K. In ref. 13 and in this study dSLP ~12–13 for the North Atlantic. This is consistent with the results of ref. 23 that demonstrate that the low-frequency variability in the Northern Hemisphere is spanned by around half-a-score empirical normal modes, with growth rates of around 3–5 days in phase space.

We also compute d for the SST fields of some of the above datasets. Hereafter, we specify whether we are referring to dSLP or to dSST. We do not adopt this notation for θ as we compute it for the SLP field only. Finally, we note that the multi-model and ensemble-average d and θ discussed below are not the metrics computed on the mean SLP fields, but rather the mean values of the metrics computed on the SLP fields of each model or ensemble member. This avoids introducing spurious trends in the data16. Since we are interested in long-term changes in predictability, we apply a 5-year running average to each quantity.

Observed and projected predictability changes

We first comment on the historical period and the reanalysis datasets (Fig. 1 and Table 1). The absolute values of dSLP and θ for all datasets are reported in Supplementary Fig. 1. The 20CRv2c ensemble shows a significant decreasing trend in dSLP, whereas the inverse persistence increases only over the early period of the reanalysis, before stabilising. This is possibly an effect of the scarcity of observational data in the 19th century and the consequent large spread of the members24. The ERA20CM and CERA20C reanalyses broadly agree with 20CRv2c in θ, as they show a weak neutral and positive trend, respectively, but only cover the period where the 20CRv2c θ stabilises. Important discrepancies emerge in dSLP: ERA20CM displays no significant trend, while CERA20C displays a significant increasing trend, opposite in sign to the significant decreasing trend seen in 20CRv2c. All three datasets also display large interdecadal fluctuations, with ERA20CM showing a lower variability than the other two. It is encouraging to note that, during the common period, CERA20C shows the same interdecadal variability as 20CRv2c; nonetheless, the discrepancy in the long-term dSLP trends between the three reanalyses remains to be explained. The main difference between ERA20CM and CERA20C is the presence, in the latter, of data assimilation and a coupled ocean21. To save computational time, it is common practice to restart the assimilation process of the reanalyses every 10 years25. We speculate that this operation may affect the ocean dynamics because it does not give them time to settle on the attractor, as the trajectory is displaced every 10 years. On the contrary, it affects only marginally the atmosphere-only reanalyses without an active ocean component. This does not, however, explain the differences between ERA20CM and 20CRv2c, which must derive from differences in the numerical models and data assimilation schemes.

Fig. 1
figure 1

Local dimension and inverse persistence for all the datasets. Five-year averages of local dimension dSLP (a) and inverse persistence θSLP (b) minus the respective values dSLP(y) and θSLP(y) computed for the years y = 2000 (or y = 2006 for representative concentration pathway (RCP) scenarios). Different colours correspond to different datasets as shown in the legend. Dots: single members or models. Solid lines: means of the ensembles. SLP, sea-level pressure

Table 1 Mann–Kendall test p values for dSLP, θ and the 5th and 95th percentiles of dSLP at the 5% significance level

The historical CMIP5 simulations (Fig. 1 and Table 1) display a significant decreasing trend in dSLP, whereas θ shows an increasing, albeit very weak, trend. Unlike the reanalyses, CMIP5 data does not display large interdecadal fluctuations in the two dynamical quantities. This is partly due to the fact that, since we are considering a multi-model mean, the interdecadal fluctuations between individual members average out. However, even when individual historical model runs are analysed, the standard deviations of dSLP and θ are smaller than in 20CRv2c and CERA20C, and comparable to those of the ERA20CM dataset. These values increase to levels comparable to—or even exceeding—those of the two former datasets in the representative concentration pathway (RCP) scenarios.

