a. Intraseasonal oscillation around the time of the onset of the El Niño

Intraseasonal oscillations of the Madden and Julian timescale were prominent prior to the onset of many of the recent El Niño events. A short summary of this timescale during 1997–98 and also for several recent events is presented in this paper. These are based on the European Centre for Medium-Range Weather Forecasts (ECMWF) analysis.

Since the pioneering studies of Madden and Julian (1971) on intraseasonal oscillations (roughly the timescale of 30–60 days), many studies have elaborated on the planetary- and regional-scale aspects and implications of this timescale (Krishnamurti et al. 1992). The Climate Diagnostics Bulletin carries monthly accounts on the phase and amplitude of zonal winds over the equatorial latitudes at 850 hPa. We have noted that during several El Niño years these observational summaries suggest important teleconnections of these zonal windsover the near-equatorial southern Indian Ocean and the subsequent onset of the El Niño in the equatorial Pacific Ocean. This timescale was a prominent feature during the El Niño of 1997–98 as well. Modeling the maintenance and prediction of this timescale in global models has met with limited success. The mechanisms for the maintenance of the intraseasonal timescale are not entirely clear at the present time.

Figures 1a–d illustrate a familiar Hovmöller (x–t) diagram of the zonal wind anomaly at 850 hPa to the equatorial belt 5°S–5°N. These are departures from the 16-yr averages for the period 1980–95. The anomalies were constructed using daily forecast values. A salient feature is the alternation of zonal wind anomalies over the Indian Ocean on the timescale of the Madden–Julian oscillation, that is, roughly 30–60 days. The other salient feature is the sudden eastward extension or burst of this anomaly into the equatorial Pacific Ocean. That anomaly moves into the Pacific around the month of onset of the El Niño. We noted this feature during several of the recent El Niño years: 1990–91, 1992–93, 1993–94, and 1996–97. The amplitude of the zonal wind anomaly is of the order of a few meters per second. We shall also examine the scale interactions among the Madden–Julian timescale and the interannual timescales during the El Niño of 1997–98. The zonal wind anomalies shown in Fig. 1 contain many timescales. From the perspective of scale interactions it is of interest to view these Hovmöller diagrams from bandpassed datasets that include the zonal wind anomalies on the 30–60-day timescale (i.e., the Madden–Julian timescales). Figures 2a–g illustrate the bandpassed zonal wind anomalies over the equatorial belt 5°S–5°N at 850 hPa. We show these anomalies for El Niño years 1968–69, 1971–72, 1976, 1981–82, 1991–92, 1992–93, and 1996–97. An inset box within each panel covers the months in which the El Niño onset appeared to be taking place during these respective years. What is clearly apparent in these diagrams is the sudden burst of the Madden–Julian timescale in the zonal wind over the equatorial Indian Ocean. If we look at the period outside this inset box, we do not see a clear eastward propagation from the Indian Ocean. This timescale is stationary or westward propagative outside the inset box. The sudden burst of an eastward propagating wave (on the Madden–Julian timescale) during the onset period is also a very interesting feature.

The mechanisms for this sudden burst of activity over the equatorial Indian Ocean and its propagation into the Pacific Ocean deserve further study. The amplitude of the zonal wind anomalies on this timescale (within the inset box) is of the order 2–4 m s⁻¹. Following our previous studies (Krishnamurti and Subrhmanyam 1982), we have mapped the velocity potential (χ) anomalies on the timescale of 30–60 days during the months January–July 1997 (Fig. 3a). This is a Hovmöller diagram for the equatorial belt 5°S–5°N. This illustration clearly shows the eastward propagating divergent planetary scale wave at 200 mb. It is interesting to note that this wave acquired a large amplitude over the equatorial Indian Ocean as early as February 1997, that is, well in advance of the onset of the El Niño of 1997. Figure 3b shows the west-to-east gradient of the velocity potential anomalies. This is a measure of transient intensity of the east–west circulations. This illustration very clearly shows the teleconnections between the equatorial Indian Ocean and the Pacific Ocean prior to and around the time of the onset of the El Niño of 1997. The amplitude of the east–west circulation is of the order 1–3 m s⁻¹.

