Differentiable Simulations for Joint Parameterised Optimisation of Antenna Arrays

Many modern antenna systems rely on active electronically scanned array technology to reconfigure antenna systems in real-time. The scanning of arrays requires correct array excitation coefficients. To obtain excitation coefficients that satisfy an ample space of reconfigurability, a considerable number of inverse problems usually need to be solved. This is either done with heuristics, optimisers or through costly physical measurements. In this work, we discuss our experience with implementing a batchable and differentiable antenna array simulation tool and utilising it in combination with neural networks to solve large-scale excitation synthesis using joint parametric optimisation. The method enables us to train neural networks to solve a wide range of beam-shaping problems in an end-to-end manner. We have observed that the method is competitive with direct optimisation methods when a large number of problems need to be solved. However, the current approach does not yet achieve the same solution quality as a well-tuned direct optimisation method.