In this work, we propose two methods to design zero constellations for binary modulation on conjugate-reciprocal zeros (BMOCZ). In the first approach, we treat constellation design as a multi-label binary classification problem and learn the zero locations for a direct zero-testing (DiZeT) decoder. In the second approach, we introduce a neural network (NN)-based decoder and jointly learn the decoder and zero constellation parameters. We show that the NN-based decoder can directly generalize to flat-fading channels, despite being trained under additive white Gaussian noise. Furthermore, the results of numerical simulations demonstrate that learned zero constellations outperform the canonical, Huffman BMOCZ constellation, with the proposed NN-based decoder achieving large performance gain at the expense of increased computational complexity.