Metamaterials with particularly designed microstructure may exhibit unconventional physical properties, such as negative index of refraction (NIR), negative Poisson’s ratio (NPR) or negative thermal expansion coefficient (NTEC). By introducing rotational degrees of freedom at each material point, deformation-mode couplings between tension/compression and torsion or bending can be obtained, as predicted by the non-centrosymmetric Cosserat mechanics for chiral materials. Such couplings are of great importance to the development of novel sensors. Traditionally, the design of metamaterials relies on human experiences through trial and error. By using deep convolution neural networks, such as VGG, we have developed a methodology to create metamaterials with desired chiral microstructures. The geometric data of chiral microstructures are provided, along with effective mechanical properties, to train the DNN model. The effective properties are from finite element calculations, as well as experimentally measured data. The geometry of chiral samples for DNN training is generated from a generative adversarial neural network. With our successfully trained DNN model, the inverse problem of searching a microstructure geometry for a given set of chiral properties can be solved efficiently. Optimization of chiral microstructure for maximum deformation mode coupling is hence accomplished with the trained DNN model, and verified by experimental data and brute force finite element calculations.