MERRA is a global atmospheric reanalysis that uses version 5.2 of the GEOS model and assimilation system. It utilizes a combination of 3D-Var assimilation and incremental analysis update to apply corrections from analysis to the forecast model. The MERRA system operates natively on a 0.5 × 0.667∘ latitude/longitude grid (361 × 540 grid points) and uses a hybrid sigma-pressure scheme with 72 vertical levels up to 0.01 hPa; the vertical resolution in the lower stratosphere is near 1 km. Further details about the MERRA system are given by . The MERRA data files available from NASA's Global Modeling and Assimilation Office (GMAO) are described by . Except where specified otherwise, the MERRA temperature data used here are from instantaneous daily files at 12:00 UT on the model levels and grid. However, the potential vorticity data from MERRA are available from GMAO only on a reduced 1 × 1.25∘ latitude/longitude grid (181 × 288 grid points) with 42 pressure levels up to 0.1 hPa; for the purposes of this study, MERRA PV is linearly interpolated to match the model levels and grid as was done in . Although these interpolations of the MERRA PV data cause some smoothing in the resulting PV fields, they preserve the strong PV gradients that define the polar vortex edge. All of the PV-based polar processing diagnostics we use depend strongly on the vortex edge and PV gradients (see Sect. 2.4.1), so these diagnostics are unlikely to be significantly affected by the errors introduced from interpolating MERRA PV to the model grid and levels.

ERA-Interim (hereinafter ERA-I) is another global atmospheric data assimilation system. The goal of the ERA-I project was to improve upon ECMWF's previous reanalysis, ERA-40, in advance of their planned next-generation reanalysis. It uses 12 h cycles of 4D-Var assimilation and a T255 spectral model with 60 vertical levels; the vertical resolution in the lower stratosphere is comparable to that of MERRA. The ERA-I system is described in detail by , and the data sets provided by ECMWF from the ERA-I archive are described by . Here, ERA-I data are used on the highest resolution regular latitude/longitude grid publicly available at 0.75 × 0.75∘ (241 × 480 grid points) on the 60 model vertical levels that go up to 0.1 hPa; this grid has spacing closest to that of the Gaussian grid associated with the spectral model. As with MERRA, unless stated otherwise, the ERA-I temperature and PV data used in this study are instantaneous at 12:00 UT. However, in this case the PV is derived from the provided relative vorticity, temperature, and pressure fields.

Since satellite observations are the primary constraint on reanalysis products at stratospheric levels, it is useful to consider how the data evolve with the introduction of new missions and instruments: noted the effect of the TOVS (Tiros Operational Vertical Sounder) to Advanced TOVS (ATOVS) transition in 1998 on middle and upper stratospheric global temperature anomalies from MERRA and ERA-I. At 5 hPa, he shows that ERA-I temperatures dropped suddenly by about 2 K, while in MERRA the mean annual cycle changed noticeably. Pawson asserts that these distinct discontinuities suggest that more work is needed to properly handle Stratospheric Sounding Unit radiances in reanalyses. mention that the introduction of COSMIC (Constellation Observing System for Meteorology, Ionosphere, and Climate) GPSRO (Global Positioning Satellite Radio Occultation) temperature data into ERA-I in 2006 caused a temperature shift in the tropics of about 0.5 K at 100 hPa. discuss in detail the various effects of the different satellite missions and instruments on ERA-I temperature data and perform intercomparisons with MERRA and the Japanese 55-year Reanalysis (JRA55).

In many of our diagnostics, we examine how well MERRA and ERA-I agree over the 1979 to 2013 time period. The observations assimilated in the reanalyses change dramatically over this period, and in some cases there are differences between MERRA and ERA-I in the timing of the changes and the observations included. Figure 1 shows a comparison of the primary satellite data sets assimilated in MERRA and ERA-I, compiled from information in , , and . (See Table 1 for the full names of the instruments/satellites listed in Fig. 1 and throughout this paper.) The colored regions indicate periods of years between which the data input streams change significantly; in other words, marking times when the differences between MERRA and ERA-I might be expected to shift. In this case, we have chosen boundaries at 1987 (when SSM/I was introduced), 1998 (when ATOVS was introduced), 2002 (inclusion of CHAMP in ERA-I, and AIRS and AMSU-A in both reanalyses), and 2007 (when COSMIC was included in ERA-I).

