This paper presents a convolutional neural network approach for the design and optimization of single-inputmultiple-output (SIMO) structures for the sub-terahertz spectral range (140 GHz). Two SIMO structures with two output channels have been designed using the proposed neural network approach and an iterative algorithm as a reference. Both structures have been manufactured by means of fused deposition modeling 3D printing technique and verified experimentally. A new method of 3D modeling of the designed phase maps has been developed and applied to manufacture unintuitive structures optimized with neural networks.