Neural operators offer fast approximations of PDE solution maps, but their accuracy can deteriorate as optimization, inference, or changing operating conditions take them beyond their training data. Their use therefore requires assessing individual predictions and understanding how approximation errors affect the quantities used in a design or decision. I will discuss our work on these questions, beginning with goal-oriented error estimation and residual-based correction. The corrector improves a neural-operator prediction through a linearized variational solve, without retraining. Under suitable regularity and stability assumptions, it yields local quadratic error reduction. Examples in PDE-constrained optimization and inference will illustrate its benefits, followed by the challenges of learning the correction itself.
I will then present computational-twin studies of magnetically actuated soft structures under changing stiffness and loading. These studies combine prediction-specific error assessment with selective finite-element evaluations and reuse the acquired solution paths to adapt the operator and its estimator. Adaptation improves predictions and subsequent magnetic-command selection, but improved forward accuracy does not consistently produce reliable error assessment. The talk will conclude with a discussion of what these results establish, where the methods remain limited, and the questions they raise for reliable scientific machine learning.