We next discuss future climates using CMIP5 simulations run under the RCP 4.5 and 8.5 scenarios, as well as the Community Earth System Model (CESM) RCP 8.5 large ensemble. The CMIP5 models show a significant decreasing trend in dSLP and a weak increasing trend in θ for the RCP 8.5 scenario (Fig. 1 and Table 1). CESM shows similar results for dSLP but no significant change in θ. There is a large spread among the CESM members—larger in fact than that among the different CMIP5 models. We also note that there is no large difference between the RCP 4.5 and RCP 8.5 dSLP trends. With the exception of the two ECMWF reanalyses, all datasets therefore display a decrease in dSLP throughout the period 1850–2100, albeit in some cases modulated by a marked inter-decadal variability. This trend is primarily associated with the summer (June–August, JJA) and autumn (September–November, SON) seasons (see Supplementary Fig. 2 and Supplementary Tables 2 and 3). Overall, the datasets with decreasing trends show a relative variation of dSLP of about 5–8%.

In order to understand the origin of this trend, we revert to the analysis of the atmospheric configurations prevailing in each dataset. In general, high d are associated with low-predictability meridional configurations, such as blocking or mid-Atlantic ridges. Low d match instead high-predictability zonal configurations13,14. A shift of the upper and lower percentiles of the dSLP distributions would therefore correspond to an enhanced zonality or meridionality of the flow. To test this idea, we repeat the analysis shown in Fig. 1 for the top and bottom five percentiles of the dSLP distributions (see Supplementary Fig. 3). We find that the trends are consistent with those found for the mean values, implying an enhanced zonality of the flow associated with a decrease in mean dSLP. These results are also consistent with the dominant role of the summer and autumn seasons in the d trends. Indeed, the summer NAO displays a long-term trend towards more positive values in the future, corresponding to an enhanced zonality of the flow26. A future year-round increase in flow zonality over the North Atlantic, peaking in autumn, was also found by de Vries et al. 27. The dynamical significance of our results is illustrated by Faranda et al.28, where a low-dimensional model of the jet dynamics over the northern hemisphere is derived. The authors found that the jet dynamics is sensitive to small shifts in a control parameter (the meridional temperature gradient) and that the dimension d is proportional to the number of discontinuities of the jet (a proxy of blocking). Finally, the observed jet dynamics correspond to a region of the parameter space that is sensitive to perturbations, where small changes in the dimension correspond to large dynamical changes. In the dynamical system jargon, this means that the system is close to bifurcation points.

The hammam effect

What could be the root driver of the above changes? To investigate this, we analyse two sets of forced AMIP simulations: the first with 4 × CO2 and the second with a 4 K warmer ocean. In the former, energy is injected immediately everywhere into the atmosphere, whereas in the latter the energy is stored in the ocean and may then affect the atmosphere through surface interactions. In the 4 × CO2 runs, there is no significant change in dSLP or θ (Fig. 2). Therefore, the greenhouse enhancement does not appear to radically alter the dynamical properties of atmospheric motions. There is instead a significant change in the dynamical properties when the ocean is 4 K warmer. In this world, which one may imagine akin to a Turkish bath, or hammam, the local dimension decreases (Fig. 2) and the inverse persistence increases (not shown). An intuitive explanation for this phenomenon could be that a warmer ocean implies a stronger mid-latitude jet and the partial suppression of meridional patterns, such as blocking or Atlantic Ridges29. This leads us to hypothesise that the long-term trends in dSLP discussed above may be associated to changes in SSTs, with higher SSTs in the North Atlantic corresponding to lower local dimensions in the atmosphere. We confirm this inference by computing the local dimension of the SSTs from the subset of CMIP5 models that have both pre-industrial control and 4 K warmer ocean runs. We find a significant increase in the average dSST across the warmer simulations, pointing to concomitant and opposite in sign variations in the atmospheric and oceanic local dimensions, at least in the models analysed here (Fig. 2b).

Fig. 2
figure 2

Differences of local dimension in 4 × CO2 and +4 K Atmospheric Model Intercomparison Project (AMIP) simulations. Differences of Δd between average local dimension dSLP for daily sea-level pressure (SLP) data (a, b) and dSST for monthly sea-surface temperature (SST) fields for the 4 × CO2 and +4 K AMIP simulations with respect to the control runs. Error bars indicate the standard deviation of the mean. Lines: means of the ensembles, indicated in the legend by angular brackets

