‘Subseasonal-to-seasonal (S2S) forecasts with CNRM-CM : a case study on the July 2015 West-European heat wave’, Advances in

Abstract. An intense heat wave struck West Europe in early July 2015. The degree of anticipation of that event is assessed through the new CNRM near-real time subseasonal to seasonal forecast system. A warm anomaly over France was detected for the first week of July in all the successive forecasts issued in June, even up to one month ahead. On the other hand, the positive 500 hPa geopotential anomaly observed during that period was little anticipated. Despite the limited skill of the forecast system beyond twelve days, the relatively successful anticipation of that event pleads for a predictability study based on a multi-system assessment.

The following should cover the entire project duration. (10 lines max) The main objectives of this special project are to assess beyond conclusions from the FP7-SPECS project the impact of the initialization of land-surface and sea-ice components of the CNRM-CM model on sub-seasonal and seasonal predictability. Two main questions are addressed: -What is the extent of initial condition information needed to properly initialize the sea ice and landsurface components of the model? -Can improvement in model initialization impact the predictability of specific events?

Summary of project objectives
The objective is to study these questions using land surface and sea ice initial conditions and climatologies built with the corresponding CNRM-CM components run in forced mode, and studying specific test cases with initial conditions representing extremes of the climatologies.

Summary of problems encountered
(If you encountered any problems of a more technical nature, please describe them here.) We originally intended to work with a more recent version of NEMO-GELATO but due to delays had to switch back to NEMO 3.2. As a consequence, the forced NEMO-GELATO runs initially planned were not run, because we were unable to recompile NEMO 3.2 on the new Broadwell nodes with the changes we wished to implement in the code, and we resorted to using initial conditions for sea ice derived by Mercator-Ocean by nudging the NEMO-GELATO model towards the GLORYS2V4 reanalysis. This restricts the re-forecast period to 1993-2012 instead of 1979-2012.
The most recent version of the SURFEX interface enabling nudging of land surface towards reference data is not yet ported on the Cray. We therefore chose to work on land surface initial condition sensitivity using the ERA-Land reanalysis, and an offline SURFEX run.

Experience with the Special Project framework
(Please let us know about your experience with administrative aspects like the application procedure, progress reporting etc.) The application procedure is quite straightforward. For the progress reporting, I understand that reporting end of June is convenient for ECMWF but for the intermediate reports it would probably be more efficient to report at the beginning of the following year. Reporting on the use of storage was not very easy for us as we already have national accounts and it was tricky to disentangle storage for the special project from storage from other experiments done with our other account.
During the end of the Special Project, we faced unforeseen deviations from the work plan (out of the three people involved in the project, one left the CNRM, and another changed position). This is the main reason why resources for the last year of the project were not entirely used.

Summary of results
(This section should comprise up to 10 pages, reflecting the complexity and duration of the project, and can be replaced by a short summary plus an existing scientific report on the project.) The aim of this project was to further explore the impact of land-surface and sea-ice initial conditions in the CNRM-CM model on predictions at sub-seasonal and seasonal time scales. Indeed, results from the FP7-SPECS project have shown the impact of soil moisture initialization is significant in terms of correlation for summer temperature over areas of Europe including the Balkans already identified as a "hotspot" for soil-atmosphere coupling (Ardilouze et al. 2017a). The initialization of sea ice was found to have little impact in prediction skill over the mid-latitudes in a multi-model framework, with most positive impacts found in the prediction of sea ice itself (Guemas et al. 2016).
The first step of the work consisted in reconstructing sea ice and land surface initial conditions and climatologies with the CNRM-CM model run in forced mode. We then revisited specific case studies with initial conditions from extrema years of the reconstructions to test the influence of land surface and sea ice on results from a control re-forecast with CNRM-CM using the actual initial condition of the given year.
This report is organized as follows: we first present the model version used for the project and how the initial conditions were constructed. We then summarize the main results for the land surface impacts and for the sea ice impacts.

