Toward a world model for corrosion science

Corrosion is a trajectory, not an event, yet current machine learning maps fixed inputs to outputs and cannot model reliably how a corroding interface responds to intervention. We argue that world models, deep learning frameworks that learn latent physical dynamics from observation sequences, can close this gap. A physically interpretable latent space anchored to corrosion mechanisms can enable prediction, counterfactual reasoning, and discovery. We chart a path toward realising it.