Answers are grounded in this paper and cite their sources.
Resolving the spatially high-dimensional hydrodynamic behavior of pump-jet propulsors (PJPs) through experiments or numerical simulations imposes prohibitive computational costs, especially during iterative design. While deep learning surrogates offer substantial acceleration for flow-field prediction through data-driven modeling, architecture selection remains largely empirical and often overlooks the alignment between neural inductive bias and geometric topology. To this end, this study compares three representative deep learning models: a node-based model ( MLP ), a point-cloud-based model ( PointNet ), and a mesh-based model ( GraphNet ). A hierarchical benchmarking framework evaluates these architectures across five PJP subdomains, including the PJP inlet plane, rotor-outlet wake, blade-surface manifolds, and cascade channel. Quantitative results reveal a distinct efficiency–fidelity trade-off. The MLP provides rapid global-trend prediction in smooth inlet- and wake-dominated regions, but it is less robust in boundary-extrapolation tests and on topology-sensitive blade surfaces. The PointNet incurs high memory costs for dense volumetric data. Both connectivity-agnostic models exhibit spectral bias on non-Euclidean blade surfaces and underresolve steep pressure gradients. Conversely, the GraphNet uses mesh topology to capture fine-scale spatial structures and curvature-induced flow features, showing robust performance across the tested scenarios at a higher computational cost. Besides, the GraphNet retains its relative advantage in two additional extrapolation cases. Furthermore, the mechanistic ablation shows that vectorial relative coordinates improve geometric and structural fidelity during boundary extrapolation. Ultimately, this work establishes a topology-dependent model-selection guideline for PJP flow-field prediction.
No takes yet. Share an insight, caveat, or question.
Hao et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: