AI-assisted Sensing & Control for Gravitational Wave Observatories

Gravitational wave detectors like Advanced LIGO have detected hundreds of astrophysical black hole merger events. Pushing the sensitivity of these instruments to capture fainter signals from the earlier universe crucially depends on how well we can improve the sensing and control aspects of these interferometric devices. GW observatories often strive to find the proper balance between maintaining a stable operating point, a critical factor in ensuring optimal sensitivity for extended periods, and continuously upgrading various subsystems that will help expand their astrophysical reach. There is a trade-off between observation time and time allocated for upgrades. The inherent complexities associated with the cross-coupled nature of these opto-mechanical experiments and the non-stationary nature of the ambient environment often make these tasks daunting. In this talk, I will discuss the potential of using deep neural systems for improved sensing and control at these observatories. Leveraging physics-informed deep reinforcement learning and model-based design can help improve learning efficiency while adhering to safety constraints, which is critical for real-world systems. I will discuss the potential of embedded machine learning for low-latency decision-making and highlight recent successes in deploying deep reinforcement learning-based systems in large-scale physics experiments. Finally, I will mention the challenges encountered while deploying neural systems in complex optical systems and why tackling these issues and building intelligent systems with adaptive capabilities will be crucial for next-generation GW detectors like Cosmic Explorer and the Einstein Telescope.