UC Irvine

Antenna arrays enhance communication, sensing, and imaging in fields like wireless communications, radar, and radio astronomy by improving signal quality and range through beamforming and spatial multiplexing. However, they face challenges such as mutual coupling, calibration issues, and increased complexity compared to single antennas. These issues require careful engineering to optimize array performance across applications. We present an optimization technique based on deep learning that improves the design of sparse phased arrays by minimizing grating lobes. This method starts with generating configurations for sparse antenna arrays, effectively tackling the non-convex challenges and the high degrees of freedom involved in array design. We utilize neural networks, trained on 70,000 and tested on 30,000 configurations, to estimate a non-convex cost function that evaluates the energy ratio between the main lobe and side lobe levels. This estimation is differentiable, facilitating the minimization of the cost function using gradient descent relative to the coordinates of the antenna elements, resulting in an optimized configuration. Additionally, we incorporate a custom penalty mechanism that integrates various physical and design constraints into the optimization process, enhancing its robustness and effectiveness in practical applications. The efficiency of our approach is resoundingly validated on the ten configurations with the lowest costs, achieving significant cost improvements ranging from 4.36 fold to 6.75 fold improvement, with a remarkable average improvement of 6.06 fold. The code for the method is available at https://github.com/david5010/Optimization-of-Antenna-Arrays.