Coalition to Make LIGO Sentient
Our goal is to collect a bunch of concrete ideas for what (adaptive control / advanced control techniques / machine learning techniques) we can/should do in the near term as well as what we should start to think about for the longer term. The near term output of this brainstorming will be a white paper that we write on the topic in collaboration with MIT.
Contents
Ideas for Future Things to Implement
Near Term (0 - 2 years)
- 40m: implement LMS based FF for seismic subtraction on multi DOF for LSC controls
- 40m: adaptive subtraction of OL controls to reduce angle to length coupling noise
- 40m: static + adaptive subtraction of aux. controls (PRC/SRC/MICH) from DARM
- 40m: acoustic noise subtraction from MC_F / IFO common mode
- 40m: optimal feedback design for simple loops based on cost function
- LIGO: copy/paste of eLIGO FF to multi length DOFs
- LIGO: adaptive subtraction of WFS controls to reduce angle to length coupling noise (ala Dooley)
- LIGO: static + adaptive subtraction of aux. controls (PRC/SRC/MICH) from DARM
- LIGO: optimal feedback design for simple loops based on cost function
Medium Term (2 - 5 years)
- PEM channel based feedback states (train mode, EQ mode)
- Training of LSC/ASC loop shapes to minimize glitch rate at low frequencies (I have no idea how this would be implemented)
- Optimal length of time to "rest" between locking sequences
- TCS heating - how much TCS for how much PSL power, and for how long
- SUS input matrix - constant update??
- Know when something (oplevs, bad servo, etc.) is kicking something up. Turn down gain / fix loop, or just turn off until it's fixed.
- Build a simulated IFO plant, check against the real plant. Probably we'd give it the general shape, but it would tweak filter shapes to match real IFO.
Long Term (5 - 10 years)
- Learning system uses free hanging IFO time series to design its own lock acquisition algorithms
- Online simulation uses actual PEM inputs to predict glitch rates, does MCMC to find optimal feedback/feedforward solutions and then tries them out on the real system to learn more
- Large array of seismic / acoustic sensors placed all over the vacuum system can be used to find sources of backscattered light / phase noise
- Machine has list of people / skills. Can send SMS to ask for help
- Pre-emptive prediction of failure of facility systems (e.g. power line fluctuations, weather related power failures, vibrations in HVAC indicating fan bearing failures,...)
- monotonic increase in environmental coupling indicates optics getting dirty, photodiodes getting damaged, etc.
- bad IFO alignment = operator getting sleepy
- Initial alignment of IFO, even before there is any flashing in cavities