Sensitivity of precipitation in the highlands and lowlands of Peru to physics parameterization options in WRFV3.8.1

. The performance of the Weather Research and Forecasting (WRF) model version 3.8.1 at convection-permitting scale is evaluated by means of several sensitivity simulations over southern Peru down to a grid resolution of 1 km, whereby the main focus is on the domain with 5 km horizontal resolution. Different conﬁgurations of microphysics, cumulus, longwave radiation, and planetary boundary layer schemes are tested. For the year 2008, the simulated precipitation amounts and patterns are compared to gridded observational data sets and weather station data gathered from Peru, Bolivia, and Brazil. The temporal correlation of simulated monthly accumulated precipitation against in situ and gridded observational data show that the most challenging regions for WRF are the slopes along both sides of the Andes, i.e. elevations between 1000 and 3000 m above sea level. The pattern correlation analysis be-tween simulated precipitation and station data suggests that all tested WRF setups perform rather poorly along the north-eastern

product provides half-hourly precipitation with a spatial resolution of 0.1 • (approximately 10 km). Version 6 is employed in this study, which is available for the period 2000-present. 195 CHIRPS is a high-resolution precipitation data set that covers the area 50 • N to 50 • S (Funk et al., 2015). The daily precipitation amounts are available at a 0.05 • spatial resolution for the period 1981-present. As the previous data sets, it combines satellite data with the World Meteorological Organization's Global Telecommunications System (GTS) rain gauge data.
PISCO Version 2.1 provides land-only daily precipitation amounts estimated for the entire country of Peru at a spatial resolution of 0.1 • for the period 1981-2018. This precipitation data set combines radar and comparably dense gauge measurements 200 (441 quality controlled stations) maintained by SENAMHI with the CHIRPS data set. The performance of this product is evaluated against independent weather stations from those used to develop PISCO, and the coast and the western slopes of the Andes showed the best scores (Aybar et al., 2020). These regions coincide with areas covered by the highest weather station density.

Temporal analysis of precipitation
To evaluate the performance of the different setups of the WRF model, we start by analyzing the annual cycle according to the monthly precipitation sums of year 2008. Spearman correlation coefficients (Fig. 3) and root-mean-square error (RMSE;  Schulzweida, 2019). For the maps, all the gridded data are bi-linearly interpolated to the original grid of PISCO before calculating the two statistics between the gridded data (observation based and WRF) and the PISCO data set.
PISCO is used as reference because it is generated particularly for Peru by combining weather station data with satellite data (Sect. 2.2). Box and whisker plots are estimated for different regions (Sect. 2.2): SW or NE flatlands, SW or NE slopes, and 215 the plateau.
The temporal correlation coefficients in the box and whisker plots show that the area with the highest correlation is the plateau, where the median values of all the correlation coefficients are above 0.8 (Fig. 3a). PISCO shows the highest median value with the smallest spread. The good agreement between the weather station data and PISCO confirms the good quality of the latter in the plateau, related to the available dense station network. The high temporal correlations between the weather 220 station data and the simulations with the different parameterization options are confirmed by the correlation maps against PISCO ( Fig. 3f-l).
The box and whisker plots reveal also a high temporal correlation in the two slope regions (median r > 0.75). For the SW slopes (Fig. 3b) the distribution of the correlation coefficients is rather homogeneous, with a slightly lower median for Micro13, however, not significantly different to the others. This is reversed in the NE slopes (Fig. 3c) slightly lower correlation coefficients. High temporal correlations between PISCO and all other gridded data  are also observed in the slopes at both flanks of the Andes, depicted in the maps. An exception is the region to the south in the SW slopes where the correlation coefficients with respect to PISCO are close to zero for all other gridded data sets. The fact that CHIRPS, IMERG and the sensitivity simulations correlate better with the weather station data in this southern part implies that 230 the PISCO data set is not fully able to realistically represent observed precipitation in this region.
The lowest temporal correlation coefficients in the box and whisker plots are found for the flatlands. In the case of the SW flatlands ( Fig. 3d), the correlation coefficients between the weather station data and the WRF simulations are in the range of those obtained by the gridded observational data sets, whereby Kenya and No Cumulus show the worst performance. The correlation coefficients of Micro13 with respect to the weather station data are comparable to the ones obtained by CHIRPS or 235 IMERG. Europe and South America show the best performance in the SW flatlands, i.e., in an extremely dry area, because these two parameterization options simulate rather scarce precipitation throughout the entire year (see Sect. 3.3 for more details on precipitation amounts). The maps of temporal correlation between PISCO and all other gridded data show negative correlations at the Pacific coast of Peru for all WRF setups, while IMERG and CHIRPS compare well to . This is expected as the observation based data sets are not independent from each other.

