Role of Atlantic SST in the Recent ENSO Predictability Change in NMME Model Hindcasts


 The present study analyses the possible change in the seasonal prediction skill of El Nino Southern Oscillation (ENSO) in association with the reported climate modification in the tropical Pacific during the early 21st century. The hindcasts of nine models that participated in the National Multimodel Ensemble Project (NMME) are used for the analysis. Both the boreal summer (JJAS) and winter (DJF) seasons ENSO indices from 4 months and 1-month lead for the period 1981-2018/19 are studied. The analysis shows that all the models have reduced interannual variability as observations for both seasons. There is not much skill (both actual and potential) difference for DJF season for all the models for both the lead times. Summer skill loss for Feb IC is more for models such as CanSIPv2, CCSM3 and NEMO, while it is minimum for CCSM4. There is an increase of skill for Feb IC hindcasts for three GFDL models for JJAS season. Most of the models failed to simulate the ENSO events during the second period. The summer season ENSO pattern in the recent period are influenced by spring time north Atlantic SST anomalies. The models with maximum decrease of skill after 2000 fail to simulate the tropical Atlantic SST anomalies during the initialization months and also the summer season SST anomalies induced by these SST anomalies. The models with better or close to observed patterns with Atlantic SST induced ENSO patterns are only able to maintain the same skill as previous decades.


Introduction
El Nino Southern Oscillation (ENSO) is the climate phenomenon impacting the climate around the world with impacts on health, agriculture, water resources and ecosystem (McPhaden et al. 2006). Fortunately, the occurrence of ENSO can usually be predicted 2-3 seasons ahead using modern operational seasonal prediction systems (e.g., Barnston et al. 2012), which help in preparing for the ENSO impacts. This achievement is made as a result of the improvement in observations made during the last 2 decades (Balmaseda et al 2007) and the better scienti c understanding of ENSO mechanism and improved modelling systems developed by different forecasting centres around the world ( . This decline also coincided with more frequent occurrences of Central Paci c El Niño (CP-ENSO) events, in which the SST warming is not only related to the thermocline feedback, but also some other extratropical processes.
Thus a natural climate change is observed in the tropical Paci c in the recent period, which can in uence the properties of ENSO such as its interannual variability (IAV), frequency, amplitude etc which may be associated with basic state change in the tropics. Based on signal-to-noise principles, it is generally believed that ENSO predictability is related to the magnitude of ENSO variability (i.e., the larger its variability, the less likely that the ENSO signal). It is already proven that seasonal prediction models have a weaker ability to predict the central Paci c ENSO than the east Paci c ENSO (Hendon et al 2009, Pillai et al 2017. There is also evidence that the negative IPO phase is associated with epochs of lower predictability for ENSO (Kirtman and Schopf 1998, Choi et al 2011, Hu et al 2013. Thus it is evident that there will be a change in ENSO skill between the two decades centred around 2000. The present study looks at the difference in ENSO skill in the National Multi-Model Ensemble Project (NMME, Kirtman et al 2013) models to understand the change in the predictive skill of ENSO, which can have a major role in the prediction of global climate. Here we are looking at the skill of ENSO for both its developing and mature phase, ie for boreal summer (JJAS) and boreal winter (DJF).
Along with natural variability of ENSO, it is shown that the ENSO variability is controlled by SST anomalies of both Indian and Atlantic Oceans also ( Meanwhile, other studies with Atlantic SST anomalies have shown that two types of Atlantic SST anomalies can in uence ENSO evolution. The north Atlantic SST anomalies (SST anomalies averaged over the area 90 0 W-20 0 E, 0 0 -15 0 N of the tropical Atlantic) during the spring season is supposed to in uence the evolution of ENSO in the tropical Paci c. Ham et al (2013b) have shown that the ENSO associated with these north Atlantic SST anomalies are mainly central Paci c ENSO events as there is a cancellation of wind anomalies occur over the east Paci c that induced by the north Atlantic SST and evolving ENSO. Meanwhile, the north Atlantic SST anomalies are in uenced by the previous year ENSO conditions also. Studies such as Rodriguez-Fonseca et al (2009), Ham et al (2013a, b)., Wang et al (2009) have also shown that the equatorial Atlantic SST anomalies in the form of Atlantic Nino also in uence the ENSO evolution. These SST anomalies contribute to canonical ENSO through the modi cation of Walker Circulation. Another important nding is that these relationships are not stationary and is found to be strong after the 1990s only. Thus, it may be interesting to look at the variation of in uence of these SST anomalies on ENSO during the two periods along with the changes that occurred in the Paci c basin to understand the recent variations in ENSO predictability.

