European temperatures in CMIP5: origins of present-day biases and future uncertainties

Abstract

European temperatures and their projected changes under the 8.5 W/m² Representative Concentration Pathway scenario are evaluated in an ensemble of 33 global climate models participating in the fifth phase of the Coupled Model Intercomparison Project (CMIP5). Respective contributions of large-scale dynamics and local processes to both biases and changes in temperatures, and to the inter-model spread, are then investigated from a recently proposed methodology based on weather regimes. On average, CMIP5 models exhibit a cold bias in winter, especially in Northern Europe. They overestimate summer temperatures in Central Europe, in association with a greater diurnal range than observed. The projected temperature increase is stronger in summer than in winter, with the highest summer warming occurring over Mediterranean regions. Links between biases and sensitivities are evidenced in winter, suggesting a potential influence of snow cover biases on the projected surface warming. A brief analysis of daily temperature extremes suggests that the intra-seasonal variability is projected to decrease (slightly increase) in winter (summer). Then, in order to understand model discrepancies in both present-day and future climates, we disentangle effects of large-scale atmospheric dynamics and regional physical processes. In particular, in winter, CMIP5 models simulate a stronger North-Atlantic jet stream than observed and, in contrast with CMIP3 results, the majority of them suggests an increased frequency of the negative phase of the North-Atlantic Oscillation under future warming. While large-scale circulation only has a minor contribution to ensemble-mean biases or changes, which are primarily dominated by non-dynamical processes, it substantially affects the inter-model spread. Finally, other sources of uncertainties, including the North-Atlantic warming and local radiative feedbacks related to snow cover and clouds, are briefly discussed.

Appendix

1.1 Methods

1.1.1 Anomalies and frequency of extremes

In Sect. 3, mean biases $\Updelta$ of a raw variable T, expressed as a function of the space s, the year y and the calendar day d, are simply assessed from the difference between the model M and observations O averaged over time:

$$\Updelta^{M-O} \overline{T}_s = \overline{T}_s^M - \overline{T}_s^O = \frac{1}{N_{y,d}} \sum_{y,d} T^{M}_{syd} - \frac{1}{N_{y,d}} \sum_{y,d} T^{O}_{syd} . $$

(7)

Then, in order to derive biases in frequencies of hot/cold days, we first compute daily anomalies relative to the observed climatology:

$$T_{syd}^{\prime M} = T_{syd}^M - \widetilde{T}_{sd}^O , $$

(8)

where:

$$\widetilde{T}_{sd}^O \equiv \frac{1}{N_y} \sum_y \frac{1}{31} \sum_{d-15}^{d+15} T_{syd}^O , $$

(9)

is the annual cycle of observed T obtained by running average. Thus, at each space point s, the simulated frequency of hot (cold) days is the percentage of days with $T_{syd}^{\prime M}$ above (below) the 90th (10th) percentile of the distribution of $T_{syd}^{\prime O}$.

We use the exact same approach for documenting projected changes, only replacing M and O with future (RCP85) and present (HIST). In particular, this leads to calculate annual cycles for each model separately, and use model-dependent percentiles for the definition of extreme days.

1.1.1.1 Weather regimes and daily classifications

In Sect. 4, centroids of weather regimes are obtained from clustering anomalies $Z_{syd}^{\prime O}$ of the daily Z500 of NCEP2, using the kmeans algorithm (Michelangeli et al. 1995). Before clustering, the effect of the thermal expansion of the lower troposphere is removed, by computing seasonal and spatial averages for all years y:

$$\widehat{Z^\prime}^O_y = \frac{1}{N_s} \frac{1}{N_d} \sum_{s,d} Z_{syd}^{\prime O} , $$

(10)

and then subtracting the linear yearly trend $\underline {\widehat{Z^\prime}^O_y} \sim ay+b$ to original Z500 anomalies:

$$Z_{syd}^{\prime \prime O} = Z_{syd}^{\prime O} - \underline {\widehat{Z^\prime}^O_y}, $$

(11)

with $Z_{syd}^{\prime \prime O}$ the so-called “patterned” Z500 anomalies. As the subtracted field is spatially constant, this procedure does not affect horizontal gradients of Z500 which controls the large-scale circulation.

To speed up the computation, the kmeans is performed on the first N Empirical Orthogonal Functions (EOFs) of the $Z_{syd}^{\prime \prime O}$ field, with N defined as the minimum number of EOFs required to retain 90 % of the total variance (N = 21 in summer, N = 16 in winter). For both seasons, the kmeans is found stable for a number of 4 clusters. We define the “centroids” as the cluster centers in the EOFs-reduced space.

Then, each day (y, d) is assigned to the regime whose centroid is the closest (Euclidean distance) to the day’s patterned anomaly $Z_{syd}^{\prime \prime O}$. This provides a daily classification C ^O _yd for NCEP2, in which (1) days with a spatial correlation below 0.25 between the selected centroid and $Z_{syd}^{\prime \prime O}$ and (2) regime episodes lasting less than three days are “de-classified”, i.e. assigned to a so-called “bin” class. These filters allow to retain robust and quasi-stationary episodes.

We apply the exact same procedure to CMIP5 models in order to compute the simulated occurrences of weather regimes, except keeping NCEP2 centroids as a common reference for both HIST and RCP85 ensembles. That is:

• 1. we subtract the thermal raise to each simulated $Z_{syd}^{\prime M}$ and only retain the patterned Z500 anomalies $Z_{syd}^{\prime \prime M}$;

• 2. we project the $Z_{syd}^{\prime \prime M}$ anomalies onto the four NCEP2 centroids, and apply both minimal-correlation and persistence filters in order to derive C ^M _yd classifications.

