Hybrid Optimization of Electromagnetic Devices Enhanced with Physics-Informed Machine
Robust optimization of microwave and RF devices has been very challenging because of the complicated electromagnetic behavior of the devices. A wide variety of optimization techniques have been proposed in the past, which can be classified into two main categories: search-based and gradient-based methods. While the search-based methods offer the potential to find global optima in the presence of a large number of design variables, they can be prohibitively expensive and time-consuming. Gradient-based methods, on the other hand, can find optimal solutions quickly, but they require initial designs to be in the vicinity of the solutions. All these optimization techniques are based on efficient forward solutions, which is particularly true for the search-based 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 are difficult to obtain. To overcome these challenges, we propose a hybrid algorithm enhanced with a machine learning approach. The hybrid algorithm combines the slow-but-global search-based method with the fast-but-local gradient descent method. In this algorithm, the search-based method is used as pre-screening to filter out bad initial designs and generate diversified better designs, with a rather loose ending criterion. Once these better designs are found, the gradient descent procedure is invoked to refine the designs to reach their optima quickly. This process is repeated until a desired design is obtained. To facilitate the proposed optimization scheme and improve its efficiency, we use a neural network-based surrogate model in both search and gradient descent processes. The neural network model can speed up the forward evaluation and provide analytical gradients using standard back-propagation in a very fast and accurate manner. To alleviate the computational burden associated with the time-consuming generation of training data, we further propose the use of a physics-informed machine learning model to improve modeling efficiency and reduce the training data generation cost.
