Seminar series
Date
Thu, 29 Oct 2026
Time
12:00 -
12:30
Location
Lecture Room 4, Mathematical Institute
Speaker
Joseph Webb
Organisation
Mathematical Institute University of Oxford
Joseph Webb (Mathematical Institute, University of Oxford) is going to talk about "An optimisation framework for the well-conditioned training of physics-informed neural networks"
Physics-informed neural networks (PINNs) have emerged as a promising route to solving partial differential equations, yet they have long struggled to achieve high accuracy. The obstacle is increasingly understood to be one of optimisation, owing to a severely ill-conditioned loss landscape. This talk presents a new optimisation method that enables a vanilla PINN formulation to solve a wide range of PDE problems many orders of magnitude more accurately, and faster, than existing methods.
We introduce Taro, a new family of optimisers that, rather than focusing on the objective, navigates parameter space towards regions where the data can be best utilised. We describe its novel step-selection strategy and how an optimiser can be designed to leverage it accurately and efficiently. We also discuss recent advancements with our framework, and the bottlenecks that emerge once the optimisation barrier is lifted.
Further Information
For Joseph's bio please follow this link