Work Packages

WP1 — Pushing Neural-Network Performance for Next-Generation AO Control

Adaptive optics increasingly relies on advanced control strategies to deal with rapidly evolving turbulence, strong non-linearities, and the demanding requirements of astronomical and space-surveillance systems. Recent progress in reinforcement learning and world-model architectures offers major opportunities, but their performance, robustness, and predictive capability must be significantly improved to meet operational constraints. The goal of WP1 is to develop advanced AI-based controllers, integrating improved reward strategies, multi-sensor data fusion, RL directly from wavefront-sensor images, and predictive world models for turbulence and Point-Ahead Angle estimation. WP1 will deliver optimized RL architectures, predictive modules based on JEPA/SimVP/meta-RL, and a unified framework that significantly improves control stability, accuracy, and adaptability over state-of-the-art AO controllers.

WP2 — Co-Design and Wavefront-Sensor Optimization Under Strong Turbulence

In strong turbulence regimes, conventional wavefront sensors suffer from reduced sensitivity, scintillation, and pronounced non-linearities that severely degrade AO performance. Deep learning and optical co-design now make it possible to rethink the sensing architecture itself, moving beyond traditional Shack-Hartmann and pyramid concepts. WP2 aims to develop new non-linear WFS designs—such as curvature or diffusion-based sensors—and to explore focal-plane sensing strategies that can maintain performance under severe atmospheric conditions. This work combines physical modelling with AI-driven optimisation to jointly redesign the optical element and the reconstruction algorithm. WP2 will deliver novel WFS concepts, validated reconstruction models, and performance evaluations in simulated strong-turbulence environments with clear pathways toward on-bench implementation.

WP3 — Real-Time Implementation and Hardware Optimization for AO Systems

To be operationally useful, the innovations from WP1 and WP2 must be embedded into real-time AO systems capable of running at kilohertz frequencies with minimal latency. Current RTC architectures are not designed for deep neural networks or predictive models, and require substantial optimisation. WP3 aims to integrate the new sensing and control modules into real-time pipelines, focusing on GPU/FPGA acceleration, reduced-latency execution, and architectural restructuring for deterministic performance. The WP will deliver fully functional real-time modules, integrated into the DAO/RTC framework, along with benchmarked hardware implementations demonstrating feasibility for next-generation AO instruments.

WP4 — On-Sky Demonstration of AI-Enhanced Adaptive Optics

Laboratory testing is insufficient to validate AI-driven AO systems, which must prove their robustness under genuine atmospheric conditions and operational constraints. WP4 provides this critical validation step by deploying the new WFS and control solutions on real telescopes using AO4FEELINGS, PAPYRUS, and associated platforms. The goal is to assess performance improvements in terms of stability, sensitivity, predictive accuracy, and image quality, while demonstrating operational readiness for future ELT and PROVIDENCE-class instruments. WP4 will deliver on-sky performance reports, validated datasets, and a final demonstration showing the practical impact and maturity of the AI-enhanced AO solutions developed throughout the project.

Page updated

Google Sites

Report abuse