Neural operators aim to learn mappings between function spaces from data, offering a flexible surrogate for parameter-to-solution maps of partial differential equations. With sufficient data and careful design, they can perform well on relatively simple problems; however, their accuracy deteriorates sharply in optimization and inference, which are central to our work and represent a broad class of PDE applications. This difficulty arises from their limited ability to generalize beyond training data and the instability of their predictive accuracy.
We address these challenges through two complementary approaches: our earlier residual-based correction method and an agent-based strategy. In the residual-based approach, neural operator predictions are refined by solving a linear variational problem for the model error. This method substantially improved accuracy in topology optimization and Bayesian inference, but at the cost of an additional linear solve at each inference. The agent-based approach rethinks correction as a learning problem: instead of auxiliary solves, agents are trained to correct operator predictions directly. These agent-based methods preserve efficiency while delivering notable accuracy gains. In applications such as optimization problems constrained by nonlinear Poisson equations, topology optimization with nonlinear elasticity, and inverse problems, they show promising improvements.
A key difficulty in learning corrections is that the errors frequently involve highly oscillatory modes that standard neural networks struggle to represent. We will highlight preliminary results from reinforcement learning agents, as well as parallel efforts on multi-level neural networks, developed by groups at Stanford University and Polytechnique Montréal, that aim to address the same bottleneck. The overarching goal of this talk is to connect these complementary directions and to invite discussion on open questions related to oscillatory corrections and the generalization of reliable neural operators.