Among all of the variables such as zonal wind, surface pressure (not shown), and 200-hPa velocity potential, the velocity potential anomalies exhibit the most robust eastward propagating structure.

b. Energetics in the frequency domain

In this section we shall illustrate scale interactions among the intraseasonal timescales (the Madden–Julian timescale) and the interannual timescales (the El Niño timescales) as viewed from energetics in the frequency domain. Although not central to the coupled model study, it is an important issue for the onset of the El Niño. In order to address these issues, energetics in the frequency domain following Hayashi (1980) and Sheng (1986) were evaluated. Appendix A outlines the mathematical treatment for these energy transformations in the frequency domain. Here we have derived an expression for just the nonlinear energy transfer into a frequency from its interaction with other frequencies. This kind of intrafrequency energy transfer arises from triad interactions in the frequency domain, which follow certain selection rules for permissible transfers. As may be seen in appendix A, the nonlinear transfer to a frequency n arises from triple products. The cospectra P_n needs to be extracted from real data (or from model output data) in order to evaluate these exchanges. Again following Hayashi (1980) and Sheng (1986), we have performed global integrals, through the depth of the troposphere (here from 1000 to 100 hPa). This required once daily datasets for a 3-yr period from 1994 through 1997. These datasets were obtained from the ECMWF data archived at the National Center for Atmospheric Research (NCAR) on a 2.5° lat × 2.5° long mesh. The required datasets for the calculations of frequency cospectra were the u, υ components of the daily winds and the vertical velocity, which were obtained using the kinematics method following Krishnamurti and Bounoua (1996). In the study, we have selected four frequency windows to explore these energy exchanges: (a) all scales larger than 1 yr, (b) the annual cycle, (c) the Madden–Julian timescale, and (d) the synoptic disturbance timescales from 1 to 10 days.

Addressing specific issues on the onset of the El Niño of 1997 in the above context is not a simple matter. If we were to describe these scale interactions from roughly 5 yr of data where one of the frequency windows is greater than 1 yr, we end up looking at both the 1997 and the 1993–94 El Niño, that is, all in one sweep. To address what went on during the El Niño of 1997 required us to follow a data length extension strategy. Energetics in the frequency domain were calculated using six different data lengths, all centered close to 3 yr of data:

1. March 1994–March 1997: 37 months

2. March 1994–April 1997: 38 months

3. March 1994–May 1997: 39 months

4. March 1994–June 1997: 40 months

5. March 1994–July 1997: 41 months

6. March 1994–August 1997: 42 months

We have noted that these frequency–frequency interactions, of interest for this study, generally invoke two members with close periods in the intraseasonal timescales and a member from the interannual timescales. With a 3+ year-long time series, all one can do is calculate the energy exchange in the frequency domain centered on the intraseasonal timescales. This can be done following Sheng (1986). There are two types of energy exchange in the framework. On one hand, there is an exchange of available potential energy into eddy kinetic energy, which involves a quadratic nonlinearity (i.e., a covariance of the vertical velocity and temperature). That exchange by definition (of a quadratic nonlinearity) occurs in the same frequency and not among two different frequencies. Thus, this cannot play a role in exchanging energy between the intraseasonal and the interannual timescale. On the other hand, the exchange of kinetic energy among different timescales invokes triple products primarily among horizontal wind components, which can be computed with 3+ years of data. All that tells us is whether or not the intraseasonal timescales are gaining or losing energy to timescales larger than a year. Table 1 illustrates examples of such interactions among the Madden–Julian and interannual timescales. Some of the salient triad interactions among three frequencies r, s, and n are illustrated. The timescales for r, s, and n are shown in Table 1. When the dataset covering the 3-yr period from March 1994 to March 1997 is considered, we note that the interacting members of the timescales 43 days, 45 days, and 36 months show a net loss of energy (0.51 × 10⁻²) for the interannual timescale of 36 months. Such an energy exchange of the order 0.51 × 10⁻² becomes significant since it is being applied to low-frequency modes. The accumulative effect of the positive sign implies a significant energy exchange over the timescale of several months. However, as we add sequentially an additional month, that is, March 1994–April 1997, March 1994–May 1997, and so on, we note that triad interactions contribute to a net exchange of energy from the intraseasonal to the interannual timescales. These results are shown in Table 1. A positive entry in the third row of each box denotes a gain of energy by a frequency n as it interacts with frequencies r and s. In most instances the interannual timescales gain from such interactions with members of the Madden–Julian timescales. As the datasets from the period of the El Niño (onset) are included sequentially in these computations, we find that the transfers from the Madden–Julian timescales to the interannual timescales appear to become fairly robust. This suggests that during the onset of the El Niño of 1997 the Madden–Julian timescale must have had an important role.