Data And Methodology
The present study mainly uses SST data from different models that participated in National Multi-Model Comparison Project (NMME, Kirtman et al 2013) for the period 1981-2019. The NMME is a multimodel forecasting system consisting of a series of coupled climate models from U.S. modelling centres, including NCEP, GFDL, NASA, NCAR, and the Canadian Meteorological Centre. In these models, whose hindcast is available up to at least 2018 is only taken in the present study as we have to consider two periods P1 (1981)(1982)(1983)(1984)(1985)(1986)(1987)(1988)(1989)(1990)(1991)(1992)(1993)(1994)(1995)(1996)(1997)(1998)(1999)(2000) and P2(2002-2018/19) as per the availability of the data. SST data for all these model hindcasts are available on the IRI data server: http://iridl.ldeo.columbia.edu/ SOURCES/.Models/.NMME/., which hosts mainly SST and rainfall data sets only. All data has more than 10 ensembles and a lead time of more than 9months. Here x denotes observations and xi denotes forecast/hindcast and for variance and for ensemble mean, which is only for model hindcasts. The rst two terms on the right side of the RE equation will give contribution from model spread as these are dominated by model and observed variance and third term is governed by ensemble mean contributed by signal size to RE. Thus, RE will have value for every year and is a prognostic prediction parameter (Pillai et al 2018). Average of RE will give average predictability called as mutual information (MI) and potential skill is calculated from MI as SST anomalies of north Atlantic (NA_SST, 90 0 W-20 0 E, 0 0 -15 0 N) and Atlantic Nino (ATL, 20 0 W-0 0 , 3 0 S-3 0 N) are used to nd the in uence of Atlantic Ocean on tropical SST. Regression analysis and composite analysis for both the periods are conducted to understand the role of different forcing. Long term seasonal trend is removed from all the data before performing the analysis.

Results
a) Mean Paci c SST state difference for individual models. Table 1 below shows the SST bias and standard deviation for model hindcasts for both the seasons and lead hindcasts. In all the model hindcasts, the standard deviation (i.e., interannual variability) is decreased in the tropical Paci c (table 1) as similar to the observed decrease. Meanwhile, the mean SST difference between the two periods is not much signi cant for the model hindcasts, if the seasonal trend for the period is removed from the data (not shown). Similar to the observations, all the model hindcasts have decreased interannual variability of ENSO for the second period irrespective of the initial condition and target season. Models such as CanCM4i, CanSIPSv2, CCSM3, CCSM4, GFDL_aero4, GMAO6 and NEMO has negative bias over the Nino3.4 region for all ICs but is reduced for the second period and is changed to positive bias for CCSM4. The positive bias of the second period is stronger for the DJF target season from both the Aug IC and Nov IC hindcasts. Models such as GFDL_FLORA and FLORB have a positive bias in both periods and the positive bias increases in the second period. For all the initial conditions, the standard deviation is always overestimated by models such as CanCM4i, CanSIPSv2, CCSM3, GFDL_aero4 and GMAO6. GFDL_FLORA has slightly reduced interannual variability than observed variability except for NOV IC hindcasts. CCSM4 also has reduced interannual variability for all ICs considered except for MAY IC.
Thus all the models have decreased interannual variability similar to observations, but ther is more intermodal difference for both the target seasons and lead hindcasts. b) ENSO skill-actual and potential-.
