APPSolver: Adaptive Patch Partitioning for Point-Wise Ship Flow Prediction on Unstructured Meshes
This study introduces APPSolver, a point-wise flow prediction framework based on Adaptive Patch Partitioning (APP) to reduce computational cost in ship hydrodynamics. The method uses a deterministic quadtree representation with finer patches near the hull and coarser farther away, enabling efficient prediction through downsampling and recovery. Experiments show that APP offers computational benefits but does not universally outperform baselines in accuracy, highlighting a tra
01 ABSTRACT
The paper presents APPSolver, whose core is Adaptive Patch Partitioning (APP), a quadtree-based representation for handling non-uniform point sets from ship CFD simulations. APP allocates finer patches near the hull and coarser patches farther away, using downsampling and recovery for efficient prediction. Under a corrected evaluation protocol, APP-Transformer achieves lower computational cost (1.815 GFLOPs, 1.309 ms) and reduces MAE by 16.4%-24.9% compared to uniform partitioning. However, a persistence baseline achieves lower one-step MAE, indicating APP's benefit is mainly computational efficiency rather than predictive superiority. The authors note that the current absolute next-state objective does not establish accurate long-horizon dynamics.
02 KEY FINDINGS
- Propose APPSolver, a point-wise flow prediction framework based on Adaptive Patch Partitioning (APP)
- APP uses a deterministic quadtree representation with finer patches near the hull and coarser farther away
- Corrected evaluation protocol: temporal splitting, reuse training normalization, report three seeds
- APP-Transformer computational cost: 1.815 GFLOPs and 1.309 ms per forward
- Compared to uniform partitioning, APP reduces MAE by 16.4%-24.9%
AI GENERATED SUMMARY / DISCOVERED BY ARXIV CS.LG