We shall next design two global model experiments, one that assimilates and retains the Madden–Julian timescale in its initial state and one that systematically excludes this timescale in the assimilation via time filtering of this timescale for all of the atmosphere–ocean variables at all vertical levels.

a. The atmospheric model

The atmospheric component of this system is a global spectral model at the resolution 42 waves using triangular truncation and 14 vertical layers between the earth’s surface and the 30-hPa surface. This model utilizes a semi-implicit time-differencing scheme, smoothed orography at a transform grid resolution T42 (i.e., triangular truncation, 42 waves) for this study. Along the vertical coordinate the model utilizes a centered differencing scheme for all variables except the humidity variable where an upstream differencing scheme is deployed. The physical parameterizations of this model include deep and shallow moist convection, dry convective adjustment, surface similarity theory [including the Beljaars and Miller (1990) corrections for surface winds less than 5 m s⁻¹], vertical disposition of surface fluxes using the mixing length theory, and a Richardson number–dependent vertical diffusion. Short- and longwave radiative transfer is based on a band model, and clouds within this radiative transfer algorithm are diagnostic, that is, defined using threshold values of specific humidity, the so-called diagnostic cloud scheme. Surface energy balance computations are coupled to the surface similarity theory and define the diurnal change over land areas. Air–sea interactions are defined via the surface similarity theory where the Charnock formula defines the oceanic roughness. Data initialization and assimilation issues are addressed in section 3d. Details of the atmospheric and ocean models are presented in Fig. 5.

b. The formulation of the ocean model

The present study uses the Max Planck Institute’s global ocean model (Latif 1987). This is a primitive equation model that makes use of the Boussinesq and hydrostatic approximations on a staggered grid. An outline of the ocean model is presented in Fig. 5. This includes 17 vertical levels and a ½° latitude grid near the equator. Realistic bottom topography is included in the model. Ten of the vertical levels reside within the top 300 m. The main equations of this system are provided in LaRow and Krishnamurti (1998).

The ocean model includes two main mixing processes that affect the heat, salinity, and momentum—a Richardson number–dependent mixing scheme based on the formulation of Pacanowski and Philander (1981), and the other is an oceanic mixed layer parameterization, which follows the study of Latif et al. (1994).

c. Coupled atmosphere–ocean model

The coupled ocean–atmosphere model utilizes the two components described above. The coupled model is used both in a data assimilation mode and for free seasonal and annual forecasts (without any flux corrections). In the following section we shall address the components of coupled data assimilation.

d. Physical initialization-based data assimilation for the coupled model

In this study we perform data assimilation using physical initialization for an entire year. That constitutes the coupled assimilation of our model. This assimilation procedure is illustrated in Fig. 6.

The data requirements for this procedure are as follows.

• Yearlong, daily global atmospheric fields (here we have used the ECMWF daily analysis covering the period 1 April 1996–31 March 1997). A list of acronyms appears in Table 2. The ECMWF data, archived at NCAR, was available at 2.5° × 2.5° long resolution. This includes all basic variables at all vertical levels of the analysis from the earth’s surface to 30 mb, which are interpolated to the vertical discretization of The Florida State University global spectral model (Krishnamurti et al. 1991).

• Outgoing longwave radiation (OLR)-based rainfall rates along the swaths of the National Oceanic and Atmospheric Administration (NOAA) satellites, the data are analyzed using a method proposed by Krishnamurti et al. (1983).

• Twice-daily fields of the OLR derived from NOAA satellites [archived by the National Environmental Satellite Data and Information Service (NESDIS)].