The diagnostics we use are designed to assess a wide range of conditions related to polar processing; many of them have been described previously . These diagnostics fall into three categories, focusing on assessment of temperatures, vortex characteristics, or air parcel histories. Taken together, the diagnostics used here provide a comprehensive evaluation of the meteorological conditions pertinent to chemical processing and ozone destruction in the polar stratosphere. Because our focus is on assessing the effects of reanalysis differences on studies of polar processing that takes place in the lower stratosphere, we focus in this paper on isentropic levels below about 600 K (approximately 25 km, or 30 hPa).

Figure 3 shows an example of the type of differences we calculate for most of the diagnostic intercomparisons (in this case Tmin) for 1979/80–2013/14. The orange line showing the differences for 2012/2013 provides an example of the magnitude of differences in an individual recent year. Note that there are large day-to-day variations in differences and that the differences are largest in December 2012 and early January 2013, when the polar vortex was unusually disturbed prior to/during a major sudden stratospheric warming (SSW) with a vortex split . The departure of the daily average differences (from 1979/80 to 2013/14; thick black line) from zero suggests that there exist persistent differences between the two reanalyses that dominate for much of the 34 years. Since these differences vary over the season, we define “monthly comparison period average differences” (monthly CPADs) between the data sets as the monthly means of the average daily differences over the comparison period years defined in Fig. 1. These quantities provide a compact means of summarizing the agreement between the data sets for a given month within a subset of years and of comparing the magnitude of differences between comparison periods for a particular diagnostic. We emphasize that monthly CPADs are meant to diagnose significant differences between the two data sets; they are not an assessment of the absolute accuracy of either. For the rest of this paper, we use the convention of subtracting ERA-I from MERRA (that is, MERRA minus ERA-I) to calculate the monthly comparison period average differences. Thus, differences in a diagnostic greater (less) than zero indicate that, on average, MERRA is greater (less) than ERA-I.

The seasonal progression of polar minimum temperatures provides an indication of when conditions favor the development of PSCs. The dependence of PSC formation on a temperature threshold implies that conclusions drawn from the Tmin diagnostic are most sensitive to differences at the beginning and end of the season when minimum temperatures first drop below or rise above PSC thresholds. For the Arctic, these periods are typically around the beginning of December and mid-March, respectively, but large interannual variability and the common occurrence of mid-winter SSWs can result in much earlier or later threshold dates in individual years and references therein. In the Antarctic, the threshold periods tend to be in the first half of May and in mid-October. Figure 4 shows the monthly CPADs between MERRA and ERA-I at the 580 K level (∼ 30 hPa, corresponding to an approximate TNAT of 193 K). For years preceding 2002, MERRA consistently has lower minimum temperatures than ERA-I. The Antarctic monthly CPADs in this period are particularly large, with differences between -5 and -6 K in some months, in contrast to the largest differences in the Arctic, which are about -1.4 K. Differences in 1998–2001, after the introduction of the ATOVS instruments but before the introduction of the Aqua instruments, are significantly reduced over those prior to 1998. In both hemispheres, there is a distinct shift in agreement after the introduction of the Aqua instruments from 2002 onward (likely due to the vast increase in the number of observations included by assimilating AIRS data (see, e.g., , and ). This is especially easy to see for the Antarctic, where the differences are reduced to values akin to those in the Arctic. In the Northern Hemisphere (NH), the shift marks the first period in which ERA-I minimum temperatures become consistently lower than those from MERRA. At lower levels down to about 460 K (not shown), the Tmin monthly CPADs are smaller in magnitude (e.g., at 490 K they are between -0.4 and 0.7 K in the Arctic and -2 to 2 K in the Antarctic) and have different seasonal variations, with ERA-I having comparatively more months in the first three time periods with lower minimum temperatures.

