Applied Maths Seminar: Luo -- Decision-ready surrogates models via derivative-informed operator learning

Dingcheng Luo (Queensland University of Technology) will be visiting us next Wednesday
and will give a seminar in Carslaw 275 at 1pm.  We will go to lunch leaving Level 2 at
12:05 sharp.  All welcome, students get a free lunch.  

Title: Decision-ready surrogates models via derivative-informed operator learning 

Abstract: Decision-making with computational models often involves two interrelated
tasks: Bayesian inference (calibrating models from data) and stochastic optimisation
(optimising decisions under uncertainty).  These tasks typically require repeated
evaluations of the underlying computational model, which can be prohibitively expensive
when the model is based on PDEs with high-dimensional inputs.  Operator learning offers
a promising alternative by constructing surrogate models that learn the mapping from
input fields to PDE solutions, enabling rapid and repeated evaluations.  However, for
downstream decision-making tasks, not only the accuracy of the operator approximation
but also the fidelity of its derivatives are crucial, as gradients drive outer-loop
optimisation and inference procedures.  

In this talk, we present derivative-informed operator learning: an approach that
incorporates derivative information during training to improve both the predictive and
differential accuracy of learned surrogates.  We discuss the theoretical underpinnings
of this framework and practical training strategies that enable the efficient solution
of computationally intensive outer-loop tasks.