Dr. Nikhil Mukund

In gravitational-wave (GW) observatories, suspended optics are often susceptible to alignment perturbations, particularly to slow drifts over time due to environmental changes. Such misalignment affects the coupling of the incident laser beam into the optical cavities. This drift leads to degradation of circulating laser power, affects optical squeezing, and thus decreases the overall astrophysical sensitivity to merging compact binaries. Traditional alignment sensing is carried out using differential wavefront sensing techniques, but these are often bandwidth-restricted and are susceptible to sensing noise. In this talk, I present the first successful implementation of a neural sensor-based alternative architecture for drift control at a gravitational wave observatory and the corresponding improvement in sensitivity and duty cycle observed at the GEO600 detector. The talk will also cover the recent results of designing multi-input-multi-output (MIMO) controllers using deep reinforcement learning (RL). This experience-based learning has the potential to bring us a step closer to achieving artificial general intelligence. The convergence in performance similar to classical controllers demonstrates deep RL's capabilities as a viable tool for real-time control for current and next-generation GW detectors.