Prediction of vertical optical turbulence profile from assimilated meteorological characteristics and optical measurement data using Machine Learning

Introduction

Optical inhomogeneities of the atmosphere limit the resolving power of ground-based telescopes. To improve the quality of observations, knowledge of the altitude distribution of the structural constant of the turbulent fluctuations of the air refractive index Cn2 is necessary. In this paper, we have tested machine learning methods[1] to predict the altitude profile of Cn2.

Methodology

The training dataset was prepared using ERA-5 atmospheric reanalysis data and meteorological characteristics obtained in the surface layer. These data were combined with archival optical observations of Cn2 at the Caucasus Astronomical Observatory of the Moscow State University GAISH.

The ERA-5 data have hourly resolution and are obtained by assimilation of atmospheric characterization measurements. The input parameters for model training were: boundary layer height, turbulent near-surface stresses, gravity wave stresses, and average wind gust speed[2].

Meteorological characteristics included air temperature, surface wind speed and direction, and humidity. Optical measurements were made with a telescope using the MASS-DIMM technique[3].

Data and results

Measured Cn2 values were available for altitudes: 0, 0.5, 0.71, 1, 1.41, 2, 2.82, 4, 5.66, 8, 11.3, 16, 22.6 km. The data archives cover the period from November 15, 2007 to June 16, 2013. Decimal logarithms of lgCn2 were used for training.

The study resulted in an optimal configuration of the Random Forest model linking the vertical profiles of atmospheric optical turbulence to the ERA-5 reanalysis data. The Random Forest method showed the best results: the correlation coefficient between measured and predicted Cn2 values was about 0.78.

For an altitude of 0.71 km, the mean absolute and squared deviations between the measured and predicted lg(Cn2) values were about 0.44 and 0.32, respectively. For an altitude of 22.6 km, these values were 0.13 and 0.03.

Conclusions

The most important input characteristics for modeling the vertical profiles of AOT are the ERA-5 reanalysis data, including wind speed components and air temperature. This suggests that AOT vertical profiles can be determined using machine learning for other astronomical observatories.

We estimate that the use of local surface data improves the quality of reproduction of Cn2 variations by 5-8%.

This work was supported by the Russian Science Foundation grant No. 23-72-00041. Measurements and primary data processing were supported by the Russian Ministry of Education and Science.

List of sources

1. Shikhovtsev A.Yu. et al. Estimation of astronomical visibility using neural networks at the Maidanak Observatory. Atmosphere 2024, 15, 38. https://doi.org/10.3390/atmos15010038

2. Shikhovtsev A.Y., Kovadlo P.G. Atmospheric boundary layer and free atmosphere: Dynamics, physical processes and measurement methods. Atmosfera 2023, 14, 328. https://doi.org/10.3390/atmos14020328

3. Shikhovtsev A.Y. et al. Vertical distribution of optical turbulence at the Terskol Peak Observatory and Mount Kurapdag. Remote Sensing. 2024, 16, 2102. https://doi.org/10.3390/rs16122102