Difference between revisions of "Detector Control"
| (3 intermediate revisions by the same user not shown) | |||
| Line 1: | Line 1: | ||
= Earthquake early warning for GW detectors and earthquake controls modifications = | = Earthquake early warning for GW detectors and earthquake controls modifications = | ||
| − | Based on real-time earthquake parameter estimation ([[https://usgs.github.io/pdl/ USGS PDL client]]), we realized an earthquake early warning for GW detectors, which is used today at LIGO and Virgo detectors. The purpose is not only to alert the operators and commissioners of an incoming earthquake, but also to predict the probability that an earthquake will interrupt the operation of the detector. If such a "lock loss" is predicted, the operators can decide to modify the controls into a more robust (and slightly noisier) state with the goal to ride out the earthquake and remain operative. The gain is to increase the duty cycle of a detector. | + | Gravitational-wave detectors rely on careful control of their mechanics including the suspended test masses to make sure that optics are properly aligned and distances between suspended optics do not change by more than a small fraction of the laser wavelength (1064nm) due to low-frequency forces like tides or ground motion. This way, the detector can be continuously operated and in a state of high light power inside the arm cavities. |
| + | |||
| + | However, occasionally, ground motion is strong enough to cause a lock loss, which means that the external forces acting on the alignment and positioning of optics cannot be compensated by the control system anymore. This typically means that optics start to move with increased amplitudes and uncontrolled, and the light power inside the arms drops to zero. Such events limit the duty cycle of the detectors, and it takes time to get the interferometer back online for a restart of observation mode. Most of these strong-ground motion events are associated with high-magnitude earthquakes with epicenters, for example, in the Pacific Ocean, Alaska, or South America. | ||
| + | |||
| + | Efforts to make the detectors more robust to earthquakes have led to the development of earthquake early-warning systems. Based on real-time earthquake parameter estimation ([[https://usgs.github.io/pdl/ USGS PDL client]]), we realized an earthquake early warning for GW detectors, which is used today at LIGO and Virgo detectors. The purpose is not only to alert the operators and commissioners of an incoming earthquake, but also to predict the probability that an earthquake will interrupt the operation of the detector. If such a "lock loss" is predicted, the operators can decide to modify the controls into a more robust (and slightly noisier) state with the goal to ride out the earthquake and remain operative. The gain is to increase the duty cycle of a detector. | ||
Details of the controls modification at LIGO can be found here: | Details of the controls modification at LIGO can be found here: | ||
| − | [[https://iopscience.iop.org/article/10.1088/1361-6382/aaa4aa | + | [[https://iopscience.iop.org/article/10.1088/1361-6382/aaa4aa Biscans et al (2018)]] |
| + | |||
| + | Improved prediction of lock loss was achieved with machine-learning techniques: | ||
| + | [[https://iopscience.iop.org/article/10.1088/1361-6382/ab28c1 Mukund et al (2019)]] | ||
| − | + | = Nonlinear feedback control = | |
| + | The goal of nonlinear feedback control is to improve robustness of a detector to external disturbances and to reduce controls noise relative to the current, linear controls methods applied at the detectors. Improvements can be expected in the presence of nonstationary data, or when the system has important nonlinearities in its dynamics. Our approach is to develop machine-learning techniques for feedback control, with current focus on reinforcement learning. | ||
Latest revision as of 19:58, 18 April 2020
Earthquake early warning for GW detectors and earthquake controls modifications
Gravitational-wave detectors rely on careful control of their mechanics including the suspended test masses to make sure that optics are properly aligned and distances between suspended optics do not change by more than a small fraction of the laser wavelength (1064nm) due to low-frequency forces like tides or ground motion. This way, the detector can be continuously operated and in a state of high light power inside the arm cavities.
However, occasionally, ground motion is strong enough to cause a lock loss, which means that the external forces acting on the alignment and positioning of optics cannot be compensated by the control system anymore. This typically means that optics start to move with increased amplitudes and uncontrolled, and the light power inside the arms drops to zero. Such events limit the duty cycle of the detectors, and it takes time to get the interferometer back online for a restart of observation mode. Most of these strong-ground motion events are associated with high-magnitude earthquakes with epicenters, for example, in the Pacific Ocean, Alaska, or South America.
Efforts to make the detectors more robust to earthquakes have led to the development of earthquake early-warning systems. Based on real-time earthquake parameter estimation ([USGS PDL client]), we realized an earthquake early warning for GW detectors, which is used today at LIGO and Virgo detectors. The purpose is not only to alert the operators and commissioners of an incoming earthquake, but also to predict the probability that an earthquake will interrupt the operation of the detector. If such a "lock loss" is predicted, the operators can decide to modify the controls into a more robust (and slightly noisier) state with the goal to ride out the earthquake and remain operative. The gain is to increase the duty cycle of a detector.
Details of the controls modification at LIGO can be found here: [Biscans et al (2018)]
Improved prediction of lock loss was achieved with machine-learning techniques: [Mukund et al (2019)]
Nonlinear feedback control
The goal of nonlinear feedback control is to improve robustness of a detector to external disturbances and to reduce controls noise relative to the current, linear controls methods applied at the detectors. Improvements can be expected in the presence of nonstationary data, or when the system has important nonlinearities in its dynamics. Our approach is to develop machine-learning techniques for feedback control, with current focus on reinforcement learning.
