Point Cloud CycleGAN

We develop a graph convolutional CycleGAN model that translates between simulated and experimental event representations in the Active-Target Time Projection Chamber. The model simultaneously learns two taks: 1) modeling detector response and noise behavior and 2) removing noise and completing tracks for data cleaning. We modified an existing TreeGAN architecture to create a CycleGAN which can be used for point clouds. The point cloud CycleGAN can be used to convert simulated data to experimental data and vice versa. Through this transformation noise can easily be removed from experimental data, and simulations can be made more realistic.