Broad-Range Inverse Design of Antennas Using Deep Neural Networks with Material Parameter Embedding and Rapid Fine-Tuning

Abstract

Contemporary antenna design increasingly depends on numerical optimization to manage multiple design variables, objectives, and constraints. In many applications, such as metasurfaces, antenna arrays, or redesign over wide operating ranges, global optimization is no longer optional but essential. However, conventional bioinspired algorithms are notoriously inefficient, while machine learning-based methods struggle to construct sufficiently accurate metamodels. This communication presents a new broad-range optimization strategy based on inverse data-driven surrogates implemented with deep convolutional neural networks (CNNs). A custom embedding process is applied to allow constructing models covering diverse material properties (e.g., substrate permittivity). The surrogate maps antenna frequency responses to geometry parameters. Training data are generated in a dimensionally reduced space that captures the most influential directions in the design space, enabling accurate model construction with a limited number of samples. In the first phase, the inverse model produces an initial design directly from the performance specifications (e.g., desired operating bandwidth) mapped into a target frequency response, which is subsequently refined using an accelerated gradient-based optimizer. The presented approach is verified with two planar antennas optimized for the best reflection response and gain. The outcomes indicate that the inverse surrogate can be applied across a broad spectrum of frequencies, geometries, and substrate materials at a low cost of fewer than 30 electromagnetic (EM) analyses on average. Benchmarking demonstrates the benefits of our method and its practical edge over a broad spectrum of bioinspired, machine learning (ML), and surrogate-assisted population-based approaches.