Deep Loop Shaping: AI-designed controllers improve LIGO sensitivity
September 18, 2025
Our collaboration with Google DeepMind has produced a new reinforcement-learning method called Deep Loop Shaping that designs better feedback controllers for LIGO’s mirror stabilization systems. The technique uses frequency-domain rewards to train controllers that reduce low-frequency control noise by over 30x across the 10–30 Hz band at LIGO Livingston — surpassing the quantum-noise-limited design goal.
This improvement opens a new observational window for intermediate-mass black hole mergers and gives earlier warning for neutron star collisions. Scaling the approach to all of LIGO’s mirror control loops could yield hundreds of additional detectable gravitational-wave events per year.
Published in Science 389, 1012–1015 (2025). Preprint: arXiv:2509.14016.