With increasing computational power, scientists often employ high-fidelity simulations to study complex scientific phenomena. However, these simulations remain slow and costly in terms of computation and storage, prompting the use of faster low-fidelity alternatives. Yet, the distribution gap between low- and high-fidelity data, due to missing fine-scale details and simplified physics, hampers accurate scientific understanding. To overcome these challenges, we propose a neural operator-based framework for multi-fidelity prediction on unstructured data. Our framework leverages a graph neural operator to effectively map low-fidelity data to the high-fidelity counterparts. It further incorporates a spectral-based module that captures fine-scale details, enhancing the reconstruction of high-fidelity fields. Extensive experiments across diverse datasets demonstrate that our method consistently surpasses strong baselines, including state-of-the-art neural operators and learning methods for unstructured data, highlighting its robustness and effectiveness in bridging fidelity gaps.
Chen et al. (Thu,) studied this question.