High-throughput virtual screening (HTVS), which utilizes computational evaluation of the potential properties of materials, has led to the discovery of a large number of new compounds in the fields of electrocatalysis, battery cathode materials, and electrolytes. In order to reduce the cost of computational time required for structural relaxation in the HTVS process, researchers have developed models such as the "domain switching" model, which implements a "many-to-many" task for structural relaxation, however, this method is prone to overfitting in the case of multiple unrelaxed structures converging to the lowest energy structure. However, this approach is prone to overfitting in the case of multiple unrelaxed structures converging to the lowest energy structure. A team lead by Prof. Yousung Jung from Department of Chemical and Biomolecular Engineering, KAIST, South Korea, proposed a data-driven structural relaxation ML model, Cryslator (Crystal Structure Translator), based on the domain translation model. Unlike the conventional domain translation approach applied to the many-to-many problems as described above in chemistry, the approach mainly focuses on predicting the final relaxed structures from various unrelaxed structures (i.e., many-to-one problem). For this, the authors leverage the pix2pix domain translation model, which is widely used in the computer vision field, where the pair data (i.e., initial & final structure) are provided. More specifically, the pix2pix is adapted here to translate between the crystal hidden feature space of unrelaxed structures and that of relaxed structures, where both features are obtained from pre-trained crystal graph convolutional neural networks. To demonstrate the performance, the proposed model is applied to the X-Mn-O dataset (X=Mg, Ca, Sr and Ba) enumerated by the element substitutions. The Cryslator shows robust performance in predicting the crystal formation energies regardless of the relaxation type of input structures (i.e., unrelaxed, partially relaxed, or fully relaxed crystal structures) demonstrating a many-to-one mapping capability. Additional experiments show the reasonable performance on predicting the relaxed atomic structures as well as the considerable reduction of errors in several structural and electronic properties through translation. Finally, the authors discuss generalizability of the model with chemical space consisting of more diverse combinations of elements in periodic table by conducting additional experiments with Materials Project database.