Investigating the Role of Machine-Learning Techniques in Performance Prediction of High Power Microwave Devices

Particle-in-cell code-based finite-difference time domain (PIC-FDTD) simulations are widely used to design, simulate, and analyze high power microwave (HPM) devices. Although these simulations are quite accurate in simulating physics of the HPM devices and can generate huge amounts of data, they suffer from several limitations such as long runtimes and lack of optimization features. The aim of this paper is to apply machine learning techniques on the data generated by PIC-FDTD simulations in order to predict or forecast several performance parameters of HPM devices such as number of simulation particles, output power, current, efficiency, etc. The objective is to develop a forecasting algorithm which can be univariate or multivariate based upon the time series data generated by PIC simulations with intervals of several pico-seconds. Scope and performance of some of the popular machine learning algorithms such as ARIMA, VARIMA, Auto ARIMA, Fourier regression model, NBEATS model, Naive baselines, TCN model, Temporal fusion transformer models, etc., shall be compared. This work could be a starting point toward determination of optimum device design and performance parameters.