This program uses a collection of existing planet formation simulation data to generate a new snapshot which contains qualitatively similar but numerically distinct data. This is done by using a Generative Adversarial Network (GAN) to model the posterior distributions of the features in the dataset and then randomly drawing new points from the model. The model is powered by CTGAN.

This was used in the final chapter of my PhD thesis to generate a larger set of late-stage initial conditions for terrestrial planet formation simulations. Although the bulk of my thesis involved using high-resolution simulations to model this process starting from the smallest gravitationally bound objects, planetesimals, I found that it was too computationally expensive to run a statistically robust sample of planetary systems all the way to completion. In particular, the early stages of growth are the most calculation-intensive and the simulations become cheaper as they evolve because the particle count diminishes as objects conglomerate. At the same time, broader regions of the planet-forming disk come into communication with each other and the system becomes chaotic. To fully understand the outcome of this process, one woud need to run a large number of simulations through the chaotic phase.

To circumvent this issue, I train the GAN on some of the intermediate simulation snapshots, and use the model to generate a much larger set of initial conditions that begin partway through the planet formation process. Because the ICs are all numerically distinct, I can run them to gain a broader picture of the possible outcomes of the dynamical chaos.