In the present section, the ability of the models to predict the JJAS and DJF season Nino3.4 index is analysed using correlation for actual and based on RE values for potential skill (equations 1 and 2) and is shown in Table 2. CanCM4i model has a decreased JJAS Nino3.4 index skill for both Feb and May IC hindcasts, while the decrease is stronger (~0.1) for Feb IC. Potential skill is always higher than actual skill, but it has also decreased for Feb IC in recent years. CanSIPSv2 also has a similar reduction of Nino3.4 skill, but the actual skill drop is around 0.2 for Feb IC, while the potential skill drop is around 0.1. The potential skill remains the same in both epochs for May IC and the actual skill decrease is of the order of 0.1 for the 2001-2019 period. CCSM3 has a decrease of 0.28 for JJAS Nino3.4 simulated with Feb IC and 0.20 for May IC with a reduction of 0.15 for Feb IC and 0.08 for May IC potential skills. Feb IC skill loss during the second period is non-signi cant for models such as CCSM4, GFDL_aero4, FLORA and FLORB. But the actual value is less than 0.6 for GFDL models and is 0.63 for CCSM4. GMAO6 has a decrease of ENSO skill of around 0.15 for both ICs, while potential skill remains the same during both periods. NEMO model, which has a skill of 0.65 for Feb IC during the rst period, decreased to 0.29 after 2000 with a decrease of 0.35, while May IC, the skill change is 0.10. The potential skill is high and unchanged during both periods. Thus it is evident that actual skill drops mainly for Feb IC for all models except the GFDL family models, while the maximum decrease is for NEMO and CCSM3. The skill decrease is less for May IC for all models but is relatively higher for GFDL models. Another exciting thing is that potential skill has not much change for May IC and even for Feb IC, the change is very less and is limited to CanSIPSv2, CCSM3, CCSM4 and GFDL_aero4 only.
The same skill comparison for DJF season using Aug (4-month) and Nov (1month) lead ICs are also performed to see whether the same skill drop is applicable for the mature phase of ENSO also. Table 1 shows that compared to JJAS season, the skill drop is negligible for DJF season for both 4 months and 1-month hindcasts. The skill change is always less than 0.05 for 4 months lead itself and also potential skill has no change for both the lead time hindcasts of DJF season ENSO. Thus the skill change of ENSO is mainly concentrated to JJAS season (developing phase of ENSO) and is higher for spring initiation of the hindcasts and so JJAS season is focused rest of the paper. Here we discuss all the changes for all the models for the JJAS target season. Still, only CanCM4i (model with higher skill during period 1 and decreased for period 2), CCSM3 and NEMO (models with the large decrease of skill), CCSM4 (model with less difference between periods) and GFDL_FLORA (having increased skill for Feb IC during period 2) are shown in the following sections.
The analysis of potential skill in these two periods for individual models indicates that de nitely, models have room for improvement. Still, the decrease of potential skill for both ICs are marginal or even nill in some cases indicating that the decrease in skill is not directly related to any model deterioration instead model response to the observed climate variability may be the crucial factor. The calculation of Pot. skill itself indicates the role of ensemble spread and ensemble mean in the skill and maybe the combination of these two are not leading to decreased skill, while the actual skill based on ensemble mean only shows a decreased skill. c) Role of initial SST anomalies in ENSO skill in both the periods.
In the present section, the role of initial SST anomalies in the ensemble mean and observations of JJAS Nino3.4 are analysed with the help of correlation analysis. Here Nino3.4 SST indices from both observations and model hindcasts are correlated with the observed initial month SST anomaly. For eg, for Feb IC hindcasts, JJAS nino3.4 index is correlated with February observed SST and for May IC it is May month initial SST. The same analysis for the JJAS Nino3.4 index with February initial SST for observations and Feb IC hindcasts are shown in Figure 1 and the same with May initial SST are shown in gure 2. The observed summer Nino3.4 SST indices have a stronger relationship with off equatorial SSTA in the tropical Paci c, north Indian Ocean (IO) SSTA of February initial month during both the periods (Fig 1 a and b). Additional signi cant correlation with tropical north Atlantic SST anomalies appears in the second period for February initial conditions. Meanwhile, for May IC, the correlation appears mainly in the tropical east Paci c in both the periods and as of Feb IC, after 2000s North Atlantic SST also plays a role (Fig 2  and b).