240
The temporal correlation coefficients in the NE flatlands plotted as boxes and whiskers (Fig. 3e) show a rather homogeneous performance of the gridded observational data sets, where PISCO obtains the lowest median and the largest spread. This poor performance for PISCO is already pointed out by Aybar et al. (2020). Among the parameterization options, Kenya and Micro13 show the lowest medians of the correlation coefficients. However, only 15 stations were available in this large area, which means that the evaluation of the parameterization options against the weather stations or the gridded observational data 245 sets over the NE flatlands must be carried out with caution. In the NE flatlands, all parameterization options show a rather well correlated area in the southeastern part of the region, but this temporal correlation is lost towards the northwest (Fig. 3f, h, i, k and l). Kenya and Micro13 even show negative correlations in the northern part of the domain. However, Micro13 and South America are the only parameterization options that show good temporal correlations over Madre de Dios. A good temporal correlation between the gridded observational data sets and PISCO is also highlighted in the flatlands ( Fig. 3g and j). This 250 good agreement between observational data sets is expected as they are not fully independent from each other, particularly in the regions where the number of available weather stations is small. This is particularly true for the Brazilian part of PISCO, as only stations from Peru are considered for the creation of PISCO. Consequently, this area fully depends on the information provided by the gridded observational data sets.
The plateau (Fig. 4a), the NE slopes (Fig. 4c) and the NE flatlands ( Fig. 4e) are the three regions where the parameterization 255 options show the largest RMSE against weather station data, particularly the Micro13 and South America simulations. It is noteworthy that PISCO shows a larger RMSE than IMERG and CHIRPS in the NE flatlands and particularly in the NE slopes.
However, these are the regions where CHIRPS and IMERG are more precise than PISCO, as the median RMSE of PISCO against weather station data is higher and the spread is largest. This misrepresentation of PISCO might be related to the scarce weather station availability and the respective miscorrection of the satellite data in that area. This might lead to a loss in the 260 precipitation related to the complex atmospheric dynamics along the NE slopes and flatlands. The pattern of the RMSE against PISCO is rather similar in all the parameterization options and gridded observational data sets, highlighting the largest values in the NE flatlands and slopes of the Andes (Fig. 4f-l). The lowest RMSE for the parameterization options are shown at the Pacific coast of Peru (Fig. 4d), which are similar to those obtained by PISCO. The RMSEs shown by ERA5 and the other observational data sets are larger than the ones for the parameterization options. This agrees with the RMSE maps against 265 PISCO, because precipitation amounts are very small during the entire year.
WRF is able to maintain its skill in terms of temporal correlations and RMSE also at finer temporal resolutions (not shown).
The correlations and RMSEs of all parameterization options and gridded observational data sets show rather similar results for intervals of a months, 15-days and 10-days, independently of the region. For 5-day and daily intervals the values drop for the correlations and rise for RMSEs. The increase in the RMSEs and the reduction in the correlations are expected due to the fact 270 that capturing the exact amounts of precipitation at the same time as the observations is rather challenging for the model. The differences between the parameterization options, ERA5 and the gridded observational data sets considered are reproduced at finer temporal resolutions as seen in Figs. 3a-e and 4a-e.
In summary, the differences in parameterization options with respect to correlation coefficients and RMSE are rather small.
Nevertheless, Micro13 clearly stands out in the NE slopes, while it fails to capture the proper correlations in the SW slopes.