To examine the global variation of temperature differences between MERRA and ERA-I, Fig. 5 shows maps of mean temperature differences (averaged over the time periods of interest from Fig. 1) for the months with the largest Tmin monthly CPADs in Fig. 4. The maps for the Arctic, with the more symmetric color bar, clearly show that the regions where MERRA is colder tend to be mostly confined between ±90∘ longitude and poleward of 60∘ N. This is a preferred direction for the polar vortex and cold region to be shifted off the pole in Arctic winters e.g.,. Outside of this area, ERA-I temperatures are, for the most part, lower than those from MERRA. In the Southern Hemisphere (SH), the first three time periods show that the regions that are colder in MERRA are fairly symmetric poleward of the 60∘ S latitude circle. Examination of the mean temperature fields (not shown) for these periods suggests that larger differences are associated with lower temperatures. The maps for the following two time periods, from 2002 to 2013, demonstrate that the change seen in Fig. 4 reflects a large shift throughout the polar regions: the largest temperature differences during these periods are confined to relatively small regions, and overall the temperature differences lie between ±0.4 K in both hemispheres. show extratropical zonal-mean temperature differences between ERA-I and MERRA at 30 hPa that indicate that the extratropics (±20–90∘ latitude) are colder in ERA-I than in MERRA for most of the reanalysis period. The results shown here are generally consistent with that finding but suggest that this result does not always hold true in the polar regions (above 60∘ N and below 60∘ S) in winter/spring.

The number of days below PSC thresholds, a diagnostic derived from Tmin, for winters in the 1979–2013 period is shown at 490 K (∼ 56 hPa) for both hemispheres in Fig. 6. We show days below Tice, rather than TNAT, in the Antarctic because the periods with temperatures below TNAT and Tice are much longer than the periods below these temperature thresholds in the Arctic, making the differences in the SH more apparent for the lower (i.e., more sensitive) ice threshold. This diagnostic indicates the approximate duration of the period with conditions conducive to polar processing. Overall, the number of cold days from ERA-I is greater than the number from MERRA in both hemispheres at this level. The Antarctic differences before 2002 are quite large, with some years showing ERA-I having over 10 more days with temperatures below the ice PSC threshold than MERRA. At levels up to 580 K (not shown), the situation in the Antarctic is opposite; that is, MERRA has significantly more days (sometimes between 20 and 30 days in years before 2002) with temperatures below the ice PSC threshold than ERA-I does. These results, along with those discussed above for Tmin, suggest that not only may the choice of data set have a large influence on analysis and modeling of polar processes in the Antarctic for years preceding 2002 but also effects may vary qualitatively and quantitatively in the vertical. In contrast to those for the Antarctic, the magnitudes of differences in the Arctic at higher levels up to 580 K are largely similar to those at 490 K but with MERRA having more cold days than ERA-I.

Average values for the total area with temperatures below the NAT threshold at 580 K are displayed in Fig. 7. The Arctic maximum mean area is just above 3 % of the hemisphere for ERA-I, while MERRA reaches nearly 4 %. In a similar fashion, the Antarctic maximum mean area is about 11 % of the hemisphere for ERA-I while that for MERRA reaches nearly 12 %. Differences in these mean values between the reanalyses vary considerably with the season in both hemispheres. Figure 8 shows the corresponding monthly CPADs for ANAT. Here it is worth noting that the small differences in the last month of the season in each hemisphere are expected, as these include many days that have zero differences since minimum temperatures have risen above the NAT threshold in both reanalyses. Consistent with the Tmin monthly CPADs, there is much closer agreement after the first three periods. The relatively frequent occurrence of warm Arctic winters after 1998 e.g., suggests that the third comparison period for the NH might be less easily compared with the second comparison period which contained unusually cold Arctic winters (e.g.,, but the marked decreases of the monthly CPADs in the SH still indicate a substantial effect due to changes in assimilated observations. In most cases the monthly CPADs are positive, indicating that MERRA tends to have larger cold regions than ERA-I at this level in both hemispheres. For levels below 580 K, down into the upper troposphere/lower stratosphere region at 390 K, MERRA still tends to have larger cold regions than ERA-I, but the conditions are different: the differences are much smaller in all periods, generally lying between -1 and 1 % of a hemisphere in the Antarctic and -0.4 and 0.4 % of a hemisphere in the Arctic. Figure 9 demonstrates this behavior; it shows time series contour plots of the MERRA minus ERA-I ANAT differences averaged over the comparison periods from Fig. 1. Although the differences are smaller at lower levels, the same convergence towards better agreement we see at 580 K is not always seen at levels below 520 K, especially in the Arctic around 410 and 430 K.

