Neural Network-Based Genetic Algorithm for Complex Circuit Design of High-Power Vacuum Electron Device

Abstract

The circuits of high-power vacuum electron devices (HPVEDs) typically possess complex topologies that are crucial for efficiently converting electron beam energy to microwave energy. Due to the highly nonlinear beam-wave interactions, designing HPVED circuits generally relies on extensive particle-in-cell (PIC) simulations, making it a computationally intensive task. Especially for circuits with frequency tuning capabilities, the simulation workload is even one to two orders of magnitude higher than that of conventional circuits. To reduce the reliance on PIC simulations, this paper investigates the capability of artificial neural networks (ANNs) for modeling HPVED circuits. Given that the advantageous gene patterns are retained and recombined during the iterations of genetic algorithm, a method for HPVED circuit modeling using process data from the genetic algorithm is designed. This method avoids generating an extensive dataset for ANN pre-training before optimization. Testing on a dataset obtained by a simple genetic algorithm (SGA) shows that the ANN has good modeling capabilities for power, model evaluation, and tuning performance. Accordingly, this paper proposes a neural network-based genetic algorithm (NNGA), which significantly reduces the dependency on PIC simulations during optimization and enhances the efficiency of HPVED circuit optimization design. Preliminary tests on optimization tasks for HPVED circuits with one and two tuning parameters yielded excellent results, achieving tuning bandwidths of over 17% and 20%, respectively. In the tests, NNGA achieved optimization results comparable to SGA with half the simulation workload and better optimization results with the same simulation workload.