Study of the Applications of Machine Learning in Aligning Millimeter Wave Quasi-Optic Higher Order Mode Generators

Abstract

Gyrotrons are active devices that enable the generation of higher-power (greater than hundreds of kW) millimeter waves by employing cyclic electron beams to interact with rotating high-order modes. To simulate gyrotron performance at low-power signals (in mW), mode generators are used to produce higher-order modes mimicking gyrotron modes. This simulation aids in understanding gyrotron performance, optimization, and other necessary experimentation that does not require high-power signals. The design of a mode generator typically involves the use of of the coaxial cavity which can be excited by focusing a shaped beam on its caustic radius using a quasi-optic (Q-O) source. Focusing the beam requires the proper alignment of the Q-O source, which is very challenging and time-consuming.

This thesis introduces various machine learning (ML) techniques for optimizing the alignment of a quasi-optical higher-order mode RF generator in the millimeter band. It is the first time ML is used in the alignment of quasi-optic higher order mode generators. The ML techniques used in this thesis involve mode pattern analysis using machine-learned simulated data, which helps in calculating misalignment positions. The training data are obtained from simulating mode generator output mode patterns at several misalignment positions in a 2D plane. Mean Square Error (MSE), Scalar Correlation Coefficient (ηs), and Vector Correlation Coefficient (ηv) are used as feature extraction methods to quantify the obtained mode pattern and reduce data dimensionality. These feature extraction methods are used as inputs of the machine learning model.

To train various ML models, regression algorithms that can predict alignment positions are used. In this work, Linear Regression (LR) and Deep Neural Networks (DNN) are discussed. These algorithms help to create a model that makes inferences on misalignment positions using the extracted features (MSE, ηs and ηv) as inputs. The prediction is used to control the motion controller which moves the mode generator to a proper alignment position.

The aligned mode generator output mode pattern using this technique shows improved mode purity of 96.8% in the W-band (95 GHz T E6,2). The results demonstrate excellent repeatability and adaptation to misalignments outside the training range. The alignment time using these techniques is significantly reduced to less than 180 minutes on average, which is far less than the traditional manual alignment methods that take more than 4000 minutes.