GNN-Surrogate

GNN-Surrogate applies graph neural networks to build surrogate models that predict ocean climate simulation outputs for parameter space exploration on unstructured meshes.

Key Features:

• Surrogate modeling with GNN: Employs graph neural networks to predict simulation outputs from input parameters such as wind stress, reducing the need for repeated full simulations.
• Handling unstructured meshes: Represents unstructured-mesh simulation domains as graphs to manage irregular grids and spatial relationships.
• Hierarchical graphs and adaptive resolutions: Constructs hierarchical graphs with adaptive resolutions to focus training and computational effort where needed.
• Visualization: Supports visual mappings for exploring parameter-space effects on outputs such as temperature fields.

Scientific Applications:

• Ocean climate simulation parameter-space exploration: Enables efficient exploration of how input parameters influence ocean climate model outputs.
• Evaluation on MPAS-Ocean: Demonstrated quantitative and qualitative evaluation using the MPAS-Ocean simulation.

Methodology:

Creates graph-based representations of unstructured meshes and trains graph neural networks as surrogate models with hierarchical graphs and adaptive resolutions to predict simulation outputs from input parameters; evaluated on MPAS-Ocean.