Giant telescopes: neural networks to predict turbulence and wind above the VLT and the LBT

The team responsible for operating the forecasting systems at the LBT and VLT observatories is testing neural networks against the autoregressive method, which is currently the benchmark for one-hour forecasts.

For large ground-based telescopes, and even more so for the European ELT under construction, knowing in advance how the optical turbulence of the atmosphere will behave is crucial: it conditions the planning of observations, the use of adaptive optics and even optical communications in free space. Today, the autoregressive (AR) method offers the best one-hour forecast, with an error of about 0.1 arcseconds on the "seeing", the measurement of image quality.

The team of E. Masciadri, who leads the ALTA prediction centre for the Large Binocular Telescope and the FATE project for ESO's Very Large Telescope, evaluated a series of neural network architectures, which are considered a priori to be better suited than random forests to this type of time series. The interest of deep learning lies in its flexibility and its ability to exploit correlations between several parameters processed simultaneously, or to take over where the AR method is not applicable.

The tests, conducted here on wind speed under operational conditions, also address a practical problem: optical turbulence is not measured during the day, which leaves gaps in the training data. The authors show that having measurements covering the 24 hours could improve the forecast.