Exponential-Nemytskii Neural Operators: Learning Field Evolution Across Step Sizes (Q7414969)

We develop a novel neural operator that allows for accurate prediction at multiple time resolutions without dense observation, leveraging a first-order Duhamel approximation of semilinear dynamics to prescribe step-size dependence. Under suitable conditions, agreement between models at two step sizes controls their differences throughout the predicted range and we provide local identifiability results for both components.