Physical initialization, as originally proposed by Krishnamurti et al. (1991), for atmospheric modeling consisted of a reverse cumulus parameterization algorithm for the assimilation of “observed” rain rates, a reverse surface similarity algorithm that provides a consistency of the computed surface fluxes of moisture with respect to the observed rain and an OLR matching algorithm that modifies the upper-tropospheric humidity field such that model-based and satellite-based fields of the OLR are brought close to each other within 10 W m⁻². Within the coupled model another parameter, that is, the SST field, is continually assimilated toward the observed fields of SST. These components of physical initialization are subjected to Newtonian relaxation. Figure 7 illustrates typical examples of daily rainfall fields that are assimilated by the coupled model. These are very close to the observed rainfall totals that are based on satellite estimates. The observed and assimilated rainfall have a correlation of around 0.9. This is additional data that are being assimilated on a daily basis. In addition to the wind stress, the prescribed precipitation is also a major link between the atmosphere and the ocean since that, via reverse similarity, controls the evaporation rate over the rain areas. Furthermore, the prescribed rainfall modified somewhat via physical initialization modifies the freshwater fluxes, and thus the surface layer salinity distribution. This is a critical factor in the coupled assimilation, since it includes daily estimates of net radiation received at the earth’s surface, which is also a function of the daily cloud cover. Global clouds in the radiative transfer are defined using a diagnostic cloud algorithm, which is based on the difference between the model’s initial relative humidity, and certain threshold predefined relative humidities, Krishnamurti et al. (1990).

The deep oceans are first spun up by the initial 10-yr-long integration. The yearlong coupled data assimilation (which follows the oceanic spinup) includes a Newtonian relaxation of the SSTs; no other datasets in the deeper oceans are directly relaxed. However, the premise here is that, with the improved atmospheric datasets and the assimilated SSTs, an overall improvement of the entire ocean would be seen during the 10-yr-long oceanic spinup and the yearlong coupled assimilation. The procedure outlined here is similar to what was proposed for coupled assimilation by LaRow and Krishnamurti (1998) and for atmospheric assimilation by Bengtsson et al. (1996). Table 3 describes the Newtonian relaxation coefficients used in the present study. Here the vorticity, surface pressure, and SSTs are relaxed strongly, whereas a weaker relaxation coefficient is used for the divergence. The reverse algorithms directly alter the humidity field and the temperature field is passive.

In summary, the proposed data assimilation includes the following features. The assimilation essentially retains the rotational part of the wind from the ECMWF analysis since the Newtonian relaxation of the vorticity field is strong. The surface pressure field from the ECMWF analysis is also essentially retained. The premise here is that there are a very large number of surface observations over land and oceans that the ECMWF analysis makes use of and preserving those may be important for data assimilation. Many data void areas, especially over the tropical land and ocean regions, benefit from the daily input of observed rainfall via physical initialization in the proposed data assimilations. In these rain areas, we note an improvement in the vertical distributions of moisture, divergence/vertical velocity convection heating and surface fluxes of moisture (Krishnamurti et al. 1995). The observed assimilated precipitation input into the ocean would continually improve the salinity distributions since the precipitation is explicitly carried in the salinity equation. Several factors contribute to the improvements of SST assimilations: improved surface wind stresses resulting from physical initialization; improved surface energy balance resulting from improved surface evaporation (from the use of reverse algorithms in physical initialization); improved net radiation at the earth’s surface, which includes effects of diagnosed clouds derived from the physical initialization (Lee and Krishnamurti 1995); and the yearlong assimilation provides an overall improvement of the mixed layer and the deeper ocean thermal stratification. This is evidenced from the overall improvement of the assimilated forecasts over a control run.

a. Control experiment

Using a recursive Butterworth filter, the entire yearlong-coupled assimilation data for all variables at all vertical levels of the model (globally) and the SST over the global oceans were time filtered. This procedure was used to extract and subsequently remove the Madden–Julian timescale (30–60 days) from the assimilated data. This was carried out at the interval of 1 day. That string of filtered atmosphere–ocean data was reassimilated for a 1-yr period using the coupled model. This dataset did not include the Madden–Julian timescale explicitly. A seasonal forecast was next carried out with the control experiment. Figures 8a–c show a comparison of the equatorial (5°S–5°N) zonal winds at 850 hPa from (a) observation, (b) from the fully assimilated experiment, and (c) from the experiment that excludes the Madden–Julian timescale initially. These results clearly show that the amplitude and even the direction of propagation of the predicted intraseasonal wave was misrepresented when it was excluded in the data assimilation and the initial state. The results from the fully assimilated experiment, however, show that an intraseasonal wave is retained during the coupled model forecasts. The amplitude is reasonable and the eastward propagation is also clearly evident. The control experiment (that excludes the intraseasonal wave) failed to simulate an El Niño and the climate anomalies of 1997–98. We shall not discuss this control run any further in this paper.

