Hybrid optical turbulence models using machine-learning and local measurements.

Area of Science:

• Atmospheric science
• Optical engineering
• Machine learning

Background:

• Accurate prediction of atmospheric optical turbulence is crucial for free-space optical (FSO) systems.
• Existing macro-meteorological models may not perform well in new environments.
• Developing new models with local data is time-consuming and resource-intensive.

Purpose of the Study:

• To develop a machine-learning informed hybrid model framework for predicting atmospheric optical turbulence.
• To improve the predictive power of baseline macro-meteorological models by integrating local observations.
• To evaluate the performance of hybrid models against baseline and data-only machine learning models.

Main Methods:

• A hybrid model framework was developed by combining a baseline macro-meteorological model with local observations.
• Gradient boosted decision tree architecture was used to train both hybrid and data-only models.
• Models were trained using varying amounts of in situ meteorological observations.

Main Results:

• Hybrid and data-only models outperformed three baseline macro-meteorological models, even with minimal local data (as little as one day).
• The hybrid model achieved a 29% reduction in mean absolute error with one day-equivalent of data, increasing to 68% with 180 days.
• Hybrid models showed slightly superior performance compared to data-only models, with the advantage diminishing as more data became available.

Conclusions:

• The hybrid model framework offers a viable solution for enhancing atmospheric optical turbulence prediction in localized environments.
• This approach is particularly beneficial when collecting local meteorological data is costly or difficult.
• The performance of the models suggests that incorporating local data significantly boosts predictive accuracy for FSO systems.