Towards understanding glasses with graph neural networks

Abstraction: DeepMind GNN predicts particle mobility in glass to illuminate glass transition

Key points:

• Glass modeled as particles with short-range repulsive interactions — a relational, local structure naturally suited to graph neural networks
• Input graph: nodes = particles, edges = neighboring pairs within 2 particle diameters labeled by relative distance; output = predicted mobility per particle
• Mobility (average distance traveled over time) was regressed using simulated ground truth, then model internals were analyzed for physical insight
• Glass transition is a proxy for a broad class of complex systems: polymers, colloidal suspensions, granular materials, biological cell migration
• Practical motivation: stable amorphous glass structures could enable faster-dissolving drug delivery systems and better disordered materials
• Published in Nature Physics; glasses also serve as testbeds because they are easy to simulate and interpret via particle-based ML models

Connections: Deepmind · Graph Neural Networks · Physics Simulation · Scientific ML

Source: https://deepmind.com/blog/article/Towards-understanding-glasses-with-graph-neural-networks