b. The predicted SST anomalies

The predicted SSTAs are illustrated in Fig. 9. Here the anomalies were constructed using average SST values from an 8-yr integration period. These were adjusted such that the model-based mean anomaly for the period April 1997–May 1998 was close to the observed mean value for this period. In this illustration we see the establishment of a well-defined El Niño by July 1997 with a 4°C warm temperature anomaly. That feature extends to the equatorial west coast of South America with values of the warm anomaly as high as 5°C over the eastern equatorial Pacific Ocean by September 1997. This feature is maintained through February 1998 (not shown). Thereafter we note a rapid cooling of the equatorial Pacific waters. The cold waters are clearly seen in July 1998; however, 1°–2°C warm water anomalies do persist near the South American coast through September 1998. The observed–blended SST anomalies, based on the datasets from the Climate Prediction Center, are shown in Fig. 10. These six panels show the SST anomalies for April, July, September, and December of 1997 and for April and July of 1998. These show the formation of a strong El Niño with SST anomalies reaching 5°C by December 1997. Thereafter a weakening of the anomalies and the establishment of the cold phase is noted by July 1997. Warm anomalies of the order 2°C persist south of the equator along the South American coast during July 1998. The starting month of forecasts, that is, April 1997, already sees a weak warm SST anomaly of the order 2°C in the eastern equatorial Pacific Ocean. It should be stated that several operational groups, such as the climate-forecast center of the U.S. National Weather Service and the ECMWF, have shown similar results in the simulation of these SSTA features. The start date of our coupled run was only 2 months prior to the onset of the El Niño. However, the length of integration we present in this paper contains information through December 1998. Overall, this result appears quite promising when the observed anomalies (shown in Fig. 10) are compared to the predicted estimates of Fig. 9. In Fig. 9 we note a large cold SST anomaly over the southeastern Pacific Ocean. That was not present in the observed field displaced in Fig. 10. There was a slight delay in the initial weakening of the cold phase in the model. That cold phase preceded the warm event of 1997. This delay was, however, not very critical since the model very quickly corrected the onset of the warm phase as is evident from the comparison of the SST anomaly by September 1997. In Figs. 9 and 10 we see major difference in the first month in forecasts of SST anomalies; these are the monthly means for April 1997. We do have a spinup problem during the first month after the coupled assimilation. A smooth transition from the assimilation (which includes nudging) to the forecast (which does not include any nudging or flux correction) is desirable. We are examining this issue in further detail. This spinup issue was not detrimental to the future evolution of the SST anomalies and the skill in fact improves considerably as shown in the figures.

The predicted time history of the SST anomalies in the Niño-3.4 region are shown in Fig. 11. SST anomalies of the order +3°C were reached by July 1997. Warm anomalies, in excess of +2°C, prevail through February 1998 and thereafter a rapid decrease in temperature continues through June 1998. Given the length of this forecast, that is, 17 months, these results do seem quite promising. The corresponding observed SST anomalies over the Niño-3.4 region are shown in Fig. 11. The observed SST anomalies over the equatorial Niño-3.4 region were obtained from the Climate Diagnostic Bulletin. Our long coupled model integration started when there already existed a SST anomaly of +0.5°C over the Niño-3.4 region. By November 1997 a peak value close to 3°C is realized. Thereafter a sharp drop continues. By June 1998, the positive anomaly ceases and thereafter the cold phase is established. Overall, Fig. 11 shows a reasonable agreement. It may be important to calculate correlation between the observed and predicted fields of SST anomalies. Our modeling has not reached that stage of quantitative intercomparison. We can qualitatively note that a warm phase does form in the model at roughly the right time and a cold phase does occur rather rapidly after February 1998. A correlation of the observed and the modeled SST anomalies over the Niño-3.4 region was calculated from the data displayed in Fig. 11 and a correlation of 0.72 was noted. It would be desirable to restart the coupled model forecast somewhat earlier than April 1997. A start date of 1 December 1996 would have been more desirable; we propose to include such details on sensitivity of ENSO forecast to start dates in our future studies.

