LIGO Document T1800244-v2

Investigation of Optimal Non-linear Temperature Control

Document #:
Document type:
T - Technical notes
Other Versions:
This is an node for collecting links to documents for Shruti Jose Maliakal's 2018 SURF project on Temperature control using novel machine learning non-linear controls. Reports and relevant documents will be linked back to this document number for a central reference point.

The realisation of gravitational-wave detectors with higher sensitivity demands an improvement in their Brownian thermal noise floor. In an experiment to directly measure this Brownian noise of a novel mirror coating, readout is limited by variations in bulk temperature that cannot be controlled adequately using PID thermal feedback. The use of machine learning neural networks which actively learn to find the optimal set of control parameters as a function of the state of the system is a prospective upgrade from PID.

A physical model of the system must be developed to train potential algorithms. Initially, a simple but representative model including only the vacuum can was chosen and made into an OpenAI gym environment. A suitable reinforcement learning algorithm that can control the system in real time must be chosen. DeepQ, PIDNN, ACKTR, and Curiosity are algorithms being tested on the environment. On gauging its success, further complexity can be added to the environment to obtain a closer simulation of the real physical situation. Experiments were also performed on the system to obtain more accurate values of parameters and improve the model. The final aim is to train the network with real data and obtain the optimised heat actuation.

Files in Document:
Other Files:
Notes and Changes:
Adding final version of report submitted to SFP.

DCC Version 3.4.1, contact DCC Help