Overview

• Google DeepMind, working with LIGO and GSSI, introduced the Deep Loop Shaping controller and reported results in Science after on-instrument tests at the Livingston observatory.
• The method cut noise in LIGO’s most unstable feedback loop by 30–100 times, pushing control-induced vibrations below those from quantum radiation pressure.
• Using frequency-domain rewards, the reinforcement-learning approach outperformed traditional linear controllers in simulation and matched those gains on hardware.
• Full deployment across the many mirror control loops has not happened yet, and an outside expert notes the real-world run to date was brief, with long-duration stability still to be demonstrated.
• Researchers project that scaling the technique could enable hundreds more gravitational-wave detections per year and see broader use in vibration-sensitive fields such as aerospace, robotics, and structural engineering.