Research map / Point clouds / Point cloud classification

A Pioneering Neural Network Method for Efficient and Robust Fuel Sloshing Simulation in Aircraft

AAAI 2025 · Point clouds › Point cloud classification · approach: Scene flow and dynamic point clouds · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a point-cloud neural network with Triangle Feature Fusion (combining CConv, antisymmetric ASCC and FC paths) for fast, stable fuel sloshing simulation, and releases the Fueltank dataset of 320k frames.

Problem

Simulating fuel sloshing in aircraft tanks with complex internal structures (ribs, holes) is computationally expensive with Navier-Stokes/grid/SPH solvers, and existing neural fluid simulators are unstable in such scenarios.

Approach

Treats fuel as a particle point cloud within a position-based fluids scheme and fuses features from a continuous convolution (CConv) path, a momentum-conserving antisymmetric convolution (ASCC) path, and a global fully connected path via learned sigmoid-weighted Triangle Feature Fusion modules, trained with multi-step rollout loss (W+T steps).

Key results

• On the Fueltank dataset, the proposed network achieves CD 1.149 mm at t+1 and sequence error 25.310 mm, versus 1.232 mm and 35.318 mm for DualFluidNet.
• Max density error drops to 0.008 g/cm3 versus 0.013 for DualFluidNet, with inference time 0.158 s per step versus 0.199 s.
• Takeoff simulation takes about 2 minutes versus nearly 10 hours for Flow3D, with comparable accuracy (reported).

Measured effects reported in the paper

Surprising findings

• CConv and DMCF overfit to a fixed gravity direction and, under rotations, predict near-zero internal forces so fluid simply falls under gravity.

Limitations

• Evaluated only on water-like fluid with a single fuel density and particle radius; generalization to other fluids is claimed but not shown.
• Long-term stability and accuracy compared against SPH ground truth rather than real measurements.
• Takeoff comparison with Flow3D is only qualitative in the main text.

Experimental setup

Datasets: Fueltank dataset (generated with DFSPH solver)

Benchmarks: Fueltank dataset aircraft takeoff scenario

Metrics: Chamfer Distance Earth Mover Distance n-frame sequence error max density error inference time

Baselines: CConv DMCF TIE DualFluidNet DFSPH Flow3D

Builds on: CConv ASCC (DMCF) DualFluidNet position-based fluids DFSPH

Closest papers by meaning

• Neural SPH: Improved Neural Modeling of Lagrangian Fluid Dynamics (ICML 2024 · similarity 0.84)
• Aerodynamic Coefficients Prediction via Cross-Attention Fusion and Physical-Informed Training (AAAI 2025 · similarity 0.82)
• Deep Learning-based Algebraic Reynolds Stress Closures for RANS simulations of Turbulent Flows (NeurIPS 2026 · similarity 0.82)
• AeroGTO: An Efficient Graph-Transformer Operator for Learning Large-Scale Aerodynamics of 3D Vehicle Geometries (AAAI 2025 · similarity 0.81)
• Symmetric Basis Convolutions for Learning Lagrangian Fluid Mechanics (ICLR 2024 · similarity 0.81)
• Metamizer: A Versatile Neural Optimizer for Fast and Accurate Physics Simulations (ICLR 2025 · similarity 0.81)