We next verify whether a clear dependence between SST values and dSST   is also found in the reanalysis datasets. Above-median values of dSST correspond to widespread cold anomalies across the Atlantic basin, while below-median values display anomalies of the opposite sign (Fig. 3). This is not directly comparable to Fig. 2b, as we are looking at SST anomalies here. Nonetheless, it highlights a close coupling between dSST and basin-wide SSTs. For the historical period, all three reanalysis datasets display rapidly increasing SSTs (not shown), yet only one of these datasets (20CRv2c) displays a significant decrease in dSLP, as seen in the models. This points to differences in the ocean–atmosphere coupling as the key to the discrepancies between the three reanalyses. The CMIP5 simulations are coherent with the coupling found in the 20CRv2c reanalysis, since they display a decreasing dSLP trend and a gradual increase in SSTs from the beginning of the historical period to the end of the century30.

Fig. 3
figure 3

Composite sea-surface temperature anomalies for high and low local dimension. Composite sea-surface temperature anomalies (units: K) with respect to the climatological seasonal cycles for days with dSST above (a, c, e) and below (b, d, f) the respective median values. Datasets: COBE, as used in 20CRv2c (a, b), ERA20CM (c, d) and CERA20C (e, f)


We have shown that the atmospheric circulation’s intrinsic predictability in the North Atlantic has increased in the recent past, and will continue to do so in a future with continued high levels of anthropogenic emissions. The local dimension dSLP, which we take to be representative of large-scale atmospheric motions, decreases. At the same time, the persistence θ−1 shows a weaker decrease. Thus, the effect of increasing dimension dominates. The main driver of the trend in dSLP appears to be the warmer (Atlantic) SSTs, which act to strengthen the zonality of the large-scale atmospheric flow and reduce the meridional patterns responsible for the more unpredictable situations. Warmer SSTs also correspond to more predictable atmospheric configurations, since dSLP is lower for warmer ocean states. We named this robust response the hammam effect.

It remains to explain why an increase in the temperature of the climate system, which corresponds to an increase in the Reynolds number for turbulent flows, could lead to more predictable states. For most of the flows, increasing the Reynolds number corresponds to an increase of the number of degrees of freedom. However, there are several examples of the opposite behaviour. A first example is the von Karman flow dynamics: Faranda et al.31 showed that there is a region of very high Reynolds number where a low-dimensional dynamics emerges out of featureless turbulence. Another example is provided by Faranda et al.28 who analysed the jet stream dynamics using a low-dimensional model obtained by embedding the data. The bifurcation sequence, when increasing the control parameter (a surrogate of the Reynolds number), evolves from a high-dimensional noisy fixed point to a lower dimensional structure where the dynamics of the jet switches between blocked and zonal flows. Possible explanations of these phenomena of noise-induced order are investigated in the framework of stochastic dynamical systems32,33.

Our conclusions come with some caveats. First, the ECMWF reanalyses do not show a decrease in dimensionality, but rather show either an increase of dSLP with time (CERA20C) or no trend (ERA20CM). Although the ERA20CM trend is compatible with those observed in some of the CMIP5 models, the CERA20C case is unique amongst the analysed data. We hypothesise that this may be due to the way the observations are assimilated into the reanalyses. For example, they are reinitialised every ten years, possibly leaving the ocean in an unstable transient state. Differences in ocean–atmosphere coupling between the reanalysis datasets may also play an important role. Moreover, our analysis does not allow us to draw conclusions concerning the global circulation nor small-scale phenomena. Indeed, it relates more to the predictability of large-scale motions than to conventional weather predictability, although a relevant correlation between the variation of local dimensions and persistence and the spread in numerical forecasts of SLP fields exists13,18. Finally, the validity of our results is restricted to the North Atlantic region, chosen because of the comparative abundance of observational data during the early reanalysis period. This is, however, also a strength of our methodology: the possibility of focusing on a specific region and scale by choosing the appropriate observable fields allows for future targeted studies addressing changes in predictability occurring, for example, in tropical or monsoonal regions, as well as analyses of the role of small-scale phenomena.


