To analyze the spin-up characteristics of GRAPES_GFS2.3.1, the initial fields in F00, G21, and G00 are used to perform the integration, and the temporal variations of the average total WVT and TT at different heights from 00:00 to 12:00 UTC are calculated, as shown in Fig. 1. Seen from the figure, both the WVT and TT show sharp fluctuations at the initial stage of the integration in the three experiments, especially during the first hour. After 3–6 h of spin-up adjustment, the variation magnitudes of WVT and TT gradually become gentle, but the variation characteristics vary with different initial fields. At the early stage of the integration, the WVT is adjusted in F00 and G21, with the amplitude of -4.5 gkg-1d-1. In G21, the water vapor adjustment occurs in the lower layers of the model (850 and 925 hPa), while the WVT is relatively gentle without an obvious adjustment in the upper and middle layers (500 and 300 hPa). In F00, the water vapor adjustment occurs at the upper levels of the model at the early stage of integration. The WVT at 300 hPa can reach -4.5 gkg-1d-1, but it weakens immediately afterwards, probably due to the supersaturated water vapor in the initial field from FNL data. In F00, the WVT in the lower layers of the model is also significantly larger than that in G21. For example, at 850 hPa, the WVT in F00 maintains about 1 gkg-1d-1 for a relatively long time but in G21 mostly changes within 0.5 gkg-1d-1. The corresponding temperature adjustment processes in the two experiments present the same variation characteristics as the WVT adjustment. Therefore, the spin-up in the integration using the analysis field of GRAPES_GFS2.3.1 as the initial field is gentler than that using the FNL reanalysis data as the initial field.

In G21 and G00, the variations of both WVT and TT are very consistent, indicating that G00 has inherited the temperature and humidity structure of G21 well. However, G00 still needs to go through the spin-up during which a gradually stable adjustment process follows a sharp fluctuation at the early stage of integration; i.e., the dynamic and thermal adjustments are required to reach a statistical equilibrium state in the model. At the initial stage of integration in G00, the variation amplitudes of WVT and TT are smaller than those in G21, but greater than those in G21 after the 3 h integration. It shows that although G00 can retain the temperature and humidity structure of G21, the loss of cloud-field information in the operation still has a destructive effect on the model equilibrium state after 3 h adjustments. Based on the variation of TT, the spin-up time required for G00 is generally less than that for G21. It takes about 6 to 8 h to reach a TT equilibrium state in G21, but it is less than 6 h in G00.

In GRAPES_GFS2.3.1, the total temperature tendency of the model (ALL) is determined by dynamic core (DYN), radiation process (RAD), turbulent mixing in planetary boundary layer process (PBL), cumulus convection process (CONV) and cloud physical process (CLOUD). Among them, the total temperature tendency of all physical processes (PHY) is defined as the sum of the last four items (PHY = RAD + PBL + CONV + CLOUD). Likewise, the total water vapor tendency for ALL and PHY are same to those of temperature tendency except for the radiation process (RAD). Figure 2 shows the temporal variation of mean WVT due to dynamic and physical processes at different heights in F00, G21 and G00. In the middle and upper layers of the model (Fig. 2a and d), there is a drastic adjustment in the atmosphere at the early stage of the integration in F00. It may be due to the supersaturated water vapor in the initial field from FNL data, which causes the cloud to condense very quickly, and thus a relatively stable state is reached after three integration steps. At these levels in G21 (Fig. 2b and e), the total WVTs at the first few integration steps are slightly larger than those at the subsequent integration steps. The variations of the WVTs from dynamic core and turbulent mixing process in the planetary boundary layer are much less than those from the cumulus convection process and cloud physical process, and the latter two processes jointly determined the variation of WVTs at 300 and 500 hPa. There is not much difference in the dynamic field tendencies between G21 and F00. The magnitudes of the WVTs in the dynamic processes of the two experiments are also very close: around 0.5 gkg-1d-1 at 500 hPa and 0.25 gkg-1d-1 at 300 hPa. Therefore, the differences of the upper middle-level water vapor adjustments in the spin-ups between G21 and F00 are mainly caused by physical processes, and there is a good consistency in the dynamic process between the two experiments. At 925 hPa (the lower layer of the model), the total WVT stays around 1 gkg-1d-1 in F00 after three integration steps, humidifying the atmosphere. In G21, it reaches a relatively stable state after six integration steps, and water vapor decreases overall. As the WVTs of the dynamical processes in F00 and G21 have the same magnitude around 0.25 gkg-1d-1, the difference of the total WVT between G21 and F00 is mainly caused by physical processes. The effect of the boundary layer on the WVT is similar in both experiments, and the WVT is almost 3 gkg-1d-1. The greatest difference between the two experiments is mainly caused by the convection scheme. The convection in F00 is relatively gentle, and the WVT from convection is around -1 gkg-1d-1. In contrast, due to the strong dehumidification ability of convection in G21, the WVT is between -5 and -2.5 gkg-1d-1, which is significantly stronger than that in F00. At 925 hPa, the water vapor mainly decreases due to the strong convection process in G21. Such a significant difference in the convection processes between F00 and G21 may be related to the low-level temperature and humidity structures and the triggering conditions for convection. Meanwhile, it can be seen that the difference in the initial field of the model can significantly affect the physical processes.

