The goal of our project is to develop a physically-informed surrogate simulator for spatiotemporal phenomenon suitable for us in robotic decision-making that overcomes the computational challenges of traditional numerical models. In particular, we will be using our robot for exploration of hydrothermal plumes, typically characterized by partial differential equations (PDEs). To enable our robot to optimize sampling trajectories for plume exploration, we need to effectively forward-simulate in real-time the fluid dynamics surrounding the robot. To meet this real-time demand, we will use deep learning and statistical methods to train a surrogate simulator that approximates forward solutions of the PDE system (on scale of seconds).

I am super interested in how we can leverage statistics and deep learning to tackle modeling complex, dynamic environments. In particular, I’ve been excited to delve into this novel problem of using learning to develop real-time PDE simulators for exploration of environmental phenomena.