The same analysis with model hindcasts shows that CCSM3 and CCSM4 have off equatorial patterns in the Paci c during period1 for February initial SST anomalies, while GFDL_FLORA and NEMO have a strong equatorial pattern at the tropical east Paci c. After 2000, CanCM4i, CCSM4, GFDL_FLORA have off-equatorial patterns and also have strong anomalies in the north Indian Ocean and north Atlantic (Fig 1c-j). NEMO model has an opposite pattern in both Indian and north Atlantic regions (Fig 1j). Meanwhile, the major contribution from May initial SST is from the equatorial east Paci c for all the models during both periods, while the role of north Atlantic in the second period is captured by CanCM4i, CCSM4 and GFLD_FLORA model only (Fig 2c-j). Models such as CanSIPSv2 and GFDL_FLORB also simulates the role of Atlantic SST, while GMAO6 and GFDL_aero4 fail to do so (not shown).  Figure 2 also indicates the increased in uence of Atlantic SST initial state in ENSO evolution and here we analyses whether these SST anomalies induce difference in ENSO evolution in these two periods. Figure 3 shows the regression of MAM season NA_SST on observed and Feb IC simulated JJAS tropical SST for both periods. The NA_SST regressed anomalies during P1 shows central Paci c ENSO pattern as shown by earlier studies, while during P2, the equatorial pattern has cooling in both east and central Paci c with an extension to off-equatorial east Paci c (Fig 3 a& b). Thus the spatial pattern indicates a clear shift of NA_SST in uence from central Paci c to canonical ENSO pattern in the second period. The same is con rmed with the correlation analysis between the indices in table 3  pattern is evident for FLORA model only. Also, the SST pattern in other oceans such as the Indian Ocean is opposite for all the models (Fig 3c-j). In CanSIPSv2 and GMEO6 also has no pattern, while, GFDl_aero4 has warming pattern during period 2 and FLORB has similar pattern as FLORA. Thus the majority of the models fail to capture the NA_SSTA effect on the Paci c Ocean during period2.
The JJAS season SST anomalies obtained by the regression of JJAS ATL SST anomalies for observations and Feb IC and for May IC are shown in Figures 4 and 5. The observed summer season ATL SST anomalies have an almost similar pattern in both the periods with earlier reported canonical ENSO pattern with maximum slightly shifted to central paci c (Fig 4 a&b). The correlation analysis also indicates an almost similar correlation with Nino3 and Nio3.4 in both periods, while EMI has a very low correlation during both periods (Table 3). During period 1, all the model hindcasts from Feb IC has stronger SST pattern in the tropical east Paci c for ATL_SST induced pattern except for CCSM3 and NEMO. The second period has a weak pattern for GFDL_LORB, GMAO6, CanSIPSv2, CCSM3 and GFDL_FLORA hindcast, while others have opposite patterns to that of the observed (Fig 4c-j). At the same time, the May IC hindcasts have cool east Paci c for all the models during period1. But the pattern is weak for CCSM3 and FLORA during period2 and is opposite for NEMO model (Fig 5c-j). Thus, models have better predictability with equatorial Atlantic SST induced pattern as it has a similar pattern as canonical ENSO, while the changes associated with NA_SST are not proper in the majority of the models. The models with a better pattern of NA_SST and ATL induced SST pattern in the Paci c has the better role of initial SST from Atlantic in Figure 2 also. e) Process leading to observed change of NAT_SST in equatorial Paci c Another important question here is that why the observed central Paci c ENSO like pattern induced by NA_SST extended to the east Paci c also during the second period? The present section analyses the reason behind the observed change of NA_SST induced SST pattern from central Paci c ENSO to east Paci c after 2000. Figure 6 shows the 850 hPa circulation and 200 hPa velocity potential for MAM and JJAS seasons for the positive minus negative NA_SST composites for both periods. It can be seen that before 2000, positive NA_SST anomalies induces upper-level divergence (lower-level convergence) over the extreme east Paci c and opposite patterns in the central and east Paci c. This is associated with a westerly anomaly in the east Paci c. By summer, this pattern weakens, and convergence moves to the Atlantic but has persistent westerly wind anomalies at east Paci c and easterly anomaly at central Paci c. This is associated with the observed SST pattern in gure 3a earlier and is documented in Ham  indicates that the in uence of previous winter ENSO on NA_SST, which is the main modulator of spring season NA_SST is also weakened. Thus, we can conclude that the displacement of NA_SST pattern and associated circulation and convection during the boreal spring season during period 2 is responsible for the different effects of North Atlantic SST in equatorial SST in JJAS season. As the role of NAT_SST in CP ENSO reduces after 2000, it seems that the earlier proposed mechanisms involving the ocean role may be responsible for more number of central Paci c ENSO in the recent period (Chung and Li 2013). This needs further analysis also.