275
In the SW flatlands Europe and South America show the best results, as they generally produce little precipitation amounts, which matches for the driest region of the country. While in the plateau the performance is overall very good, related also to the high station density, the opposite is true for the NE flatlands. Over Madre de Dios, only Micro13 and South America are able to keep high correlation values.

Pattern correlation analysis of precipitation 280
Since the temporal correlation analysis does not clearly constrain an optimal parameterization option over the entire region, we use the an additional measure, i.e., pattern correlation. The pattern correlations are calculated between the different parameterization options and the weather stations and the gridded observational data sets, separately. Due to the large RMSE and poor correlation of PISCO in the NE parts of the domain, which is in agreement with Aybar et al. (2020), CHIRPS is selected as reference in those areas, while PISCO is considered for the plateau and the SW slopes and flatlands.

285
The monthly pattern correlations between the parameterization options and the reference data sets, i.e., PISCO or CHIRPS depending on the region, are shown in Figure 5. The pattern correlation value of 0.5 against stations and 0.4 against the gridded observational data sets are considered as moderate correlation values for each case, as they explain approximately a quarter of the variance (25 %). For the calculation of the pattern correlations against the weather station data, all the gridded observational data sets are bi-linearly interpolated to the grid of the second domain of the WRF simulations using CDO. Conversely, for the 290 pattern correlations against PISCO, the remaining gridded observational data sets and all the parameterization options are bilinearly interpolated to the grid of PISCO. For the correlation against CHIRPS the same procedure is applied, but taking its grid as a reference for the interpolation.
The gridded observational data sets do not always agree well with the weather station data in terms of the spatial pattern of monthly precipitation sums, particularly in the NE slopes of the Andes, the SW flatlands and the dry months in the SW 295 on the rather dry conditions, except for Micro13 and CHIRPS that show wetter patterns. The main differences in precipitation amounts between both observational data sets and the different parameterization options are identified in the NE slopes of 365 the Andes. However, the maxima are located in the same areas as in CHIRPS. Micro13 is the wettest simulation in that area, followed by Kenya, South America and No Cumulus simulations. Note that only Micro13 is able to produce similar patterns and amounts of precipitation in the flatlands compared to PISCO, which in this month, shows the highest pattern correlations against weather station data (Fig. 5). The wide agreement of the Micro13 simulation with the weather station data in the NE flatlands is also highlighted by the high pattern correlation in Fig. 5. As in February, Europe is again too dry over the entire 370 domain compared to PISCO and especially CHIRPS. Similar results are shown in the transect through the Andes towards Madre de Dios. There, the excess of precipitation of Micro13 is highlighted, and also the effect of a finer representation of the topography in precipitation for the parameterization options compared to the observational data sets. Again, both CHIRPS and PISCO show the precipitation peak over the lower part of the slopes or the flatlands.
Altogether, the parameterization options show different behaviours over the entire domain and seasons. In the wet season 375 the Europe parameterization option is too dry, which applies to some extent to Kenya as well, particularly in the rainforest.
Micro13, South America and No Cumulus provide reasonable patterns and amounts for the rainy season, except for the last which overestimates the amounts. As expected, the South America parameterization option simulates precipitation reasonably in the NE slopes of the Andes, since this configuration is designed to capture storms in this area. In the dry season, all the parameterization options simulate too dry conditions. Micro13 must be excluded from this, as its precipitation patterns and 380 amounts are similar to the ones shown by PISCO and CHIRPS, rendering Micro13 as the optimal setting for Madre de Dios.