Climate models employed for the analyses

We analyse daily output of the CMIP534 for: 27 historical simulations (Supplementary Table 1), 18 RCP4.5/8.5 projections and ten +4 K sea-surface temperature (SST) and 4 × CO2 atmosphere-only (AMIP) simulations. The historical simulations cover the period 1850–2000; the forcings are consistent with observations and include changes in: atmospheric composition due to anthropogenic and volcanic influences, solar forcing, emissions or concentrations of short-lived species and natural and anthropogenic aerosols or their precursors, as well as land use. RCP4.5 and RCP8.5 projections are projections of future climates (2006–2100) forced by two representative concentration pathway (RCP) scenarios. These result in a radiative forcing of 4.5 and 8.5 W m−2, respectively, in year 2100, relative to pre-industrial conditions. AMIP simulations cover the 1979–2005 period and are performed by prescribing SSTs and sea ice boundary conditions to an atmosphere-only model. +4 K simulations impose a uniform 4 K warming to observed SSTs, while the 4 × CO2 simulations impose quadrupled atmospheric CO2 concentrations relative to the pre-industrial control value of 280 ppm. We additionally analyse a 32-member ensemble of simulations from the CESM35. This covers the 1950–2100 period, and follows an RCP8.5 scenario starting from 2006. As the historical part of the simulations is much shorter than for the other datasets, we analyse d and θ trends over the whole CESM period as opposed to performing separate analyses for historical and future trends.

Dynamical systems notions

The attractor of a dynamical system is a geometric object defined in the space hosting all the possible states of the system (phase-space). Each point ζ on the attractor can be characterised by two dynamical indicators: the local dimension (d), which indicates the number of degrees of freedom active locally around ζ, and the persistence (θ−1), a measure of the mean residence time of the system around ζ13.

Local dimension and persistence

To determine d, we exploit recent results from the application of extreme value theory to Poincaré recurrences in dynamical systems. This approach considers long trajectories of a system—in our case successions of daily SLP latitude–longitude maps—corresponding to a sequence of states on the attractor. For a given point ζ in phase space (e.g. a given SLP map), we compute the probability that the system returns within a ball of radius \(\epsilon\) centred on the point ζ. The Freitas et al.36 theorem, modified by Lucarini et al.37, states that logarithmic returns:

$$g(x(t)) = - \log ({\mathrm{dist}}(x(t),\zeta ))$$

yield a probability distribution such that:

$$\Pr (z > s(q)) \simeq \exp \left[ { - \vartheta (\zeta )\left( {\frac{{z - \mu (\zeta )}}{{\sigma (\zeta )}}} \right)} \right],$$

where z = g(x(t)) and s is a high threshold associated to a quantile q of the series g(x(t)). Requiring that the orbit falls within a ball of radius \(\epsilon\) around the point ζ is equivalent to asking that the series g(x(t)) is over the threshold s; therefore, the ball radius \(\epsilon\) is simply es(q). The resulting distribution is the exponential member of the Generalised Pareto Distribution family. The parameters μ and σ, namely the location and the scale parameter of the distribution, depend on the point ζ in phase space. μ(ζ) corresponds to the threshold s(q), while the local dimension d(ζ) can be obtained via the relation \(\sigma = 1/d(\zeta )\).

When x(t) contains all the variables of the system, the estimation of d based on extreme value theory has a number of advantages over traditional methods (e.g. the box counting algorithm38,39). First, it does not require to estimate the volume of different sets in scale space: the selection of s(q) based on the quantile provides a selection of different scales s, which depends on the recurrence rate around the point ζ. Moreover, it does not require the a priori selection of the maximum embedding dimension as the observable g is always a univariate time-series.

The persistence of the state ζ is measured via the extremal index \(0 < \vartheta (\zeta ) < 1\), an adimensional parameter, from which we extract \(\theta (\zeta ) = \vartheta (\zeta )/\Delta t\). θ(ζ) is therefore the inverse of the average residence time of trajectories around \(\zeta\) and it has unit of a frequency (in this study 1/days). If ζ is a fixed point of the attractor \(\theta (\zeta ) = 0\). For a trajectory that leaves the neighbourhood of ζ at the next time iteration, θ = 1. To estimate θ, we adopt the Süveges estimator40. For further details on the extremal index, see ref. 17.