Physics-Informed Machine Learning for Hybrid Optimization of Microwave and RF Devices

Design optimization of microwave and RF devices has been of great interest to researchers and engineers in the microwave and RF communities for decades. To achieve an effective optimization, one often faces two challenges. The first is to choose a robust optimization technique. In the past, a wide variety of optimization techniques have been proposed, and they can be classified into two main categories: search-based algorithms and gradient-based methods. The genetic algorithm (GA) is one of the most popular search-based algorithms, based on the principle of natural evolution. While GA offers the potential to search for global optima in the presence of a large number of design variables, the method can be prohibitively expensive. Gradient-based optimization methods, on the other hand, are based on the analytical evaluation of gradients of the objective function with respect to design variables. Fast result can be achieved provided that the initial design is in the vicinity of the optimal solution. The second challenge is to obtain very efficient forward solutions required in the optimization process. This is particularly true in the GA optimization which requires the evaluation of many designs. In the gradient-based optimization, one not only needs to provide forward solutions, but also their gradients with respect to design parameters, which is by no means trivial.