Feature engineering driven estimation of C_n² from deformed optical signals using neural networks

Atmospheric turbulence distorts the optical wavefront and deteriorates the bit rate in free-space optical communication or quantum key distribution. Adaptive optical methods are limited by the feedback timescales in the hardware. Here, we estimate the atmospheric refractive index structure constant Cn2 from distorted optical signals using a neural network that integrates physics-informed feature engineering. The advantage of feature engineering was reconfirmed by principal component analysis of the input samples. The model estimates Cn2 to an accuracy of up to 99%, within the coherence time relevant to the particular Cn2. This technique will provide ML-based augmentation in the reduction of link-error probability in free-space optical communication systems.