Seasonal cycle over the northeastern flatlands
To investigate the reason for the different performances of the parameterization options over the northeastern flatlands, which covers the region of interest, field means of different variables are analyzed by means of a seasonal cycle and daily cycles.
The Europe parameterization option simulates especially low monthly precipitation sums in the wet seasons, while the other 385 parameterization options show almost no precipitation during the dry season (Fig. 8a). The Micro13 parameterization option simulates more precipitation than the others in all months as already shown in Fig. 6 and 7. The same is also true for the precipitable water, i.e., the vertically integrated water vapour content of the whole atmospheric column. Here, the difference between the Micro13 parameterization option and the other options is evident in the first half of the year (Fig. 8c). The combination of cooler 2-meter temperatures (Fig. 8b) and a higher water vapour content results in a higher relative humidity at 390 2-meters, especially in July, but also during the first half of the year. Even though the average 2-meter temperature of the Europe option is similarly low as in Micro13, the relative humidity is not higher compared to the others, even if the water vapour content is similar. The reason for this might lay in the daily cycle of 2-meter temperature (Fig. 9). While the monthly average is similar to the Micro13 option, the daily cycle looks completely different. The temperature is lower than for the Micro13 option for Europe, while during the day the temperature is much higher and hence, the relative humidity decreases on average compared to 395 the Micro13 option. The relative humidity and the precipitable water of the No Cumulus parameterization option is especially low, even though the precipitation is comparable to the other options, which might indicate that this parameterization option has 12 https://doi.org/10.5194/gmd-2021-307 Preprint. Discussion started: 23 November 2021 c Author(s) 2021. CC BY 4.0 License. an efficient process to remove moisture from the atmosphere, i.e., convective processes. This is also supported by the fact that precipitation occurs mainly in the afternoon, while the other options have precipitation distributed over the whole day (Fig. 9).
A striking difference between the Micro13 and the other parameterization options is the fraction of mid-level clouds and the 400 soil moisture content. A high cloud cover allows for more precipitation, but also to reduce atmospheric temperature, which is especially true for a high fraction of high-level clouds (not shown). The difference in soil moisture is induced by the higher precipitation amounts, which then allows to fuel the water vapour content in the atmosphere, leading to more precipitation and supporting a positive feedback. Hence, for the simulation of the convection permitting domain, the change in the microphysics parameterization results in the largest changes of precipitation and related processes. year 2008 already shows that this option is too dry during the entire year. The temporal correlation of monthly precipitation sums against weather station data shows that, in general terms, the performance of the parameterization options under wetter 410 conditions is rather similar to that of year 2008 (Fig. S1a). The overall performance of the parameterization options in terms of temporal correlations is improved in the SW slopes, the plateau and the NE flatlands, while the performance is slightly worse in the SW flatlands and the NE slopes. The RMSE of all the different parameterization options and observational gridded data sets is slightly larger than in 2008 in the NE parts of the domain (Fig. S1b). CHIRPS should be still preferred as a reference in the NE slopes of the Andes in 2012, but not anymore in the NE flatlands since PISCO shows higher correlations and lower RMSEs.

415
The poor agreement between the gridded observational data sets and weather station data in terms of the spatial patterns of monthly precipitation sums in the NE slopes and the SW flatlands is also observable in 2012 (Fig. S2). Micro13 and Kenya still remain the best parameterization options for the NE part of the domain in 2012. The same is true for the No Cumulus option in the SW part of the Andes.
Micro13 captures the precipitation patterns relatively well in both rainy and dry seasons of 2012. In the rainy season, i.e., in 420 February (Fig. S3a), Micro13 is a bit too dry in the northern part of the domain. Kenya shows a larger deficit of precipitation over the same area. Note that February 2012 is an anomalously wet month compared to the climatology of 1981-2010 (see Fig. 2), which highlights the ability of Micro13 to even simulate correctly anomalous monthly precipitation amounts. During the dry season, i.e., in July (Fig. S3b), Kenya and No Cumulus show rather dry conditions over the Amazon, and only Micro13 shows comparable amounts to those in PISCO (and CHIRPS -not shown).

