The construction of a noise eater using a neural network to stabilize laser power output

This thesis project is the first in a total of two thesis projects. The main focus of this project was the design and implementation of a noise eater circuit which utilizes a neural network (NN) for its control element. The NN replaced a proportional-integral-derivative (PID) controller in a laser-stabilization feedback loop, inspired by Cheon et al. [1]. In this paper, the training and implementation of the neural network is discussed. Initially, a simulated feedback loop was built in MATLAB’s Simulink environment to test the NN’s ability to replace the PID controller. The NN was able to successfully stabilize the simulated laser, though not as well as the PID controller. Then the NN replaced the PID controller in the real set-up. The NN was able to successfully stabilize a laser for a short time (<1 >s) before error build-up caused the stability to fail. The second part of this process, which will form the next thesis, is replacing the neural network with a reinforced neural network, which will be giving the goal of maintaining stable laser power rather than simply mimicking a PID controller. It is believed that a reinforced neural network will not suffer from the same error-build-up problems as a traditional neural network, allowing it to achieve long-term laser stability.