Maximum PV gradients indicate the strength of the polar vortex as a barrier to transport and mixing of air from lower latitudes with the cold vortex air where chlorine activation takes place e.g.,. The seasonal evolution and mean values of MPVGs at 490 K are shown in Fig. 10. The primary difference between the hemispheres is seen in the average values: the Arctic maximum gradients tend to level off at around 10-5 s-1deg-1 early in the season, while the Antarctic maximum gradients steadily increase up to approximately 1.7 ×10-5 s-1deg-1 near the end of the season. The MPVG monthly CPADs are shown in Fig. 11. Note that the scaling of the values is 10-6 s-1deg-1. This means that a difference of 1 would be one-fifth the height of a grid box from Fig. 10. As such, the monthly CPADs are overall quite small, but the monthly and comparison period variations still provide useful information. For instance, the greater MERRA MPVGs in the months when the vortex usually weakens (March or April in the Arctic, November in the Antarctic) suggest that the polar vortices as represented by ERA-I tend to weaken earlier than those in MERRA. Furthermore, a shift in agreement contemporaneous with those seen in Figs. 4 and 8 is also present. In this case, however, the monthly CPADs increase considerably for the NH rather than decrease. This behavior, with agreement in the SH (NH) improving (degrading) slightly, is also seen at other vertical levels between 460 and 580 K (not shown).

Similar to APSC, the sunlit area of the vortex provides an approximate quantitative measure of the area on a given vertical level where chlorine-catalyzed ozone destruction can take place. Different sunlit area diagnostics have been used in detailed polar processing studies to establish correlations between the coverage e.g., and duration e.g., of sunlight and ozone loss; these diagnostics are usually somewhat computationally intensive, whereas the diagnostic used here (described in Sect. 2.4 above) is simple enough to compute for multiple long-term data sets. Figure 12 shows time series averages and ranges of the sunlit vortex area diagnostic as a percentage of a hemisphere at 490 K. The seasonal variations of the sunlit vortex are very similar in both data sets. Small differences can, however, be seen: the ERA-I Arctic polar vortex tends to be filled with slightly more sunlight than the MERRA vortex, while in the Antarctic the differences change over the season. Consistent with this, the sunlit vortex area monthly CPADs in Fig. 13 show predominantly negative values in the Arctic (in agreement with Fig. 12, which showed ERA-I values greater than those from MERRA) and differences that change sign in the Antarctic. Like the NH MPVG monthly CPADs, the NH sunlit vortex area monthly CPADs increase to a maximum by the last time period (2007–2013) of Fig. 1. The overall maximum monthly CPADs occur in November and February of this final time period, both months when a substantial portion of the vortex is typically in darkness and varying sunlight may affect chlorine activation whenever temperatures are low enough for PSC formation. In contrast to those for the NH, the SH monthly CPADs improve slightly over the reanalysis period. Differences of the small magnitude shown here are unlikely to have large effects on polar processing in either hemisphere, especially in the Antarctic, where the vortex is larger, colder, and less variable from year to year than in the Arctic. The vertical structure of the sunlit vortex area differences (not shown) is complicated, but for recent years the ERA-I NH polar vortex tends to have a larger area in sunlight than that in MERRA at levels between 460 and 550 K, while the opposite is true for the SH. The differences in sunlit vortex areas shown here primarily arise from corresponding differences between MERRA and ERA-I in the total vortex area (Avort, not shown). However, small discrepancies between the monthly CPADs of sunlit and the total vortex area indicate that differing vortex positions in the two reanalyses also affect the monthly CPADs slightly.