-Model, initial conditions and re-forecasts
The CNRM-CM model version used for this special project corresponds to Météo-France seasonal forecasting system 5, operational at the time of the beginning of the project. This version of the model is also used for the CNRM contribution to the S2S project. A schematic of the model components, resolution and coupling frequency is shown in Figure 1. Initial conditions for the land surface component (SURFEX) were prepared with an offline SURFEX run with the ISBA-DIF multilayer soil scheme forced with ERA-Interim data (except for precipitation data calibrated using GPCC). An illustration of results for the Balkans region is shown in Figure 2 for the 1 st of May root-layer soil wetness index (SWI). These conditions are used in the re-forecasts discussed in section 2.
Ocean initial conditions were provided by Mercator Ocean International and are derived from the GLORYS2V4 reanalysis upscaled to the NEMO ORCA1° grid using a nudged NEMO run. The sea ice initial conditions are directly computed from this nudged run which uses the GELATO sea ice model. No sea ice concentration assimilation is used, the sea ice formation and melt is piloted by the ocean conditions in GLORYS2V4 and ERA-Interim atmospheric forcings. The run covers the 1993-2012 period, we therefore used this period as our re-forecast period for this study. Figure 3 shows the sea ice area on November 1 st found over different regions with this run.
Re-forecasts with 30 members, extended to 60 members in the last year of the project for the winter runs, were initialized in May and November using initial perturbations to the ERA-Interim atmospheric conditions on the 1 st of May and 1 st of November, respectively. This is a different approach to the one used in the Météo-France seasonal forecasting system 5 which uses in-run stochastic perturbations of the atmospheric prognostic variables. These re-forecasts were the baseline for a set of sensitivity experiments described below.

Soil moisture experiments
For soil moisture the focus season was boreal summer (JJA for the seasonal time scale, and a case study in July for the sub-seasonal time scale).
A first set of paired experiments aimed to revisit the study by Ardilouze et al. (2017a) with the version of CNRM-CM used for the special project, and therefore compared the skill of the reforecasts initialized in May with soil moisture conditions from the offline SURFEX run (SINI) to those initialized using the climatology of 1993-2012 1 st of May conditions from the same run, therefore suppressing the inter-annual variability of the surface conditions in the initial state (SCLI).
Several sets of case study experiments comparing results to SINI were run for the 2012 summer. Indeed, as shown in Figure 2, the SWI over the Balkans area was exceptionally low in May 2012 (record low over the time period). We therefore investigated the impact of this low SWI in the initial condition by comparing the JJA re-forecast for 2012 in SINI and SCLI, but also re-running the 2012 season starting from "average" SWI conditions (SNEU, using 2009) and anomalously wet conditions (SWET, using 2006 from the reconstruction).

Ocean and sea ice conditions
The impact of sea ice conditions was investigated at the seasonal time scale, focusing on the boreal winter season (DJF initialized on 1 st of November).
The reference re-forecast (IINI) uses initial sea ice conditions from the nudged NEMO-GELATO run constrained by GLORYS2V4. The case study focused on the 2009/10 winter which was marked by a strong negative NAO at the seasonal time scale. We chose two years from the last decade of the reanalysis period to avoid too much impact from the sea ice concentration trend over the reforecast period and ran two seasonal re-forecasts with minimum and (local) maximum sea ice area conditions, IMIN using 1 st November 2011 sea ice conditions, and IMAX using 1 st November 2005 conditions.
For the summer season, we also investigated the role of the ocean conditions over the North Atlantic region in the 2015 summer season during a master's thesis. Unfortunately, these simulations had to be run on the Météo-France supercomputer for technical reasons (code compilation issues) but since results are relevant for this project they are also summarized in this report.  However, in the specific case of 2012, the land surface initialization does not improve (and even slightly degrades) the near-surface maximum temperature anomaly found over the Balkans area, in contradiction with previous results (Figure 6).
In a more recent version of the model, a correction of precipitation fed into the land-surface component is shown to partially alleviate biases over key regions for atmosphere-soil coupling and improve skill (Ardilouze et al. 2019). However, this wasn't investigated with the version of CNRM-CM ran for the special project, and the lack of improvement found for a specific case study is in contradiction with average skill over the region, making it difficult to draw a general conclusion in the case of the Balkans.