c. Kelvin waves and equatorial ocean temperatures

The onset of the El Niño in this experiment was monitored from the stage of the initial eastward passage of the equatorial Kelvin wave in the ocean. For this purpose we have displayed the depth anomalies of the 20°C isotherm over the Pacific Ocean starting on 15 April 1997 at 15-day intervals (Fig. 12). The depth anomalies, shown here, are departures from the initial values of depth at the start of the seasonal forecast. It should be noted that the ENSO is not first a Kelvin wave. ENSO is a coupled atmosphere–ocean phenomenon and the Kelvin wave is a part of the coupled system. The Kelvin wave was already excited by 15 April 1997 (the start day of the forecast is 1 April 1997). This wave initially moves eastward at the speed of 1° longitude day⁻¹. However its eastward motion slows down as the wave first arrives at the equatorial west coast of South America. The depth anomalies are of the order 10–60 m. By July 1997 the Niño-3.4 region had encountered a warm anomaly of 2°C; at this time the Kelvin wave had already reached the South American coast and reflected and spread northwestward as a Rossby gravity wave. The Kelvin wave first reached the coast by around 30 June 1997. During the initial stages, eastward propagation in the Kelvin wave depth anomaly is stretched along the equator (e.g., 15 June). After its reflection (e.g., 15 August) the reflected Rossby gravity wave is located above the equator around 5°N.

The oceanic mixed layer structure was analyzed for the observed and the predicted temperature anomalies over the equatorial latitudes between 10°S and 10°N. The observed [i.e., based on the Tropical Ocean Global Atmosphere–Tropical Atmosphere Ocean (TOGA–TAO) array datasets] and the predicted (i.e., based on the coupled model forecasts are shown in Figs. 13 and 14, respectively. The positive temperature anomaly, in April 1997, for the observed fields are found roughly 100 to 150 km below the surface. This anomaly moves eastward by December 1997 and this region is replaced by a negative anomaly during August 1997. An eastward propagation is seen through this period of 16 months. The coupled model results, Fig. 14, show a dominant positive anomaly during the months April, August, and December 1997 and a negative anomaly during the months April, August, and December 1998. The predicted amplitudes are somewhat less than the observed anomalies. The largest discrepancy seems to occur in the very first month of integration, that is, April 1997. The observed field shows a single dominant positive anomaly during this month, whereas the 1-month forecast with the coupled model shows a positive and a negative center. We believe that this discrepancy could be removed if we assimilate the temperature anomaly datasets from the TOGA–TAO buoy datasets. The current coupled assimilation includes only the surface SSTs. A comparison of Figs. 13 and 14 shows a considerable phase difference. This difference can possibly be reduced by the direct assimilation of the TOGA tower array subsurface temperatures.

d. Indonesia and Florida fires and divergent wind anomalies

When we look at the divergent wind anomalies through the fields of the velocity potential, we note that the areas of persistent descent occupy a much larger area in comparison to areas of ascent. Mislocation of areas of ascent, such as the ITCZ, by several degrees latitude is common in these coarse-resolution models. However, most regions of descent seem to be covered with the same phase errors. This suggests that perhaps the areas of persistent descent, related dry belts, and regions of drought may be more predictable compared to the narrower areas of ascent and heavy rains. The fields that we noted to be most interesting in this regard were that of the velocity potential anomalies. In order to construct the observed fields of this anomaly, we have taken a 17-yr monthly mean climatology (1980–96). These were based on the ECMWF analysis, archived at NCAR, at 2.5° lat × 2.5° long resolution.

Major fires occurred over Indonesia during September and October 1997. This was a major catastrophic event related to the El Niño. The relationship of these fire events to regional climate are best seen from the maps of the divergent wind anomalies at 200 mb. A sequence of these, based on the analysis of the Climate Prediction Center, are shown in Figs. 15a–f. A striking feature is a divergent inflow center that is noted over the central Indian Ocean (near 20°S) in July 1997. This feature exhibits a slow northeastward motion and is locatedsouth of Sumatra near 15°S in August 1997. During September and October 1997 this anomaly (of the inflowing divergent winds) expands and moves directly over Indonesia, Malaysia, Borneo, and New Guinea. It is apparent that the related large-scale downward motions must have contributed to a sustained dryness of the lower-tropospheric air and the resulting forest fires. Thereafter the divergent inflow center appears to split into two centers, one slowly moving toward 10°S and 90°E, and the other moving toward 10°N and the south of Philippines by December 1997. It is of considerable interest to ask if the monthly mean divergent wind from the coupled model forecasts exhibits some of these same characteristics.