The winter mean vortex fraction of cold air diagnostic helps to identify years with conditions favorable for PSC formation and ozone loss. In the Arctic, years with low values of VPSC/Vvort correspond to the winters with very short cold periods usually associated with a disturbed vortex and midwinter SSWs. Because of its relationship to column ozone loss, VPSC has been used extensively in climatological ozone loss studies e.g.,. Vvort as a stand-alone diagnostic is much less common; however, the volume of air within the vortex is an indicator of the absolute size of the region over which chemical ozone loss can occur. Figures 14 and 15 show winter mean VPSC/Vvort for the Arctic and Antarctic, respectively, along with the corresponding winter mean VPSC and Vvort values separately. In general, MERRA tends to have a larger volume of cold air in both hemispheres, consistent with the larger ANAT monthly CPADs shown in Fig. 8 and the average differences shown in Fig. 9. The only major exceptions are the years from 1979 to 1987 in the Antarctic, where ERA-I Vice is a bit larger than that in MERRA. The lack of improvement in agreement for VPSC follows directly from the prior discussion of ANAT; since the agreement of ANAT between MERRA and ERA-I below 520 K generally does not improve much over time, neither does that of VPSC. Consistent with the differences discussed in the sunlit vortex area (Figs. 12 and 13), MERRA's Vvort values are very similar to those of ERA-I. Thus, on average, the differences in VPSC/Vvort between MERRA and ERA-I reflect the differences seen in VPSC. Exceptions to this are in 2007–2013 in the NH, when slightly smaller MERRA Vvort values emphasize the ∼ 0.01 larger VPSC/Vvort values seen in MERRA, and in 1979–1987 in the SH, when the ∼ 0.02 larger ERA-I VPSC/Vvort values represent larger cold fractions of smaller vortices. This analysis demonstrates several caveats for comparisons of diagnostics that rely on vertical integration, time averages, or combinations of both. Their dependence on time and/or altitude can lead to cancellations and smoothing when integrated and/or time averaged, making comparisons of the final results difficult to interpret since agreement (or lack thereof) can come about for the wrong reasons. In this case, the agreement of winter mean VPSC, Vvort, and VPSC/Vvort between the reanalyses was affected by the representation of several conditions. Because the differences between MERRA and ERA-I in APSC and Avort are not uniform in time and altitude (see Fig. 9), better agreement at some levels does not necessarily correlate with higher confidence in our knowledge of VPSC and consequently VPSC/Vvort. This suggests that these diagnostics function better as qualitative measures and argues for considerable caution in interpretation of their time variations or trends.

The concentricity of the vortex with regions of air with T ≤ TNAT (VTC) has not been widely used in polar processing studies. However, , using a concentricity diagnostic different from that defined here, found that different values led to dramatically different patterns of denitrification in model simulations of NAT particle growth and evaporation, because differences in the position of the cold region relative to the strong winds bounding the vortex can result in large differences in the amount of time air spends in cold regions of comparable size . Using idealized model simulations with a constant vortex field and cold regions ranging from highly concentric to non-concentric, showed that concentricity affected denitrification independent of any variations in the vortex itself. Since our concentricity calculations require both the polar vortex and a cold region to be defined, intercomparisons of the resulting discontinuous time series can be difficult to interpret. Therefore, we present this diagnostic as occurrence frequencies of the range of VTC values, normalized by the total number of days on which VTC was calculated for each reanalysis. Figure 16 shows the results at the 490 K level. The total number of days shown corresponds to the number of days with a valid VTC value summed over the years from the comparison time periods defined in Fig. 1 and used in the monthly CPAD plots. As evidenced by the small regions of non-overlapping colors, in most cases the ERA-I vortices spend more time at higher concentricity values (within the 0.8 to 1.0 range) in both hemispheres. Consistent with the number of days below PSC thresholds shown in Fig. 6, ERA-I also tends to have more valid VTC days than MERRA in both hemispheres. The detailed implications of differences in concentricity are difficult to assess, but the results of and suggest that greater concentricity in ERA-I could lead model runs driven by that reanalysis to simulate longer-lasting PSCs, greater denitrification, and enhanced chlorine activation.

