Location
MC 5501
Candidate
Jonathan Gallagher | Applied Mathematics, University of Waterloo
Title
Towards Control-Oriented Neural Surrogate Models for PDE-Governed Systems
Abstract
Obtaining solutions to governing PDEs remains computationally intensive in many applications. Advances in deep learning have produced architectures capable of approximating PDE solution operators, offering a potentially advantageous alternative to conventional solvers in some settings.
Such surrogate models possess properties that are potentially attractive for applications in control, including a low inference cost, differentiability, and capacity for deployment on resource-constrained hardware. Motivated by this, we aim to study the integration of fully or partially learned dynamical models within control frameworks, with particular emphasis on controller design, uncertainty quantification, inverse problems and real-time decision making.
We provide some evidence in this direction in the form of a custom operator-learning architecture that predicts the behaviour of a wetted-wall-column (WWC) carbon capture device across varying operating conditions. We demonstrate its use as a fast, differentiable surrogate for online process monitoring and optimization.