In summary, in the middle and upper atmosphere, the fluctuation of WVT in G21 is weaker than that in F00, indicating the advantage of using the data assimilation cycling as the initial field. Both experiments quickly reach a quasi-equilibrium state after dramatic adjustments over several integration steps. The water vapor adjustment in spin-ups mainly occurs in the lower atmosphere of the model. The difference is mainly caused by different convection schemes. At the same time, different initial fields of the temperature and humidity structure may lead to a great difference in the dehumidification ability of convection. For G00 and G21, the WVTs of the dynamic and physical processes have roughly the same characteristics. At all of the three levels, the WVTs in G00 are slightly lower than those in G21.

In the middle and upper layers of the model, the dramatic change of the TT in F00 mainly occurs within the first half-hour of the integration (Fig. 3a and d). Among all the TTs at the first integration step, the cloud physical process leads to the largest one, followed by convection process, and they are related to the water vapor condensation process (Fig. 2a and d). For example, at 500 hPa, the global average heating produced by the cloud microphysical condensation process at the initial time can exceed 5 Kd-1, and it takes four integration steps to reach a relatively stable state. However, at this level, the TT caused by the convection process is 3 Kd-1, and it only needs one integration step with the drastic adjustment to get relatively stable. In addition, the TT caused by the dynamic process fluctuates greatly at the first half-hour of the integration. For example, at 300 hPa, the TT fluctuates between 1.1 and 1.5 Kd-1, and it requires extra 3 or 4 integration steps to reach a relatively stable state compared to the physical processes. Nevertheless, after half an hour of severe fluctuations, the TT caused by dynamic and physical processes tends to be relatively stable. Overall, the temperature increases by 0.25 to 0.5 Kd-1 in the middle and upper atmosphere in F00. Compared with that in middle and upper layers, the TT variation caused by the dynamic and physical processes in the lower layer of the model (Fig. 3g) shows a relatively small and rapid adjustment at the first integration step. However, no drastic adjustment is shown afterwards, and its variation is relatively stable. The TT of the convection process at 925 hPa in F00 varies between 1.5 and 2 Kd-1, which is mainly caused by condensing and dehumidifying of the atmosphere (Fig. 2g). Except for the cloud physical process, which has a relatively small positive tendency in the first four time steps, the TTs of dynamic core and other physical processes are all negative. Overall, in F00 the total atmospheric temperature is reduced with an amplitude of about -1.2 Kd-1 in the first hour of the integration at 925 hPa.