Sub-seasonal time scale
At the sub-seasonal time scale, soil moisture seems to play a key role in the correct S2S forecasts for the 2015 heat-wave (Ardilouze et al. 2017b). Indeed, when the dry soil moisture conditions over France are replaced by anomalously wet conditions from the following year, the ensemble forecasts with CNRM-CM yield distributions of Tmax averages over France during the peak of the heatwave (July 1-6 2015, noted 6DS) that are shifted towards cooler values, even for the last initialization of the 1 st of July (see Figure 7). 1 st of June and 4 th of June initializations are no longer shifted towards warmer values, and the tails of the distributions no longer encompass anomalies as extreme as +10 K. Note that since surface conditions have been modified around the globe, and not only over the region of interest, we cannot exclude that some remote effects are also at play (see e.g. van den Hurk et al. (2012)). These results suggest that over Western Europe, soil moisture did play a role in the amplification of the heat wave and a correct initialization is one key ingredient to a successful forecast.

-Main results on sea ice and ocean conditions Sea ice sensitivity
Interpretation of results for the sea ice sensitivity experiments is quite tricky because of the choice of actual past years for sea ice conditions instead of an idealized framework where sea ice concentration can be entirely suppressed over specific regions. For instance, although 1 st November 2011 does correspond to a minimum in total Arctic sea ice area in the GLORYS sea ice conditions used over the 1993-2011 period, it is not the case for the Atlantic sector seas (see Figure 3).
We focused on the impacts of these sea ice conditions on a specific winter, DJF 2009/10. Results suggest that sea ice only had a very limited influence on the NAO re-forecast with our model, as shown by distributions of the NAO index in the IINI, IMAX and IMIN re-forecasts for 2009 shown in Figure 8 (a). The tail of the IMAX distribution is slightly shifted towards negative NAO values and on the other end, IMIN has slightly higher positive NAO values but these differences are very small and only related to one or two members of a highly dispersed ensemble.
Similar results are found when computing the Tibaldi and Molteni (1993) blocking index for the different experiments. The model clearly underestimates blocking frequency over the Atlantic sector with respect to ERA-Interim, and completely misses the very strong blocking frequency anomaly over the Atlantic in the 2009/10 winter seen with ERA-Interim data. The three sea ice initial conditions seem to play a very minor role in the winter circulation in the re-forecasts, since little to no clear difference is found between these simulations, suggesting that possible impacts in sea ice initial conditions are very quickly drowned in noise over the mid-latitudes in this version of the model.

Impact of ocean conditions on summer 2015
We present here additional results based on experiments run on the Météo-France supercomputer, but related to the topic of this special project.
The aim of these experiments were to investigate with the CNRM-CM model used for System 5 the hypothesis relating the "cold blob" anomaly in the North Atlantic ocean to the abnormally warm summer 2015 (see e.g. Duchez et al. 2016). To this end, sets of ensemble forecasts starting from May 1 st 2015 were run up to end of August, with surface relaxation of the NEMO model towards the GLORYS ocean reanalysis, either only the cold anomaly region, or the entire globe.
In Figure 9, we show results in terms of box plot distributions of the ensemble members nearsurface temperature anomalies over the Central Europe region defined by Duchez et al. (2016). Our reference forecast (without relaxation) failed to correctly capture the warm anomaly over this region, and despite nudging the ocean over the cold anomaly region (RANO) to correctly represent SSTs over this area, the temperature distribution is not significantly shifted. Only by implementing a global relaxation (RGLO) does our model represent an anomaly closer to the reference data.
We conclude from these results that if the cold anomalies over the Atlantic did play a role, the mechanism is not correctly represented in the coupled forecast system, and other phenomena may likely have played a key role in the occurrence of a warm summer.