Optimization of Antenna Array Configurations Using Deep Learning

We introduce a deep learning-based optimization method that enhances the design of sparse phased array by reducing grating lobes. Our approach begins with a generation of sparse antenna array configurations, efficiently addressing the non-convex challenges and high degrees of freedom in array design. We then employ neural networks, trained on 70,000 and tested on 30,000 configurations, to approximate a non-convex cost function that measures the ratio between the energy of the main lobe and the side lobe level. The approximation is differentiable and allows for the minimization of the cost function by gradient descent with respect to the antenna’s elements coordinates, yielding a new optimized configuration. A custom penalty mechanism is also implemented, integrating various physical and design constraints into our optimization method. The effectiveness of our method is tested on the ten configurations with the lowest initial cost values, showing a further reduction in cost by 699% to 816%, with an average of 740%.