The construction of model-based velocity potential anomalies requires the definition of a (model based) monthly mean climatology. Lacking a long run from the model, we have used an 8-yr average (based on the years 1997–2005 model run). We noted that reasonable model results were obtained when we averaged the divergent wind anomaly over two successive months. These are shown in Figs. 16a–e for the pairs of months July–August, August–September, September–October, October–November, and November–December 1997. The results show that an Indian Ocean divergent inflow center at 200 mb during August–September 1997 over the Indian Ocean migrates over the Maritime Continent and over northern Australia. This feature is clearly seen during September–October and October–November 1997. Thereafter this feature is seen to migrate toward the southern Indian Ocean. These appear to be reasonable results, suggesting that coupled models can perhaps provide some guidance on the occurrence of such a persistent divergent inflow center in the upper troposphere that contributes to extreme dryness in the lower troposphere and possible forest fires. The SSTs over the Indian Ocean in the coupled forecast were generally warm (i.e., >28°C). We could not pin down a direct relationship of these descending lobes over Indonesia to any specific SST anomalies over the Indian Ocean.

Given that scenario, one could ask whether the same experiment extended further in time to May and June of 1998 (i.e., 13 and 14 months of forecasts with the coupled model) would provide any clues to the recent Florida fires. Figure 17 shows these results.

The fires over Florida caused considerable loss of property and forestland along the east coast of central and north Florida. This region was characterized by extremely warm and dry conditions, especially through the latter half of May and the entire month of June 1998. During the period 2 May–10 July 1998, some 2126 fires erupted in Florida, affecting 29 counties and burning close to 500 000 acres. The estimated damages totaled around$276 million, as stated by the U.S. Federal Emergency Management Agency. Most of this was attributed to continued dryness and lack of rain. The velocity potential anomaly (χ′) at 200 hPa again appeared to be a good indicator of the large-scale descent over this region. Thatwas apparent in the observed field of the χ′ during these months. We noted that a two-month averaging provided the best results for comparing the model-based χ′ and those based on observation. The large-scale convergent belt (emphasized by arrows) at 200 hPa implies a region of persistent descent. This region was located near 20°N over the western Atlantic Ocean in April and May, and moved very close to the Florida coast by May–June. The computations of χ shown in Fig. 17a bear a remarkable similarity to the observed field shown in Fig. 17b. Considering that these results are obtained some 13–14 months after the start time, they are quite encouraging. It suggests that given a detailed data assimilation, the coupled atmosphere–ocean model could perhaps provide useful information on the prediction of fires and possibly dry conditions (droughts) via the depiction of velocity potential anomalies. Furthermore, we feel that since areas of descent are much larger than areas of ascent, these dry regions may hold higher prediction skill. The predicted SST anomalies for June and July 1997 were examined over the Atlantic and the Gulf of Mexico. We noted a warm anomaly of 2°C over the central Gulf of Mexico during 1997 that weakened somewhat by July 1997. This SST anomaly may be important for the maintenance of large-scale descent over Florida and the western Atlantic Ocean.

e. North American monsoon rainfall patterns

Higgins et al. (1996, 1997) revealed a North American precipitation pattern for the summer season. This pattern appears to be excited by the Arizona–Mexico monsoon with three interesting lobes. A precipitation maximum extending northward from Mexico (with the heaviest rainfall amounts along the western foothills of the Sierra Madre Occidentale), a relative minimum of precipitation over the Great Plains of North America, and a lobe of rainfall maximum along the eastern United States. This pattern is best seen when one contrasts the monthly mean July rainfall (after the onset of the North American monsoon) with those of June (the month of the onset of rains). Figure 18 illustrates the nature of this pattern from a composite dataset based on Xie and Arkin (1996) and those of Higgins et al. (1996). The Xie and Arkin datasets shown in Fig. 18 cover the period 1987–95 and those of Higgins et al. (1996) cover the period 1963–94.

