7 Conclusion and perspectives

We demonstrated that our simulations give realistic results. A non-hydrostatic model, well adapted to put in evidence the orographic effects on the development of the gravity waves, reproduces a seeing characterized by an extended turbulent region over the Chilean coast and by general low turbulence levels over the Paranal mountain site. This numerical technique could be thus efficiently used for site testing studies.

Analyzing the whole PARSCA93 campaign we conclude that the adaptation time to the orography is never smaller than 30 min. In most of the cases the model fluctuates around a stationary mean value and it does not diverge. Moreover, the time window [$2^{\rm h}-3^{\rm h}$] seems to give the best results following two different criteria: the correlation between the Scidar measurements and the Meso-Nh simulations is maximum and the standard deviation to the regression straight line is minimum.

We demonstrated that the model has a good sensitivity and can discriminate the worst from the best seeing of the campaign i.e. from 0.38 to 1.38 arcsec.

The turbulence rate simulated by the model is well estimated both in the low and high part of the atmosphere. We remember briefly the turbulent layer developed in 16 May night after 4 hours simulations time (Fig. 8 in Masciadri et al. 1998) and the layer resolved by the model at 10 km in 25 Mars 1992 nigh (Fig. 17).

For the first time we show that is necessary to retrieve $C_{\rm N}^2 \$ profiles and thence the seeing. Using only integrated values without knowledge of the vertical turbulence distribution, one could expect a wrong estimation of the simulation time.

This study can provide complementary information for the feasibility of the seeing forecasting with ECMWF products (Benzi et al. 1996). In this paper it was proved that the ECMWF products give a good ground temperature prediction. We have shown that, to use a numerical technique, it is not sufficient to forecast the ground temperature well. It is necessary to assure the quality of the observed data in the whole atmosphere.

We proved that the poor quality of the Antofagasta radiosoundings is the main cause of unstable simulations obtained during a few nights. In these cases, unrealistic dynamic instabilities are produced by the model in the middle and high part of the atmosphere. The Antofagasta radiosoundings have a sufficient vertical resolution but are not completely reliable, at least in the limited temporal period analyzed (only 8 nights). We compared our numerical technique with the forecasting-by-persistence one but no conclusive statements can be provided for the following reasons:

• The sample analyzed is statistically poor.
• A comparison with the forecast by persistence method might be significant only on a longer period.

Antofagasta radiosoundings seem not to be of extremely good quality but it is true that:

• The Antofagasta radiosoundings are not the only products supplied by the Meteorologic Centers and others kinds of data could be analyzed. We know that analysis sampled on 31 pressure levels instead of 15 levels are provided by the ECMWF.
• During this study radiosoundings and ECMWF analysis of May 1993 have been used. Today the operational meteorological centers dispose of more and more heterogenous data in type and quality and they are not necessarily provided at synoptic hours (00:00 U.T., 06:00 U.T., 12:00 U.T. and 18:00 U.T.). The developement of new observational systems such as geostationary satellites provides a lot of informations with a high spatial and temporal resolution that can add to the data in the regions in which the observation density is poor. New methods of data assimilation are necessary to treat this new kind of data. In the last years consistent progress was made in developing new and more powerful assimilation schemes (3D and 4D variational schemes) that can integrate all the heterogeneous measurements which we have at our disposal today (Bouttier & Rabier 1998). We know that the ECMWF forecasting quality has become better and better in these last years. It could be interesting to test the model performance on more recent measurements to estimate if this progress in the forecasting science is valuable for our simulations.