In G21, the TT in the middle and upper layers also experiences a dramatic adjustment in the first half-hour of the integration (Fig. 3b and e), and the main reason for the fluctuation is the dehumidification and heating in the convection process, which is different from that in F00 caused by the cloud physical process. The temperature increase caused by the convection process in G21 is 1 to 2.5 Kd-1, which is about twice that in F00. The TT caused by the cloud physical process in G21 varies relatively gently. Similar to F00, the TT caused by the dynamic process in G21 also shows obvious fluctuations, which may be caused by the drastic variations of physical processes. In the lower layer of 925 hPa (Fig. 3h), the positive TT in G21 is also caused by convective dehumidification and heating, while other processes lead to cooling. In terms of the total TT (dynamic core plus all physical processes), F00 has a cooling effect with a value of -1 Kd-1, while G21 has a warming effect with a value within 1 Kd-1. The temperature increase rate of G21 gradually decreases with the integration step.

The characteristics of the TT variation in G00 are consistent with those in G21 (Fig. 3c, f and i). In the first few time steps, G00 also has an adjustment process, with the adjustment amplitudes of TT close to half those in G21 at all levels. After half an hour, the temperature tends to be relatively stable. The TT variation in G00 indicates that although G21 has undergone a 3 h spin-up, G00 needs to undergo it again due to the loss of cloud-field information during the restart, and its fluctuation amplitude is not substantially smaller than that of G21.

The comprehensive adjustment effect of the dynamic and the physical processes on the water vapor and temperature in the numerical model can be presented by the cloud state. To reveal the dynamic and thermal adjustment processes in GRAPES_GFS2.3.1 system at the beginning of the integration and the time required for the model to reach the statistical equilibrium state (spin-up time), this section uses the total grid number of cloud (TGNC) in the model as the index for analyses. Although the cloud is changing locally, the total area covered by cloud can be regarded as a constant globally on average. Therefore, TGNC is used as the analysis index, and the model is considered to have completed the spin-up when the TGNC gets relatively stable. The total hydrometeors content (THC, THC = cloud water + raindrop + cloud ice + snow + graupel) greater than 1.0 × 10-4 gkg-1 in GRAPES_GFS2.3.1 is defined as the grid with cloud, and the TGNC at a global scale or a certain height is the sum of all the grids in the corresponding cloud area.

Figure 4 shows the vertical distributions of TGNC at different lead times in three experiments. It can be seen that regardless of whether the GRAPES_GFS2.3.1 model is cold-started with reanalysis data (F00, Fig. 4a) or warm-started with the 4D-VAR analysis field as the initial field (G21, Fig. 4b), the TGNC experiences rapid generation and growth during the 3 h after the beginning of integration in the two experiments, especially in the middle- and low-cloud regions below 300 hPa. After 3 h of integration, the TGNC grows relatively slowly, while after 6 h of integration, the TGNC becomes basically stable. However, the time required for the TGNC to reach the equilibrium state is slightly different at different heights. In F00, the integration time required for the TGNC to gradually reach the statistical equilibrium state below 850 hPa is 6 h. Note that the statistical equilibrium state is defined when the difference of TGNC with respect to the 24 h integration is insignificant (the difference is less than 20% of TGNC at 24 h). However, it takes 6–12 h for the TGNC to get stable and it completes the spin-up above 850 hPa. For G21, the TGNC of the middle and low cloud below 300 hPa needs 6 h to reach the statistical equilibrium state, while the TGNC of the high cloud above 300 hPa needs 6–12 h. It can be seen that the GRAPES_GFS2.3.1 using the analysis field from its own data assimilation cycling enables the cloud field in middle and upper layers to reach the equilibrium state earlier than that using FNL data for the cold start. In addition, GRAPES_GFS2.3.1 is gradually adjusted from the lower to the upper layers of the model to reach the equilibrium state, which is consistent with the evolution characteristics of the thermodynamic process in the troposphere. For the cloud above 500 hPa, the TGNC in F00 is significantly more than that in G21, which is related to a higher relative humidity of the initial field. Compared with G21, F00 has a wetter water vapor environment at the upper levels (Fig.6d), which tends to quickly condense the water vapor into more hydrometeors through the cloud scheme to eliminate supersaturated water vapor at the beginning of the integration (Fig. 2a). Thus, F00 has a higher hydrometeor content value and a wider distribution of cloud region (Fig. 5a and e), and its TGNCs are also larger than those of G21 at the upper layers.