The coupled model simulation for 1997 and 1998 clearly shows this pattern. In Figs. 19a,b we show the July minus the June monthly rainfall differences. The dynamics of this North American summer rainfall pattern has not been clarified by any of the current research papers. There is evidence from 500-hPa heights that this is indeed a stationary wave pattern that would support heavier rainfalls over the western slopes of the southwestern and eastern United States and lesser rainfall amounts over the Great Plains. The climate model prediction of the monthly mean rainfall distribution for the summer months of 1997 and 1998 is shown in Figs. 20a–f. Here we shall focus on the North American monsoon rainfall pattern, that is, an axis of rainfall extending from west of the Sierra Madre Occidentale toward Arizona, a weaker rainfall belt to the east over the great plains, and an axis of heavier rain along the east coast of North America. The North American monsoon is usually weak in June (the onset month) and this pattern is expected to become more robust during the months July and August. The coupled model simulations for 1997 and 1998 show these features very clearly. Furthermore, we also note that this North American rainfall pattern for the El Niño year 1997 is somewhat weaker than that for the post–El Niño year 1998. Higgins et al. (1996) has noted this feature in his observational study. This is more clearly borne out in the coupled model predictions for July. Features seen here are not unlike those that are seen over the Asian summer monsoon. Yihui (1994) has noted a very similar Asian summer rainfall pattern where one sees periods of heavy rainfall over India coinciding with above-normal rainfall over northeastern China with below-normal rainfall distribution over western China and Indochina.

f. California rainfall

Some 10 months after the start of the integrations (i.e., January and February of 1998), the coupled model forecasts exhibited an active family of Pacific Ocean storms impinging on the North American west coast. The nature of these storms were much like those seen on observed weather maps during January and February 1998. The characteristic feature was the tropical oceanic moisture that entered frontal systems extending into the subtropical latitudes. Model rainfall contains the convective and nonconvective components. The extended troughs carried heavy convection rainfall with maximum rainfall amounts approaching 25–50 mm day⁻¹ rates. Figures 21a–d show sequences of coupled model-based rainfall charts illustrating the arrival of west coast storms. Storm A (Fig. 21a) appeared over the date line and near 30°N on 3 January 1998. This storm moves eastward at a phase speed of roughly 10° long per day, and is seen located near 40°N and 140°W by 6 January. The West Coast experiences heavy rain from the passage of this disturbance between 7 and 10 January. By 10 January most of California experiences heavy rain, although the maximum rainfall was located north of California.

System B (Fig. 21b) is noted at lower latitude near 30°N and 170°W on 12 January. This system affects most of California and the West Coast by 17 January and moves northward thereafter.

System C was noted near 40°N and 170°W on 12 February (Fig. 21c). This system moved eastward and produced rainfall along the west coast of North America by 17 February. The maximum of the coastal rainfall on 17 February occurred north of California. The maximum rainfall from this system was of the order 24 mm day⁻¹.

A system marked by the letter D (Fig. 21d) was noted at 30°N and 175°W on 6 February. This system moved eastward, like several others, and arrived at the west coast of North America on 10 February. By 11 February coastal rainfall maximum of the order 14 mm day⁻¹ was noted.

All of these systems show a bias on the location of the maximum rainfall toward the northern part of the West Coast. An examination of the predicted monthly mean rainfall clearly shows episodes of heavy rainfall during January and February 1998. At the resolution T42 these rain-producing systems are displaced somewhat to the north. In a future study we shall show that increase of resolution does correct this problem when we use a one-way nested regional spectral model within this lower-resolution global model. That information content of the global model (episodes of rain-producing system during January and February 1998) is an important ingredient for the success of the higher-resolution prediction. Thus we feel that the illustration of the California rainfall shown here is important.

Figures 22a–d show the variability of the daily rainfall totals during January and February 1998. Figure 22a shows those based on observations; this includes the daily history for San Francisco, Los Angeles, and San Diego, the city names are identified in the figure inset. Peaks in the observed maximum rainfall amounts vary roughly from 40 to 80 mm day⁻¹. The coupled model-based forecasts for the period are shown in Figs. 22b, 22c, and 22d for San Francisco, Los Angeles, and San Diego, respectively. Although the frequency of disturbance activity is similar to the observed passages, the maximum rainfall amounts are only of the order 5–15 mm day⁻¹. The predicted individual rainfall patterns cover larger areas compared to what are seen from Special Sensor Microwave/Imager. However, they seem to meet a reasonable overall global hydrological budget of E − P (evaporation minus precipitation), that is, of the orders of 2.75 mm day⁻¹ each. We have in fact noted that higher-resolution models are needed to enhance and sharpen the areas of these maximum amounts. This aspect is covered in a separate paper.