Twentieth century ENSO characteristics in the IPCC database

Abstract

In this paper, we assess and compare to observations the spatial characteristics of the twentieth Century ENSO SST variability simulated by 23 models of the IPCC-AR4/CMIP3 database. The analysis is confined to the SST anomalies along the equatorial Pacific and is based on the use of a non-linear neural classification algorithm, the Self-Organizing Maps. Systematic biases include a larger than observed proportion for modelled ENSO maximum variability occurring in the Western Pacific. No clear relationship is found between this bias and the characteristics of the modelled mean state bias in the equatorial Pacific. This bias is mainly related to a misrepresentation of both El Niño and La Niña termination phases for most of the models. In contrast, the onset phase is quite well simulated. Modelled El Niño and La Niña peak phases display an asymmetric bias. Whereas the main bias of the modelled El Niño peak is to exhibit a maximum in the western Pacific, the simulated La Niña bias mainly occurs in the central Pacific. In addition, some models are able to capture the observed El Niño peak characteristics while none of them realistically simulate La Niña peaks. It also arises that the models closest to the observations score unevenly in reproducing the different phases, preventing an accurate classification of the models quality to reproduce the overall ENSO-like variability.

Appendix

In preliminary analyses of this study, a first Kohonen map was computed onto filtered but non normalized data. The general physical organization of this map is the same as the one presented in the normalized case in this paper. The minimum or maximum of variability are located in the same areas on the Kohonen map. In particular, the upper part of the map is dedicated to cold conditions, the lower part to warm conditions, and the centre of the map to neutral conditions. Due to the amplitude diversities in our data base, the response for some data set is dominated by its intensity. Figure 11 displays the projection of selected data sets representative of particular behaviours. Indeed, observation behaviour is quite similar to the map in the normalized case. Models which display an amplitude close to the observed one (MIUB, IPSL, BCCR, MRI,...) tend to present a repartition on the map similar to the normalized case. However, models which simulate too weak or to high an amplitude are projected onto very few neurons which prevents the study of ENSO characteristics in detail. The CCCMA-T63 data are projected onto very few neurons in the centre of the map (Fig. 11), which represent neutral conditions. The same behaviours occurred for other models (CCCMA-T47, GISS-ER, MIROC-MR, GISS-AOM...not shown). Due to the strong ENSO amplitude in the CNRM model, its ENSO variability is saturating on the edge of the Kohonen map, onto the very few neurons summarizing the more intense conditions, both warm and cold. The same behaviour is also observed for the INGV and GFDL-1 models (not shown).

Fig. 11

Appendix: number of data projected onto each neuron of Kohonen map for observations and 5 selected models in the non-normalized case. The first measure beside the models acronyms corresponds to the percentage of model data in the region of the Kohonen map that do not collect any observation data. The second measure is the percentage of observation data that belong to neurons where no model data is projected. This information represents how much of the observed variability is captured by the model: realistic models are expected to have smaller values for these two measure

An intercomparative analysis in this case is rather difficult as some models’ projections are dominated by their ENSO intensity. Indeed, for some models, the El Niño or La Niña conditions will be concentrated onto very few neurons, preventing a detailed analysis of their characteristics. As previously stated, in this paper, we aim at analysing and comparing to observations the spatial characteristics of a multi model ensemble twentieth century simulations.

In consequence, the use of normalized data allows us to intercompare models and observations, and models together, as it constrains the amplitude range and makes possible the definition of common thresholds to highlight main spatial characteristics of ENSO variability for all data sets.