In G00 (Fig. 4c), the growth of TGNC is found to be much slower than that in G21, especially for the TGNC of the middle and upper cloud. For example, at 3 h after the beginning of G00, the TGNC of the middle cloud is mostly between 15 and 20, while the TGNC in G21 can reach 25–30. The reason may be that the humidity and temperature fields of the model in G21 are already in a relative equilibrium state after 3 h spin-up. Meanwhile, as the restart of GRAPES_GFS2.3.1 has lost the cloud-field information (dotted light blue line) from the 3 h integration, the TGNC cannot reach the previous magnitude in the middle and upper layers even if it has been integrated for 24 h in G00 (Fig. 4c, solid purple line).

Figure 5 shows the distributions of THC at 400 hPa at different forecast time in the three experiments. It can be seen that the temporal variation characteristics of THC and its horizontal distribution at 400 hPa have consistent results with those shown in Fig. 4. In F00 and G21, as supersaturated water vapor is removed from the initial field, the cloud is quickly generated at the first integration step of the model. The THC rapidly increases within 1 h, and the cloud area with high hydrometeor content is constantly expanding. For example, at 1 h into the integration in F00, the THC in most areas of the Pacific Warm Pool is 0.2 gkg-1. With the further adjustment of the spin-up, the THC in this area gradually decreases and maintains a relatively equilibrium state after 6 h of integration. The variation characteristics of the THC in the storm track area (60–30∘ S) in the Southern Hemisphere are similar to those in the warm pool area but are less significant.

Experiments using the 4D-VAR analysis field to provide the initial field (Fig. 5e–h) show that the variation characteristics of THC at 400 hPa are generally consistent with those in F00. After the first integration step of GRAPES_GFS2.3.1, cloud areas are quickly generated in tropical and midlatitude areas. Due to the rapid development of convection processes in tropical areas, more cloud with THC of 0.0–0.05 gkg-1 appears. After 3 h of integration, the development of the cloud area gradually weakens. After 6 h of integration, the variations of the range and shape of the cloud area are no longer obvious, and it can be considered that a relatively equilibrium state is reached. From the view of absolute value of THC in the cloud area, although the difference in the distribution range of the cloud is insignificant, the THC in G21 is significantly less than that in F00 due to the different temperature and humidity conditions in their initial fields (Fig. 6).

Since G00 does not retain the cloud-field information after 3 h of integration in G21 (the THC in Fig. 5g), the model needs to undergo a new cloud-generation process when restarting the integration. However, as the dynamic and thermal fields are obtained after 3 h of adjustments in G21, the relative humidity has undergone a condensation process, making the atmosphere of G00 have a much weaker supersaturation at the initial time than that in G21. Therefore, unlike F00 (Fig. 5a) or G21 (Fig. 5e), in which large-scale cloud appears instantaneously, the cloud field in G00 can only be gradually generated by the dynamic and physical processes of the model. It can be seen from Fig. 5i–k that this process is relatively slow, and a relatively stable cloud distribution does not appear until 3 h after the integration. The cloud range in G00 at that time is smaller than that in G21, and it generally reaches the equilibrium state after 6 h of integration. The influence of slower generation and smaller range of the cloud in G00 on the model forecast results will be analyzed and explained in Sect. 3.2.

To reveal the reason why the TGNC (Fig. 4) and the THC (Fig. 5) in the upper layers of the model in F00 are significantly higher than those in G21, the difference of water vapor content and relative humidity at 400 hPa is analyzed, and the results are shown in Fig. 6. Figure 6c shows that the specific humidity in the initial field of F00 is generally higher than that of G21 in the tropical areas and the midlatitude and high-latitude areas of the Northern Hemisphere, especially in the tropical warm pool area where the difference is mostly over 0.2 gkg-1. The relative humidity reflects the degree of water vapor saturation. Figure 6d shows that the humidity of the initial field from the FNL reanalysis data is high relative to that from the 4D-VAR analysis field in the tropical warm pool, Intertropical Convergence Zone (ITCZ), and midlatitude and high-latitude areas at 400 hPa. This means that the water vapor is more likely to get saturated using the FNL reanalysis data as initial field. Thus, the cloud area is larger and the THC is higher at the beginning of the integration. It is not difficult to conclude that there are differences in the structure of atmospheric temperature and humidity among different initial field data, which significantly impacts the spin-up characteristics of the model and the cloud formation and development. It also suggests that we need to pay more attention to the analysis quality of water vapor in data assimilation (DA). It has been also confirmed by previous studies (Weygandt et al., 2002; Ge et al., 2013) that having an accurate moisture initial field by DA is an effective way to improve the forecast performance of supercell storms in numerical weather prediction models.

