Future Research

---

Initial Discoveries

We attempted to develop the surrogate model as one would in a traditional machine learning scenario. However, the dataset being used with this model was generated from a simulation that extensively used random number generation. With the guidance of our research advisor, we came to the conclusion that deterministic models would not adequately find a correlation in the data.

Next Steps

After learning that deterministic models would not fit well with the data, we decided to turn to generative models for our surrogate. Although creating a generative model can be quite complex, we found several options for relatively smaller and simpler models that could feasibly be developed in the near future. These models include: Variational Autoencoders (VAE), Generative Adversarial Networks (GAN) (specifically the mini-GAN versions), and Monte Carlo Simulation methods. While the scale of these models are much larger than those of our initial designs, we believe that leveraging our existing resources with the University of Virginia’s Rivanna supercomputer would make these options more realistic.