The Periodic AutoEncoder aims to identify the periodicity of motion, resulting in the generation of a latent that aids the effective training of neural networks. garment simulation datasets also have periodic patterns, thus It has opportunities for expansion on this domain. However, in contrast to motion data, garment data has high dimension, and each vertex composing the garment has interactions. As a result, training a Periodic AutoEncoder directly on raw garment simulation data presents daunting obstacles. To address this issue, we employ a Fully Convolutional Mesh Autoencoder. This Autoencoder helps to reduce dimension by representing complex garments as simplified graphs. In this thesis, we first validate the effectiveness of the Periodic AutoEncoder using a garment simulation dataset. We accomplish by comparing the reocnstruction of autoencoder latent and the ground truth. Subsequently, we use this latent to train neural networks and analyze the training outcomes to evaluate the potential of the Periodic AutoEncoder with the garment datasets.