Figure 7 shows the zonal mean distributions of averaged column cloud water content (CCWC), the outgoing longwave (OLR) at the atmosphere top and the downward longwave at ground (GDLW) level simulated by G21 and G00 from 00:00 to 03:00 UTC on 9 August 2019, as well as the distributions of difference between them. It can be seen from Fig. 7a that the total zonal-averaged CCWC forecasted in G00 is systematically smaller than that forecasted in G21. The areas with smaller CCWC are mainly located in the Southern Hemisphere storm track, tropical low-latitude areas, and midlatitude and high-latitude areas in the Northern Hemisphere with active cloud. Among them, the area with the smallest CCWC is the active area of Southern Hemisphere storm track, with the CCWC difference reaching 240 gm-2, and there are also some areas with the CCWC difference over 200 gm-2 in the Northern Hemisphere. From the OLR and GDLW predicted in the two experiments, it can be seen that the OLR predicted in G00 is systematically larger than that in G21, with the maximum bias (20 Wm2s-1) appearing in the Southern Hemisphere storm track. This is due to the interaction between clouds and radiation, as well as the underestimation of the CCWC. In terms of GDLW, the reduced CCWC weakens the atmospheric warming effect, resulting in systematically smaller GDLW in G00 than in G21. In most areas, the GDLW is smaller than the observation by over 10 Wm-2s-1, and the regions with the largest bias are the midlatitude and high-latitude areas of the Southern Hemisphere and high-latitude areas of the Northern Hemisphere.

The change in the calculation of the radiation flux induced by cloud would seriously affect the atmospheric temperature field and geopotential height field. Figure 8 shows the difference distributions of the 500 hPa temperature field and the geopotential height field at four lead time between G00 and G21. It can be found that as there is less hydrometeor in the cloud in G00 than in G21, the temperature field in G00 at different forecast times shows a systematic warming of more than 0.1 K in the tropical low-latitude and midlatitude and high-latitude areas with active cloud. With the increase of the lead time, the warming area is expanding and the degree of warming gradually increases. For example, after 72 h of integration, the warming in many areas is larger than 0.2 K, and it can reach 0.5 K in some areas. Systematic biases also appear in the corresponding geopotential height field. Compared with those in G21, the geopotential height fields in G00 have also systematic positive biases. For example, in the first 24 h of integration, the systematic biases in the geopotential height field are above 0.5 gpm, and the positive bias can exceed 1 gpm in areas with active cloud. After 72 h of integration, the geopotential height field in the tropical area still shows a systematic positive bias, while in the midlatitude and high-latitude areas, the bias of the geopotential height field shows the structure with an alternation of positive biases and negative biases due to the biases of the weather system location predicted in the two experiments, but in most areas the forecast fields are still higher than the observation.

This section analyzes the biases of the cloud field, precipitation field, and the track of the super typhoon Lekima (no. 1909) and typhoon “Krosa” (No. 1910) in 2019 during the forecast period to evaluate the impact of the lost hydrometeor information on typhoon forecast operation in GRAPES_GFS2.3.1. During the forecast, Lekima and Krosa appear as double typhoons in the western Pacific. Lekima made landfall in northern China, while Krosa remained offshore. Since the conclusions for both Lekima and Krosa are the same, only Lekima will be presented in this study. Here, we show the impact on the cloud and precipitation of Lekima by the lost hydrometeor information on typhoon forecast operation of GRAPES_GFS2.3.1. In the last part, the path-forecast biases for the two typhoons are both given.