As for temperatures, $Z_{syd}^{\prime M}$ anomalies are calculated relative to the NCEP2 climatology $\widetilde{Z}_{sd}^O$ for assessing HIST biases, and relative to the HIST climatologies $\widetilde{Z}_{sd}^M$ for assessing RCP85−HIST changes.

1.1.1.2 Breakdown methodology

In order to separate dynamical from non-dynamical contributions to biases/changes in temperatures, we use the approach developed in Cattiaux et al. (2012a) and also used in De Vries et al. (2012). Temperature anomalies, averaged over time, can be written conditionally to weather regimes $\Upomega_k$:

$$\overline{T^\prime}_s = \frac{1}{N_{y,d}} \sum_{y,d} T_{syd}^\prime = \sum_k f_k \cdot t_{sk} , $$

(12)

where f _k  = N _k /N _y,d is the frequency of occurrence of the regime $\Upomega_k$, and:

$$t_{sk} = \frac{1}{N_k} \sum_{y,d \in \Upomega_k} T_{syd}^\prime , $$

(13)

is the spatial composite of temperature anomalies within regime $\Upomega_k$.

In addition, we write composites of temperature anomalies (t _sk ) as a function ϕ of composites of Z500 anomalies (z _sk ):

$$\forall k, t_{sk} = \phi(z_{sk}), $$

(14)

so that:

$$\overline{T^\prime}_s = \sum_k f_k \cdot \phi(z_{sk}) . $$

(15)

This formulation allows us to write a temperature bias $\Updelta$ between O and M (or similarly a change between HIST and RCP85) as:

$$\begin{aligned} \Updelta^{M-O} \overline{T}_s = \, &\Updelta^{M-O} \overline{T^\prime}_s \\ = & \sum_k f_k^M \cdot \phi^M(z_{sk}^M) - \sum_k f_k^O \cdot \phi^O(z_{sk}^O) , \end{aligned} $$

(16)

which can be split into:

$$\begin{aligned} \Updelta^{M-O} \overline{T}_s & = \sum_k \left[f_k^M-f_k^O\right] \cdot \phi^O(z_{sk}^O) \\ &\quad + \sum_k f_k^O \cdot \left[\phi^O(z_{sk}^M)-\phi^O(z_{sk}^O)\right] \\ &\quad + \sum_k f_k^O \cdot \left[\phi^M(z_{sk}^M)-\phi^O(z_{sk}^M)\right] \\ &\quad + \sum_k \left[f_k^M-f_k^O\right] \cdot \left[\phi^M(z_{sk}^M)-\phi^O(z_{sk}^O)\right], \end{aligned} $$

(17)

or, using shorthand notations $\left\lfloor x \right\rfloor_k = \sum_k x_k$ (adapted from De Vries et al. (2012)), and simplifying the writing of differences:

$$\begin{aligned} \Updelta \overline{T}_s & = \left\lfloor \Updelta f \cdot \phi(z_s) \right\rfloor_k + \left\lfloor f \cdot \phi(\Updelta z_s) \right\rfloor_k \\ &\quad + \left\lfloor f \cdot \Updelta\phi (z_s) \right\rfloor_k + \left\lfloor \Updelta f \cdot \Updelta \left[\phi(z_s)\right] \right\rfloor_k , \end{aligned} $$

(18)

with:

• $BC \equiv \sum_k \left\lfloor \Updelta f \cdot \phi(z_s) \right\rfloor_k$ (Between-Class), the contribution of biases/changes in regimes’ frequencies,

• $WCd \equiv \sum_k \left\lfloor f \cdot \phi(\Updelta z_s) \right\rfloor_k$ (Within-Class, Dynamics), the contribution of biases/changes in regimes’ circulation composites,

• $WC\phi \equiv \sum_k \left\lfloor f \cdot \Updelta\phi (z_s) \right\rfloor_k$ (Within-Class, Physics), the contribution of biases/changes in the transfer function ϕ between circulation and temperatures,

• $\varepsilon \equiv \sum_k \left\lfloor \Updelta f \cdot \Updelta \left[\phi(z_s)\right] \right\rfloor_k$, a second-order residual.

The term ϕ^O(z ^M _sk ) in Eq. 17 corresponds to the temperature anomaly that would produce the observed (present-day) transfer function from model (future) circulations. As proposed in Cattiaux et al. (2012a), this term is evaluated by searching analogues of model (future) circulations within observed (present-day) circulations, the analogy being assessed from maximizing the spatial correlation between fields of Z500 anomalies. Thus, we consider:

$$\phi^O(z_{sk}^M) = \phi^O\left(\left\langle z^O\right\rangle_{sk}\right) , $$

(19)

where $\left\langle z^O\right\rangle_{sk}$ is the composite over the observed analogues of model circulations assigned to the regime $\Upomega_k$. There is an asymmetry in our decomposition (Eq. 17): both BC and WCd terms are expressed as functions of f ^O, ϕ^O and z ^O, and WCϕ is estimated by seeking analogues of $Z^{\prime \prime M}$ within $Z^{\prime \prime O}$, while we could have done the exact opposite. We opt for taking observations O (or present-day HIST) as the reference in order to ease the interpretation of each term.