# Deep Learning Wavefront Sensing from Object Scene for Directed Energy HEL Systems

## TL;DR

This paper introduces a deep learning method to sense wavefront distortions in HEL systems using scene imagery, eliminating the need for traditional wavefront sensors and reference beams.

## Contribution

A novel deep learning approach for wavefront sensing that uses scene imagery and generalizes across turbulence levels and UAV types.

## Key findings

- The DL model accurately predicts Zernike coefficients from aberrated imagery of the Reaper UAV across various turbulence levels.
- The model generalizes well to turbulence levels beyond its training range and to unseen UAV types like the Mongoose UAV.
- Training across multiple turbulence levels improves model accuracy and practical deployment potential compared to single-level training.

## Abstract

Atmospheric turbulence significantly degrades the performance of High Energy Laser (HEL) systems by distorting the laser wavefront as it propagates through the atmosphere. Conventional correction techniques rely on Adaptive Optics (AO), which preserve beam quality at the object. However, AO systems require wavefront sensors, such as Shack–Hartmann, and a reference beam, increasing system complexity and cost. This work presents a Deep Learning (DL)-based wavefront sensing approach that operates directly on scene imagery, thereby eliminating the need for dedicated wavefront sensors and a reference beam. A DL model was trained to predict wavefront distortions, represented by Zernike coefficients, from aberrated imagery of the Reaper Unmanned Aerial Vehicle (UAV). Reaper imagery utilized in training was aberrated at different levels of turbulence, D/r0, with D=30 cm being the aperture diameter of a telescope capturing the object scene and r0=3, 5, 7 cm the Fried parameter that defines weak turbulence for higher values and strong turbulence for lower values. The proposed model, trained across all these turbulence levels, outperformed models trained on a single level by providing superior accuracy and offering practical advantages for deployment. The model also demonstrated strong generalization capabilities for two practical scenarios: (a) Reaper imagery with turbulence levels beyond the training range, and (b) Mongoose UAV imagery not included in the training set. The model predicts turbulence accurately in both cases. The results confirm that if the model is trained for a UAV model for a certain turbulence level, it provides accurate predictions for turbulence levels outside its training range and for other UAV aberrated images.