[Paper] Active Convolved Illumination with Deep Transfer Learning for Complex Beam Transmission through Atmospheric Turbulence

Source: arXiv - 2512.19540v1

Overview

Atmospheric turbulence scrambles light beams, degrading performance in imaging, remote sensing, and free‑space optical links. The paper introduces Active Convolved Illumination (ACI)—a physics‑driven beam‑shaping technique—and explores how it can be combined with deep transfer learning to further suppress turbulence‑induced distortions. By marrying a deterministic optical strategy with data‑driven models, the authors lay groundwork for hybrid systems that could make high‑fidelity optical communication more robust in real‑world conditions.

Key Contributions

• Conceptual integration of ACI and deep learning: Proposes a framework where a convolutional neural network (CNN) supplies learned priors to ACI’s correlation‑injection process.
• Transfer‑learning pipeline: Demonstrates how a pre‑trained CNN can be fine‑tuned on turbulence‑affected beam data, reducing the amount of domain‑specific training required.
• Feasibility study with simulated turbulence: Shows that the hybrid ACI‑DL approach can recover structured beam profiles more accurately than ACI alone in challenging turbulence regimes.
• Guidelines for coupling physics‑driven and data‑driven methods: Identifies conditions (e.g., similarity of statistical turbulence statistics, sufficient SNR) under which the learned model meaningfully augments ACI.

Methodology

1. Active Convolved Illumination (ACI) – ACI pre‑processes the transmitted beam by convolving it with a carefully designed kernel that injects spatial correlations tailored to counteract expected turbulence distortions.
2. CNN with Transfer Learning – A standard CNN architecture (e.g., ResNet‑18) is first trained on a large synthetic dataset of turbulence‑corrupted beams. The network learns to map distorted intensity patterns back to their clean counterparts.
3. Hybrid Loop – The CNN’s output (a refined estimate of the beam’s phase/amplitude) is fed back into the ACI kernel generation step, effectively “informing” the physics‑based illumination with data‑driven insights.
4. Simulation Environment – Turbulence is modeled using Kolmogorov phase screens with varying strength (Cₙ²) and inner/outer scales. Beam propagation is simulated via split‑step Fourier methods, allowing quantitative comparison of ACI‑only, CNN‑only, and hybrid performance.

Results & Findings

The hybrid approach consistently outperforms each component alone, especially when turbulence statistics differ from the training set—a key indicator that the physics‑driven ACI layer supplies useful regularization to the learned model.

Practical Implications

• Free‑Space Optical (FSO) Links: Deploying an ACI‑DL module on transmitter hardware could maintain high data rates over longer distances without expensive adaptive optics hardware.
• Lidar & Remote Sensing: Structured illumination patterns (e.g., vortex beams) can be recovered more reliably, improving target detection and classification in airborne or satellite platforms.
• Quantum Communications: Preserving orbital angular momentum (OAM) states through turbulence is critical for high‑dimensional quantum key distribution; the hybrid method offers a low‑latency, software‑centric solution.
• Edge‑Computing Friendly: Transfer learning reduces the training burden, enabling on‑device inference on GPUs or specialized AI accelerators already present in modern optical transceivers.

Limitations & Future Work

• Simulation‑Only Validation: Results are based on numerical phase‑screen models; real‑world atmospheric dynamics (e.g., anisotropic wind, temperature gradients) may introduce unmodeled effects.
• Latency Considerations: The feedback loop between CNN inference and ACI kernel update adds processing time; optimizing for real‑time operation remains an open challenge.
• Scalability to Multi‑Beam Systems: Extending the framework to multiplexed channels or multi‑aperture arrays requires additional architectural tweaks.

Future research directions include hardware‑in‑the‑loop experiments, exploring lightweight CNN architectures for ultra‑low latency, and integrating reinforcement learning to adapt ACI kernels on the fly as turbulence evolves.

Authors

• Adrian A. Moazzam
• Anindya Ghoshroy
• Breeanne Heusdens
• Durdu O. Guney
• Roohollah Askari

Paper Information

• arXiv ID: 2512.19540v1
• Categories: physics.optics, cs.LG
• Published: December 22, 2025
• PDF: Download PDF