Temperature histories along air parcel trajectories provide further information about polar processing potential beyond what can be obtained from the simple diagnostics described above. For these intercomparisons, we use isentropic trajectory calculations on the 490 K surface. As discussed by , although neglecting cross-isentropic motion would not be suitable for detailed polar processing studies, this simplification allows us to efficiently run and analyze a large number of parcels for extensive intercomparisons. Following , we consider only the parcels that are initialized in regions with temperatures less than 195 K. With this initial filter, we examine the total and continuous amounts of time that these parcels spend at temperatures below 195 K (TT195 and CT195, respectively) before and after the initialization date and the mean temperature of the parcels over the full run. The TT195 diagnostic captures the total exposure of an air mass to temperatures low enough for PSC formation and thus acts as a proxy for cumulative chlorine activation; CT195 better represents the potential lifetime of a PSC and is thus more directly relevant for denitrification, since sufficiently long continuous time at low temperatures allows PSC particles to grow to sizes large enough for sedimentation to occur . Here we show trajectory calculations initialized on 10 January 1996 and 24 January 2011 for the NH and 17 September 1988, 13 September 2002, and 25 May 2011 for the SH. These dates have been chosen to span a range of conditions representative of the interannual variability in both hemispheres and in some cases to compare with previous studies using the same diagnostics. Several other cases have been investigated for each hemisphere; the results shown here are representative.

Figure 17 summarizes the trajectory diagnostics for the NH. Overall, the parcel histograms for TT195 and CT195 are very similar between MERRA and ERA-I, with consistent distributions, peaks, and average values. The number of parcels included in the calculations indicate that MERRA has slightly larger regions of cold air, but the average values of the distributions only differ by approximately 0.05 days (∼ 1.2 h) for TT195 and 0.5 days (∼ 12 h) for CT195. Where there are small differences, they are consistent with minor spatial differences in maps of the parcels (not shown), which are also otherwise very similar. In investigating other NH winter initialization dates, we found that all agreed well (similar to the examples shown here) in each diagnostic for years as early as 1981 and as recent as 2013. The SH cases shown in Fig. 18 present a different picture. In particular, the mean parcel temperature plots reveal a significant improvement in agreement between MERRA and ERA-I over time. The 16 September 1988 case shows that MERRA parcels are consistently warmer, with temperatures roughly 1 K greater than those in ERA-I. The 13 September 2002 case also shows that MERRA parcels are noticeably warmer but to a lesser extent than in 1988. Although we highlight a different part of the SH winter season for 2011, there is a lack of a discernible temperature differences in comparison to the two previous dates, in agreement with the results from Figs. 4 and 5, which showed that temperature differences between MERRA and ERA-I are smallest from 2002 to 2013. Therefore, we consider the 25 May 2011 initialization date to be representative. The TT195 and CT195 histograms also reflect this evolution in the differences over the years, with the parcels on average spending fewer days below 195 K for MERRA than for ERA-I when the mean temperature of the MERRA parcels is higher.

The above results suggest two things. For the NH, isentropic trajectory runs initialized with MERRA or ERA-I should provide very similar results in the lower stratosphere, even for early years, thus lending confidence that transport calculations using winds from these two reanalyses should give comparable outcomes. The same cannot be said for early years in the SH, where many fewer observations were available in the 1980s and 1990s, allowing differences in the models and DAS to dominate the data constraints. This resulted in systematic differences in temperature histories that are consistent with the large monthly CPADs seen in direct temperature diagnostics (e.g., Figs. 4–9). Overall, the agreement of the trajectory diagnostics from MERRA and ERA-I in both hemispheres is much better than that from older analyses: in previous intercomparison studies, found very large differences in trajectory runs driven by fields from earlier analyses/reanalyses. Discrepancies in 465 K trajectories for average TT195 were as large in magnitude as 5 days for the NH and 7 days for the SH, with discrepancies in average CT195 in both hemispheres up to 2.5 days. For 10 January 1996 (shown here in Fig. 17), the time series of average parcel histories calculated by showed differences as large as 10 K on some days; in addition, the five analyses compared in that study showed qualitatively very different distributions of TT195 and CT195. Large qualitative differences were also seen between the analyses in September 2002 in contrast to the small differences and good qualitative agreement seen here in Fig. 18 (middle panels). The improvements in the agreement between MERRA and ERA-I over that between earlier analyses, which largely ingested the same data, demonstrate the degree of improvement in the models and DAS techniques over the past decade.