Soft Blends Analyzer (SOBA) was developed to link machine learning and computer simulation with Python interface. Using the SOBA, Lenard-Jones Potential particles dynamics were done to generating two dimensional images of the filler dispersion for the education data of Depp Learning (DL). The DL was performed to learn the relationship between the image features and the filler density. The technique was applied to bucky-paper (BP:Non-woven film of CNT). The tiling images of SEM images with different four magnifications were used as DL training data, and accurate classification model was obtained. The model enabled to predict the physical properties of BP not used for learning by the sum of multiplying the probability distribution of classification by the properties of the educated BP. Next, the Generative Adversarial Network (GAN) training was performed with the same SEM images. Finally, GAN's morphing method have made it possible to virtually generate images of BP mixing different CNTs and established a framework of virtual experiment method for predicting its physical properties with the DL model. By the framework, the physical properties of more than 1,700 types of bucky-paper were estimated, which has contributed to reduce the time of material development.