Behavioral modeling of microwave components in constrained domains using recurrent neural networks with attention layers

Abstract

Full-wave electromagnetic (EM) analysis is commonly used in microwave engineering. Nonetheless, EM-driven design involves considerable expenses. Surrogate modeling may mitigate this issue; however, constructing reliable metamodels is demanding due to the curse of dimensionality and design utility requirements. This paper introduces a novel data-driven modeling methodology that integrates an attention mechanism into Recurrent Neural Networks (RNNs) based on Long Short-Term Memory (LSTM) and Gated Recurrent Units (GRU). The proposed surrogate enhances the representation of sequential dependencies across the frequency range in S-parameter estimation, treating frequency responses as time-series data. The projected LSTM and GRU layers reduce the dimensionality of the hidden states through a learned projection matrix, which reduces computational complexity and enhances the capacity of the model to capture long-term dependencies. Another enhancement is the attention mechanism, which dynamically assigns greater importance to critical frequency regions and enables the model to selectively highlight the most relevant components (e.g., corresponding to the circuit's center frequencies). The predictive power is further improved by sensitivity-based dimensionality reduction. Comprehensive verification showcases the competitive performance of our technique regarding predictive power as compared to multiple benchmark methods.