In the dark depths of space, cold dense gas congeals into star-forming cores, eventually birthing bright blue stars.
Shrouded in thick clouds of gas and dust, newborn stars are initially invisible. Scientists resort to hydrodynamic simulations to understand the details of early star formation. One such simulation may take weeks to run on state-of-the-art hardware.
Ten stellar-nursery simulations were presented in Ballone et al. 2020a.
They form the basis for this art project.
For all its scientific accuracy, the output of our hydrodynamic simulations is not exactly photogenic. Form follows function, after all. Luckily, deep learning has made great strides, providing us with neural networks capable of creating realistic images from simulations.
We trained a cycleGAN on APOD pictures to translate the simulations into photorealistic images. CycleGAN is a neural architecture that has already found application in style transfer within automatic art generation (Zhu et al. 2017).
Ten neural-network generated images, one per simulation, are available on opensea.io as non-fungible tokens. These are unique digital assets that live on the Ethereum blockchain. Unlike the images themselves, they cannot be replicated and certify the ownership of the underlying art.