Abstract
Modern science and engineering rely on high-fidelity simulations and experiments that take hours to days per run, creating a fundamental bottleneck for design exploration. Machine-learning surrogates can ease this bottleneck, but most methods face serious constraints in realistic deployment. This talk presents two works addressing these challenges: (1) BSMS-GNN, a multi-scale graph neural network that preserves connectivity when learning physics on large-scale unstructured meshes; and (2) VICON, an in-context-learning neural operator that adapts across different physics and operating conditions without retraining. The talk closes with a brief discussion of recent work on LLM agents that orchestrate such surrogate toolkits for research automation.
For more info, please follow this link.