4 Summary and Conclusions
This study aims at determining the optimum setup for WRF to accurately simulate the observed precipitation patterns and amounts over the Amazon basin in southeastern Peru. The region of interest is the entire department of Madre de Dios, but because of the lack of a dense network of weather stations in that area we evaluate the performance of the model over a broader the large-scale atmospheric circulation controls and the complex topography of the region renders this part of the world a challenge for regional climate models. The novelty of this study is that this is one of the first times that such a complex region as southern Peru is resolved at a convection permitting spatial resolution down to 1 km, and consequently, several physics parameterizations must be tested in the regional climate model, which is WRF for this study. We tested different combinations First, the gridded observational data sets are compared to the weather station data. The following characteristics for the gridded observational data sets over southern Peru are identified:

445
-The temporal analysis over the monthly data shows that the agreement between the gridded observational data sets and the weather station data is high. This is true for both evaluated indices, Spearman correlation and RMSE. In the case of the ERA5 reanalysis, this is only true for the Spearman correlations, as its RMSEs are larger than the ones from the observational gridded data sets.
-The monthly pattern correlation analysis shows that all the observational data sets perform poorly over the NE slopes of 450 the Andes, and also in the SW slopes during the dry months. Additionally, the pattern correlations of TRMM, IMERG and CHIRPS are deficient at the Pacific coast from May onward.
-The quality and the station density of the rain gauge measurements over the plateau is high enough to validate the sensitivity simulations, but not necessarily in the remaining regions.
-The lack of a dense weather station network over the Amazon reduces the quality of PISCO in the NE slopes and 455 flatlands. Instead, CHIRPS should be preferred in these regions, which is in line with Aybar et al. (2020).
After evaluating the weather station and gridded observational data and identifying which data set must be used as reference in which region, the WRF simulations are then tested. By comparing the WRF simulations to the chosen data sets the following

465
-The monthly pattern correlation exhibits the poorest performance in the coastal region of Peru. Even though the general pattern is not well captured by the simulations, the scarce precipitation amounts are correctly simulated locally. The first months stand out as particularly poor which also applies to the Amazonian flatlands and the plateau, which supports the idea that a longer spin-up is needed. Nevertheless, no systematic improvements were observed when four months instead of two where employed as spin-up (not shown).

470
-The precipitation patterns in the wet season are captured reasonably well by all parameterization options, but the amounts are clearly underestimated by the Europe option and overestimated by No Cumulus. Kenya shows an underestimation in the NE flatlands, while Micro13 simulates too much precipitation in the NE slopes. South America performs well, as this option is designed to capture storms in the NE part of the domain.
-For the dry season, most of the parameterization options simulate too dry conditions. This includes the Europe, South 475 America, Kenya and No Cumulus parameterization options. The last two are able to correct the dryness in the NE flatland by some extent. Micro13 provides good precipitation results on the whole domain, except for an overestimation in the NE slopes.
-The change in the microphysics parameterization from the WRF single-moment 6-class scheme to the Stony-Brook University scheme causes a great improvement in the representation of precipitation over the Amazon, and consequently, 480 reinforces a positive feedback between the soil moisture, temperature, relative humidity, cloud cover and finally, precipitation.
-These results indicate that the configuration over Kenya is not optimal for another equatorial region as Peru, and supports the fact that parameterization options are not fully interchangeable, as also found for example by Takle et al. (2007), Jacob et al. (2012 and Russo et al. (2020).

485
-The transects through the Andes into the rainforest of Madre de Dios suggest that the spatial resolution of WRF is necessary to capture correctly the precipitation associated to the terrain, while this is not the case for PISCO or CHIRPS.
Thus, the biases in the slopes of Micro13, particularly against PISCO, may be related to the lack of high spatial resolution in this complex region.
Hence, we conclude that Micro13 should be the preferred setting for southeastern Peru in WRFV3.8.1, and particularly for    Spearman correlation coefficient with respect to the corresponding weather station. Note that the WRF domain is reduced at the eastern side to match the PISCO data set (refer to Fig. 1 for more details). The grey shading in panels (a) to (e) denotes the satellite based or reanalysis data, and facilitates the separation from the WRF simulations.    Figure 7. Same as Fig. 6, but for July, which represents the dry season.