Understanding aerosol dynamics is an important task for several applications, such as understanding the spread of infectious diseases or how fuel injection works in a combustion engines. It is very difficult to understand these dynamics at a microscopic scale, hence we generally operate at the macro-scale. This can be experiments with humans exhaling white smoke in an otherwise dark room and then observing the growth of the aerosol cloud. Such experiments have applications in testing the efficacy of masks and face-shield designs. Modeling the cloud growth will help us predict how a aerosol cloud (which can be the result of a cough) grows over time, given certain parameters.

We can model aerosol dynamics using partial differentials equations however they are quite challenging to solve numerically, especially for large data set. We also have to consider other interactions such as gas-fluid and aerosol-air interactions to model the dynamics properly.

However, in this project, we aim to learn aerosol dynamics directly from image data. The images are segmented using the U-Net architecture. And the cloud growth dynamics are modelled using Stochastic Differential Equations (SDEs), which are learnt using neural networks.