Please note: This PhD seminar will take place in DC 2314.
Lena Podina, PhD candidate
David R. Cheriton School of Computer Science
Supervisors: Professors Ali Ghodsi and Mohammad Kohandel
Physics Informed Neural Networks (PINNs) have been very successful in reconstructing entire ODE solutions using only a single point or entire PDE solutions with very few measurements of the initial condition. In this talk, I will introduce Universal Physics-Informed Neural Networks (UPINNs), a PINN-based approach that employs a neural network to learn a representation of one or more unknown terms in the differential equation. UPINNs are able to reconstruct the unknown terms in both ODEs and PDEs from sparse and noisy data. They can also be combined with symbolic regression to yield a closed form of the hidden term. In addition to the main paper, I will also discuss a follow-up work, where we applied UPINNs to discover the underlying drug action term for a chemotherapeutic in a pharmacodynamics model. Finally, I will talk about a work where we combined uncertainty quantification (in the form of conformal prediction) with PINNs.