Ultra-Stable Oscillator Stabilization using an Artificial Neural Network

We present results from experiments where we attempted to use an array of thermal sensors and an artificial neural network to predict the temperature-induced frequency fluctuations in a low-noise quartz oscillator. We trained the neural network by giving it thermal sensor readings as well as frequency measurements made against a hydrogen maser reference. The temperature vs. frequency model created by the neural network was then tested on separate data where the difference between the predicted frequency and measured frequency was collected. We found that the Allan deviation of this residual frequency difference was up to 10 dB lower than the Allan deviation of the original oscillator frequency. This result demonstrates the potential of using thermal sensors coupled with an artificial neural network to suppress temperature-induced frequency fluctuations in an oscillator without the need for bulky thermal insulation.