Nonlinear canonical correlation analysis (NLCCA) via a neural network approach was applied to the monthly surface wind stress (WS) and sea surface temperature (SST) in the tropical Pacific for the period 1961-1999. The strength of the nonlinearity varies with the lead/lag time between WS and SST. Relative to the CCA modes, the NLCCA modes explain more variance of the two sets of variables and have higher canonical correlations, particularly, at longer lead/lag times. Unlike the CCA, the NLCCA modes are capable of capturing the asymmetry in the spatial patterns between warm El Niņo and cool La Niņa episodes, with the westerly anomalies and positive SST anomalies located further east of the easterly anomalies and negative SST anomalies. With the WS lagging and then leading the SST, the roles of the predictor field and the lagging response field were interchanged--- the spatial asymmetry was found to be considerably stronger in the response field than in the predictor field.

The NLCCA was then applied to the subset data for the 1961-75 and 1981-99 periods, separately. The leading NLCCA mode between the WS and SST reveals notable interdecadal changes of ENSO behaviour before and after the mid 1970s climate regime shift, with greater nonlinearity found during 1981-99 than during 1961-75. Spatial asymmetry (for both SST and WS anomalies) between El Niņo and La Niņa episodes was significantly enhanced in the later period. During 1981-99, the location of the equatorial easterly anomalies was unchanged from the earlier period, but in the opposite ENSO phase, the westerly anomalies were shifted eastward by up to 25 degrees. According to the delayed oscillator theory, such an eastward shift would lengthen the duration of the warm episodes by up to 45%, but leave the duration of the cool episodes unchanged. Supporting evidence was found from observations, and from a hybrid coupled model built with the Lamont dynamical ocean model coupled to a statistical atmospheric model consisting of either the leading NLCCA or CCA mode.