Deep Learning for Physics Discovery Winter 2021

A variational autoencoder is employed to gain insights into two physical systems: a particle settling under gravity in a viscous fluid, and a chemical reactor. The network architecture allows the interpretation of neuron activations in the latent layer in terms of physical parameters governing system behavior. The network takes as input a set of observations, a question regarding the state of nature under particular conditions, and an answer describing the state of nature under said conditions. The architecture was shown to yield accurate, interpretable results in both physical systems of interest, correctly discerning the parameters governing the physical systems and storing them in the latent neuron activations, and providing additional physical insight into the systems at hand.