Abstract
Electromagnetic (EM) simulation is widespread in microwave engineering. EM tools ensure evaluation reliability but incur significant expenses. These can be mitigated by employing surrogate modeling methods, especially to expedite design workflows like local/global optimization or uncertainty quantification. However, building accurate surrogates is a daunting task beyond simple cases (low dimensionality, narrow geometry parameter and frequency ranges). This research suggests a new technique for dependable modeling of microwave circuits. Its main ingredient is a recurrent neural network (RNN) with the main architectural components being bidirectional Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) layers. These are incorporated to accurately represent frequency relationship within circuit characteristics as well as dependencies between its dimensions and outputs considered as vector-valued functions parameterized by frequency. The network's hyperparameters are adjusted through Bayesian Optimization (BO). Utilization of frequency as a sequential variable handled by RNN is a distinguishing feature of our approach, which leads to the enhancement of dependability and cost efficiency. Another critical factor is dimension- and volume-wise reduction of the model's domain achieved through global sensitivity analysis. It allows for additional and dramatic accuracy improvements without diminishing the surrogate's coverage regarding circuit's operating parameters. Our methodology has been extensively validated using several microstrip structures. The results demonstrate its competitive performance over a range of kernel-based regression techniques and diverse neural networks. The proposed procedure ensures building models of outstanding predictive power while using small training datasets, which is beyond the capabilities of benchmark algorithms.