Optimal modulation techniques such as optimized pulse patterns (OPPs) are essential for achieving efficient operation of electrical drives. Established OPP control methods often encounter limitations, including computational complexity and lookup table discontinuities. This work proposes a novel deep learning-based approach for OPP design using differentiable programming. The idea is to train a policy represented by a deep neural network (DNN) that maps motor working points of a Permanent Magnet Synchronous Motor (PMSM) to reference OPPs. Since there can be several locally optimal solutions distributed over the working space, we introduce a model structure that ensures optimality across the entire working space. For each working point, the DNN learns multiple candidate OPPs directly associated with a value metric term. For online operation, we use the value term to select the best candidate of the applied working point. The trained model matches the performance of state-of-the-art OPPs design approaches, emphasizing the effectiveness of the multi-solution approach in achieving a globally optimal policy over the entire working space. The work serves as a foundation for addressing the limitations of state-of-the-art concepts that arise when using complex motor models.