LatentDDM Darcy and Airfoil Datasets for Varying-Geometry Neural Operator Learning

This unified dataset release accompanies LatentDDM: Reusing Pretrained Local Neural Operators Across Domains and Operating Regimes. It supports two complementary benchmarks for adapting physical surrogate models. The Darcy collection contains learned local boundary-value problems, source-domain simulations, and fixed-grid-spacing domain extensions with 4 x 4, 8 x 8, and 16 x 16 covers. The Airfoil collection contains 256 synchronized level-5 AMReX/incflo trajectories for high-Reynolds-number pitching airfoils, together with metadata-only in-distribution and out-of-distribution pitch-rate splits. Airfoil observations use a nominal 0.025 time cadence while retaining each trajectory's terminal cycle endpoint. The archives preserve HDF5 arrays, paper-aligned split definitions, schemas, checksums, and dataset documentation. The record is prepared for the Physical World AI workshop at NeurIPS 2026.