Artificial neural networks (ANNs) have shown remarkable advantages in antenna modeling and optimization as surrogate models. However, in single‐objective modeling, the convergence accuracy of ANN‐based antenna models is often insufficient. This paper proposes an improved crayfish optimization algorithm (ICOA) to optimize the initial weights and biases of the neural network, aiming to enhance both convergence accuracy and speed. The improvement incorporates an elite opposition strategy and iterative local search while simplifying certain formulas in the original algorithm. The improved algorithm was evaluated on standard test functions, demonstrating superior convergence speed and accuracy compared to other algorithms. Subsequently, the ICOA was applied to the single‐objective modeling of two microstrip patch antennas, predicting antenna return loss S ₁₁ . The findings indicate that the proposed ICOA‐ANN approach demonstrates superior modeling accuracy compared to both the COA‐ANN and traditional ANN methods.

As essential radiating elements in RF and microwave microsystems, microstrip antennas require sufficient bandwidth to ensure stable operation, integration flexibility, and overall microsystem performance. From a microsystem optimization perspective, this paper proposes a bandwidth extension method for microstrip antennas that combines swarm intelligence and reinforcement learning. The proposed ICOA-TD3 framework is designed to enhance antenna bandwidth within target frequency bands and thus improve the performance robustness of compact RF microsystems. In the proposed method, an improved crayfish optimization algorithm (ICOA) is first used to explore the global design space and achieve global bandwidth enhancement, followed by the Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithm for local refinement and further exploitation of the antenna structure’s bandwidth potential. In Experiment 1, the impedance bandwidth ($S_{11}\le -10\mathrm{dB}$) is increased by up to 200%. In Experiment 2, the impedance bandwidth ($S_{11}\le -10\mathrm{dB}$) and axial-ratio (AR) bandwidth ($\mathrm{AR}\le 3\mathrm{dB}$) are improved by up to 27% and 250%, respectively. The results indicate that the proposed method is a feasible solution for bandwidth-oriented optimization of microstrip antennas and is promising for the intelligent design of high-performance RF microsystems.