What is a physics-informed neural network (PINN)?
Definition
A physics-informed neural network (PINN) is a neural network trained to satisfy the governing equations of a physical system, not only to fit data. The equations, such as the Navier-Stokes equations, enter the training loss as a penalty, so the network can learn a solution from little data or infer hidden quantities from measurements.
Example
In fluid mechanics, a PINN has been used to infer the velocity and pressure of a flow from images of dye or smoke, by building the Navier-Stokes equations into its training. A standard PINN solves one problem instance, so a new geometry usually means training again.
Sources
- Raissi, Perdikaris, Karniadakis, "Physics Informed Deep Learning (Part I)," 2017, arxiv.org/abs/1711.10561
- Raissi, Yazdani, Karniadakis, "Hidden Fluid Mechanics," 2018, arxiv.org/abs/1808.04327