Optimizing the core structure at the receiver end of a wireless power transfer magnetic coupler significantly impacts its power density and system efficiency. The patterns in the ferrite structure of the Double-DD magnetic coupler remain largely unexplored. This paper reveals for the first time how the mutual inductance of the Double-DD magnetic coupler is affected by the number of magnetic cores, breaking the inherent impression that the mutual inductance is maximum when the magnetic cores are fully packed. Deep reinforcement learning (DRL) was used to optimize the core at the receiving end of the Double-DD magnetic coupler. A 20 kW prototype and experimental platform were constructed for verification. Compared to a conventional fully-filled core structure, the optimized receiver end achieved an 11.4% reduction in core volume and a 1.2% increase in mutual inductance, while maintaining essentially unchanged output power and efficiency.