MSML2022: Hierarchical partition of unity networks- Fast multilevel training.
Nathaniel Trask · 2022
Problem: Applying AI/ML data-driven models in high-consequence engineering applications require guarantees to establish notions of trust (1) How to provide convergence guarantees Designing architectures + optimizers that provide notion of "grid convergence" critical for verification and validation(2) How to build surrogates that guarantee stability, physical realizability + generalizability? Unification of mimetic PDE discretization, algebraic topology and inverse problemsTo appear on arxiv later this month: "Data-driven Whitney forms for structure-preserving control volume analysis"