Graph Neural Network (GNN)

A graph neural network (GNN) is a neural network architecture designed to process data that is naturally represented as a graph: a set of nodes (entities) connected by edges (relationships). Unlike standard neural networks that operate on fixed-size vectors or grids (like images), GNNs can handle variable-size, irregularly structured data.

The core operation in a GNN is message passing: each node aggregates information from its neighbors, updates its own representation, and passes updated information outward. After several rounds of message passing, each node’s representation encodes information about its local neighborhood across the graph.

Why graphs matter in science:

Many scientific systems are naturally graph-structured:

• Molecules are graphs of atoms (nodes) connected by bonds (edges) — GNNs have become the dominant architecture for molecular property prediction
• The Earth’s surface can be represented as a mesh of connected grid cells — GraphCast represents the atmosphere this way
• Crystal structures are graphs of atoms in a lattice — used in materials property prediction
• Protein structures can be represented as graphs of amino acid residues

Scientific applications:

• GraphCast (weather forecasting): represents the atmosphere as a multi-scale mesh of nodes, passing information between them to propagate weather patterns
• GNoME (materials discovery): operates on crystal structure graphs to predict stability
• Chemprop (chemistry): uses molecular graphs for property prediction

Limitations: GNNs can struggle with long-range dependencies (nodes far apart in the graph), and their expressiveness has theoretical limits related to graph isomorphism. Transformer-based architectures are increasingly competitive for some molecular tasks.

Related guides: Climate & Earth Science, Materials Science, Chemistry