Multi-resolution representation via quad-tree hierarchy and efficient spatial attention; uncertainty branch for confidence under feature noise and point sparsity.

Feature-based Cycle-aware Time Positional Encoding (FCPE), hierarchical multi-scale temporal attention for irregularly timed clinical events.

Integrates differential geometry with graph neural networks (GNNs); U-Net-like mesh pooling/unpooling to learn direction-sensitive 3D features.

Physics-guided hidden Markv tree (forest) model for flood mapping on Earth imagery and 3D terrains, also extended with CNN/GNN co-training via variational EM.

4D Swin-based surrogate for estuary tidal wave propagation (hindcast + up to 12-day forecast), extended to decade-long multi-process simulation.

Systematic taxonomy for data vs. model uncertainty; summarizes advantages and disadvantages across methods; identifies future research directions.

Neuro-symbolic grounding with spatial knowledge base; trains with inferred label uncertainty.

Decade-long ~100 m coastal 3D mesh (~6M cells) with core oceanographic variables + external forcings to benchmark high-resolution coastal simulation models.

Structure-aware dataset for self-supervised GCL on protein-ligand binding; high-resolution ground-truth complexes & diverse computational decoys.