An ECMWF Special Project on:
Non-linear Aspects of the Systematic Errors of the ECMWF Coupled model
in collaboration with:
Laura Ferranti, ECMWF
The first issue was investigated by means of ensemble simulations, in which the ECMWF atmospheric model was forced by SST anomalies of fixed spatial structure but different amplitude and sign (as in Hoerling et al. 1997). This series of experiments focussed on the response to El Niño SST anomalies during the winter season (November to March). A composite of El Niño SST anomalies (ENC) was defined for each month as one half of the difference between the average SST in winters 1982/83, 1986/87 and 1991/92 (warm ENSO events) and the average SST in winters 1983/84, 1984/85 and 1988/89 (cold ENSO events). SST was taken from the ECMWF reanalysis (ERA15); the average SST during the full ERA15 period was assumed as climatology (CLIM).
30-member ensemble integrations were performed with the T63-L50 ECMWF AGCM (cycle 21r4), with initial conditions set on 15 consecutive days in November of two different years, and duration up to the end of March. For each set of initial conditions, five ensembles were run, using as boundary condition an SST field defined as follows:
With this setting, it is possible explore the non-linear dependency of the atmospheric response to the Niño amplitude, as quantified by the parameter a, in terms of both mean fields and intraseasonal variability.
The interactions between the systematic error in the SST simulated by the coupled model and the ENSO response was addressed in a second set of three 30-member ensembles, with initial conditions as in the ensemble above, but with SST defined by
Here, DRF represents the "drift" in SST climatology occurring in coupled model simulations. More precisely, DRF was constructed as the difference in SST between coupled ensemble simulations started in November of years 1991 to 1996 (and used as a "training set" to define the bias of the so-called "Coupled System 1" at ECMWF), and the observed SST in the corresponding periods. The DRF field was constructed on a monthly basis, to represent the evolution of the SST systematic error during a six-month winter integration.
In the following, the ensembles of the first set will be referred to as (m2), (m1), (cl), (p1), (p2) respectively, those of the second set with (m1d), (cld), (p1d) respectively, where "m" and "p" stand for "minus" and "plus" (with reference to the sign of the ENSO anomaly), while "d" stands for "drift".
The effect of the SST drift on the mean rainfall simulated in JFM is shown in Fig. 2a. As expected, the rainfall difference resembles the response to a La Niña event, with an increase in rainfall in the Indonesian region, and a decrease in the central equatorial Pacific and in the northern Indian Ocean. When such a difference is compared to the rainfall systematic error of the ECMWF atmospheric model, as estimated from the ensembles with no SST drift included (Fig. 2b), it is evident that the effect of the SST drift actually tends to compensate the rainfall bias of the uncoupled model over a large fraction of the tropical Pacific and Indian Oceans. This behaviour, as well as the structure of the SST drift in the northern oceans, is suggestive of too strong heat and moisture fluxes simulated by the atmospheric model at the air-sea interface, which may be counterbalanced in the coupled model by the SST cooling produced by such a deficency.
Moving to the analysis of the northern extratropical response to ENSO anomalies, with and without SST drift, Fig. 3 compares the ensemble-mean differences between experiments with ENSO anomalies and those with climatological SST (either observed or modified by the climate drift). The 500-hPa height differences (p1-cl), (m1-cl), (p2-cl), (m2-cl), (p1d-cld) and (m1d-cld) are shown in the six panels.
The asymmetry in the northern extratropical response, when the observed SST climatology is used as a reference, was already noted in previous reports. While the response to the (p1) SST anomaly amplitude exceeds (minus) 100 m in the North Pacific, the response to the opposite SST anomaly shows no feature stronger than 20 m in the Pacific - North American region. Quite surprisingly, the response to La Niña SST anomalies is stronger in the Atlantic than in the Pacific sector. Overall, the response in the Atlantic region is strongly non-linear, with similar signs in El Niño and La Niña simulations.
On the other hand, the response in the presence of SST drift appears to be more linear when the (p1d-cld) and (m1d-cld) difference are compared. The amplitudes of anomalies in the North Pacific are now comparable, and quite close to the amplitudes of the observational counterpars shown by Hoerling et al. (1997). On the other hand, any evidence of ENSO-related anomalies in the Euro-Atlantic sector is absent from the responses in the presence of SST drift.
The shift towards a more linear response in the presence of SST drift is confirmed by an EOF analysis of JFM mean anomalies of 500-hPa height in the Pacific-North American (PacNA) region. Figure 4 shows the first two EOFs computed from the 90 individual members of the (m1, cl, p1) ensembles, using anomalies defined either with respect to the mean of all experiments (top row), or with respect to the mean of each ensembles (middle row), to represent internal variability only. The bottom panels show the principal components (PCs) associated with the EOFs in top row, as well as the projections of the (m2) and (p2) experiments on such EOFs. It is evident that the ENSO response along EOF-1 is strongly non-linear, with the PC showing a clear dependence on the ENSO-anomaly amplitude (a) only for the positive sign. The response along EOF-2, on the other hand, seems to be limited to the extreme values of a.
Figure 5 shows the results of the corresponding analysis performed on the ensembles with SST drift (m1d, cld, p1d). The structure of EOF-1 is very similar to that in Fig. 4, but the dependence of the associated PC on a is much more linear in the range of a = [-1, 1]. On the other hand, the ENSO signal along EOF-2 remains weak compared to the internal variability of the ensembles.
In conclusion, the effect of the (rather strong) SST drift produced during the ECMWF coupled-model integrations seems to have a compensating effect on some of the deficiencies detected in experiments forced by observed SST. This is the case for the Pacific-North American sector, but not necessarily for other regions. In particular, the experiments with SST drift show no significant ENSO signal in the Euro-Atlantic region.
Xie, P. and P.A. Arkin, 1997: Global precipitation: a 17-year monthly analysis based on gauge observations, satellite estimates and numerical model outputs. Bull.Amer.Meteor.Soc., 78, 2539.
Figure 2 : Top panel: difference in JFM-mean rainfall between the (cld) and the (cl) ensembles. Bottom panel: Systematic error of the JFM-mean model rainfall with respect to the Xie-Arkin (1997) climatology, estimated from a weighted average of ensemble-mean rainfall from (m1), (cl), (p1) and (p2) experiments. Units: mm/day. Yellow/blue shading corresponds to positive/negative values respectively.
Figure 3 : Mean 500-hPa height response to ENSO composite SST anomalies of different amplitudes, for ensembles with and without SST drift. Top left: (p1-cl); top right: (m1-cl); centre left: (p2-cl); centre right: (m2-cl); bottom left: (p1d-cld); bottom right: (m1d-cld). Contour interval 20m. Yellow/blue shading corresponds to positive/negative values respectively.
Figure 4 : Top row: EOF 1 (left) and EOF 2 (right) of JFM-mean 500-hPa height for the Pacific/North American region (20N to 90N, 150E to 60W), computed from the 90 ensemble members of the (m1, cl, p1) experiments. Middle row: as in the top row, but for JFM anomalies with respect to the mean of each ensemble, to represent internal variability only. Contour interval 10 m. Bottom row: Projections (PCs) of all ensemble members in the (m2, m1, cl, p1, p2) experiments onto the two EOFs in the top row, plotted w.r.t. the amplitude of the ENSO composite anomaly. The shaded band indicates the range between +1 and -1 ensemble standard deviation.
Figure 5 : As in Fig. 4, but for the EOFs and PCs of the experiments with SST drift included (m1d, cld, p1d).
