Generation of Synthetic Solar Images with GANs

Abstract

This works presents an alternative solution to the already existing physical simulations for the Sun's photosphere which are used to output a single tex- ture image from the Sun's surface. This simulations are slow and require huge computational power, a cheaper alternative in time and resources is introduced. Architectures for a Generative Adversarial Network and a Variational Au- toeconder are considered and trained in the generation of small image tiles (128x128px) that resemble the surface of the Sun around an area of approxi- mately 2:62 107km2 which with the proposed method can be composed into an image of arbitrarily any size given enough tiles. Di erent architectures along with ne tuning is used to obtain the best net- work possible in both cases. Their results are compared, but the Generative Adversarial Network shows a powerful improvement on the generation of said tiles compared to the Variational Autoencoder. Lastly some methodologies for stitching together the generated tiles are pre- sented including a technique that uses a Genetic Algorithm approach to modify the generated tiles.