Figure 9 shows the evolutions of the averaged CCWC and column cloud ice content (CCIC) within the main cloud area of Lekima (22–34∘ N, 117–130∘ E) simulated in G00 and G21 from 00:00 UTC on 9 August to 00:00 UTC on 10 August 2019. It can be seen that the CCIC predicted in G00 at the early stage of integration is obviously underestimated. The averaged CCIC values in G21 are maintained within 850–1000 gm-2 from 00:00 to 09:00 UTC on 9 August, while the CCIC is only 480 gm-2 at the initial time of G00. G00 needs to restart the spin-up. During the spin-up, the CCIC predicted in G00 increases rapidly, with the greatest increase during 00:00 to 06:00 UTC. After 3 h of the integration, the CCIC increases rapidly from 480 to 820 gm-2. After 6 h of integration, the CCIC is close to 900 gm-2. In G00, the CCIC is not as large as that in G21 until 9 h after the beginning of integration.

Figure 10 shows the difference distributions of both 3 and 24 h accumulated precipitation (since 00:00 UTC 8 August 2019) of Lekima between forecasts of G00 and G21.The most significant difference of the 3 h cumulated precipitation appears within the first 3 h of integration in G00. The 00:00–03:00 UTC precipitation forecasted in G00 presents a systematic underestimation when compared with G21, and the biases are all above 1 mm. The precipitation bias in the center of Lekima can even exceed 5 mm (Fig. 10a). As shown in Fig. 9, after 3 h of adjustments, the total CWP and CCIC in the typhoon system in G00 grows rapidly and gets close to the magnitudes in G21. Therefore, the difference of the 3 h precipitation between forecasts of G00 and G21 is no longer significant during 03:00–06:00 and 06:00–09:00 UTC, and there is no more systematic bias (Fig. 10b and c). The phase differences of the weather system lead to the structure with alternating positive biases and negative biases for the precipitation difference.

It can be found from Fig. 10d that the lack of cloud-field information has a significant impact on the simulation of the accumulated precipitation in the first 24 h of Lekima. The negative biases dominate the central area of the typhoon; i.e., there is an underestimation of precipitation with the maximum bias of 5–10 mm. In contrast, in the spiral cloud zone around the typhoon, there is a structure with an alternation of positive and negative biases, which is related to the location bias of the weather system simulated in the two experiments in this area.

Figure 11 shows the forecast track evolution of Lekima and Krosa in G00 and G21 within the lead time of 72 h. Overall, G21 performs better than G00 in predicting the tracks of these two typhoons, and there are different characteristics for the track forecast biases of the two different typhoons. Lekima landed on the coast of Chengnan Town, Wenling City, Zhejiang Province, at 15:45 UTC on 9 August 2019. There is not much difference in the biases of the track forecast between G00 and G21 before the Lekima landing. In contrast, the biases appear to be different after the landfall (16:00 UTC), and the track forecast in G21 is slightly better than that in G00 around the landfall. After the landfall, the track biases change continuously during the 27th to 42th hour and 54th to 60th hour of the forecast, the track bias in G21 is smaller than that in G20. The maximum difference between the two track forecasts can reach 32 km. From the 65th to 72th hour, the forecast track bias in G21 is slightly larger. For Krosa, during the first 42 h, the biases of the tracks forecasted in G00 and G21 are not much different. But the forecast tracks of the two become different after the 42th hour, with the track bias in G00 becoming larger. In most forecasts after the 42th hour, the track biases in G00 are over 20 km and larger than those in G21. The abrupt track difference after 42th hour is most likely caused by the continuous accumulation of the direct cloud-radiation process and the systematic temperature bias in the typhoon peripheral cloud area during the re-undergone spin-up of G00 experiment, along with their impacts on the typhoon eye (track) through dynamic processes with the model integration.

Overall, G21 performs better than G00 in the track forecasts of Lekima and Krosa within the lead time of 72 h, especially in the forecast of Krosa. For Krosa, the forecast track on the ocean is less affected by other factors, so the forecast track biases at the later stage of the forecast are significantly smaller. It shows that GRAPES_GFS2.3.1 performs better in continuous-integration forecasts, and the interruption in the operation is destructive to the typhoon track forecast.