f) Spatial SST pattern associated with ENSO
The analysis of Nino3.4 indices for both models and observations indicates that during the rst period, there are 5 El Nino (1982, 1987, 1991, 1994and 1997 and 6 La Nina (1985,1986,1988,1989,1998,1999) events are reported during JJAS and it attained its maximum anomaly by DJF. The majority of the models was able to capture these ENSO years and is re ected in the higher skill of the models. Meanwhile during the second period, only the 2015 El Nino is captured by all the model hindcasts, while the El Nino years of 2002,2004,2006,2009,2014,2018 have not been obtained by the models, mainly during its growing stage of JJAS season. The observation pattern indicates that the above-mentioned years which the model does not capture are mainly the central Paci c events. Similarly, during period 1 also, models misrepresented the central Paci c events such as 1983,1994,1984,1999,1998

Summary And Conclusion
The present study analyses the skill difference of El Nino Southern Oscillation (ENSO) in the recent period using NMME hindcast for 1-month and 4-month lead hindcasts for boreal summer (JJAS) and boreal winter (DJF). We have used Nino3.4 region SST anomalies to represent the ENSO in both model and observations. Initial analysis indicates that all the models have reduced interannual variability during period2 in both the seasons and is in accordance with the observations. Thus, the models simulate the observed decrease in interannual variability of ENSO after 2000. The cool bias of most of the models reduced slightly during period 2, while the GFDL-FLORA model has a positive bias in both periods. The skill (anomaly correlation between the ensemble mean ENSO index and observed value) analysis indicates that there is an increase of skill with the decrease of forecast lead time in both the periods and seasons. Meanwhile, the majority of the models also have decreased skill for mainly long lead hindcast during period2, while some models have increased skill also. The potential skill also decreases for some models after 2000, but it is limited to Feb IC hindcast for the JJAS season. But some models such as CCSM4 has a very small reduction in skill (~0.04), while the models such as CCSM3, NEMO etc has a reduction of around 50% for Feb IC and the GFDL models have increased skill for Feb IC and that of May IC decreased slightly. Similarly, skill decrease of DJF ENSO for both the lead hindcasts is not signi cant for the models.
Further analysis of the in uence of February SST anomaly on JJAS SST simulation of Nino3.4 region shows off equatorial SST pattern in the Paci c and north IO basin-wide anomalies during both the periods while there is additional SST anomalies in the north and the equatorial Atlantic Ocean during period2. Similarly for May IC also there was a strong equatorial SST pattern in the east Paci c during both periods and additional stronger SST anomalies in the equatorial Atlantic. The majority of the models captures the off-equatorial pattern for Feb IC in the paci c, but the north Atlantic SST pattern is weaker during period2, except for CanCM4i, CCSM4 and GFDL_FLORA models. May IC has an equatorial ENSO like pattern for models also, but the role of north Atlantic SST anomalies is limited. Thus, many models have a problem in simulating the role of spring season Atlantic initial SST for simulation of the JJAS ENSO SST.
Earlier studies already indicated that north Atlantic SST anomalies in the spring season in uence the central Paci c SST anomalies during the summer season and ENSO evolution. Meanwhile, our study indicates that the spring NA_SST induces central Paci c pattern during period 1 and the SST anomalies tend to extend east Paci c during period2, indicating a higher correlation between Nino3 and NA_SST during that period.
This observed change in the tropical Paci c SST pattern is associated with the weakening and northward shift of NA_SST induced convergence during the spring season during the P2. Almost all the models fail to capture the relationship in both the periods, while CCSM4 and FLORA model has some cooling in the east and central Paci c during period1 and the extension of central and east Paci c cooling to off equatorial noth Paci c is only for FLORA during period 2 for Feb IC hindcasts. At the same time, cooling in the eastern Paci c is evident for most of the models during period 1 except NEMO and FLORA for Feb IC and for all models for May IC. But after 2000, it is weak for almost all the models. Also, models have opposite anomalies in the IO region for both the Atlantic SST anomalies. Thus, the interaction of Atlantic SST anomalies to the east Paci c is not proper in the models and is worse and opposite for NEMO and is somewhat close to observations for FLORA followed by CCSM4, both of which has a higher skill for ENSO during period2. It is also to be noted that these two are the models with better representation of the role of February initial SST in ENSO in gure 2.
The models also fail to capture the majority of El Nino events in the second period, which are mainly central Paci c events and the warming associated with the ENSO are shifted far westward and is south of the equator mainly. During rst period also