Optical turbulence at the VLT and the LBT: reference models to finally assess the accuracy of AI predictions

To find out whether a machine learning model is better at predicting turbulence above large telescopes, a point of comparison is needed: one team proposes two such points.

The quality of images obtained by large telescopes equipped with adaptive optics depends heavily on atmospheric turbulence. Forecasting makes it possible to better plan the nights of observation, and many models, both digital weather forecast and machine learning, try to do so. Still, we must be able to measure what they actually bring.

A team including Elena Masciadri has therefore defined two reference models: one predicts the average value of an astroclimatic parameter over a given duration, the other its evolution over time over the same duration. They were applied to several years of measurements collected at the sites of the Very Large Telescope in Chile and the Large Binocular Telescope in Arizona, and the authors are examining how to extrapolate these reference values to other observatories according to their turbulence statistics.

Conclusion: Basic accuracy varies according to site, parameter, time scale and approach. The performance of a forecasting model cannot therefore be expressed in absolute terms, but only in relation to these references.