FLUID: A Neural Operator-Based Framework for Learning Multi-Fidelity of Unstructured Data.

Area of Science:

• Computational Science
• Data Science
• Scientific Machine Learning

Background:

• High-fidelity simulations are crucial for complex phenomena but are computationally expensive.
• Low-fidelity simulations offer speed but suffer from data distribution gaps due to missing details.
• Bridging this fidelity gap is essential for accurate scientific understanding and prediction.

Purpose of the Study:

• To introduce a novel neural operator framework for multi-fidelity prediction on unstructured data.
• To effectively map low-fidelity simulation data to high-fidelity counterparts.
• To enhance the reconstruction of fine-scale details in scientific fields.

Main Methods:

• Utilized a graph neural operator to learn mappings between different fidelity data.
• Incorporated a spectral-based module to capture and reconstruct fine-scale details.
• Applied the framework to diverse datasets with unstructured data.

Main Results:

• The proposed framework consistently outperformed strong baselines, including existing neural operators.
• Demonstrated robust and effective bridging of the fidelity gap in scientific data.
• Achieved superior reconstruction of high-fidelity fields from low-fidelity inputs.

Conclusions:

• The neural operator-based framework effectively addresses the challenges of multi-fidelity prediction.
• This approach enhances the accuracy and reliability of scientific simulations and data analysis.
• Offers a promising solution for leveraging low-fidelity data to approximate high-fidelity results.