The work carried out reports on a new machine learning-based software prototype for designing and optimising phased array antennas. The prototype can rapidly and accurately compute excitation coefficients for large-scale antenna arrays across diverse configurations and far-field requirements. The prototype relies on tailored neural networks integrating structural physics and domain knowledge using an encoder-decoder structure. Results indicate that the prototype has a significant potential to outperform traditional synthesis methods, offering a promising new tool for antenna engineers, enabling near real-time design and optimisation for phased antenna arrays.