3D Underwater Path Planning via Generative Flow Field Surrogates

AUV LAR operational activities within the hull of an advancing host platform involve the complexity of navigating the three-dimensional wakes created by propellers. Planning paths around these hydrodynamic structures is difficult because uniform current simulation models are inadequate. Although high-fidelity RANS CFD simulation approaches offer the required accuracy, prohibitive computational cost precludes real-time use on-board. By addressing this problem, this work is the first to combine two conditional generative adversarial networks (cGANs), specifically a regularized PatchGAN and a self-attention 2D3DGAN, to act as a RANS CFD data surrogate. These generative models are integrated with a 3D energy weighted A* path planner. A hierarchical pipeline drives both generators to synthesize complete

128^3 voxel flow field volumes from scalar operating condition inputs. This method demonstrates end-to-end inference times of 28 to 146 microseconds, significantly lower than the hours it takes to compute a RANS solution. Methodology benchmarking is conducted against the four levels of uniform current, ground-truth CFD, PatchGAN, and self-attention 2D3DGAN. Using CFD wake surrogates and knowledge led to a 5.7 – 12.5% reduction in energy expenditure and a 77.8% reduction in high-velocity wake-core encounters in the 19,800 uniquely generated trajectories over 550 flow conditions.

Message: Hi FE, I need help with “3D Underwater Path Planning via Generative Flow Field Surrogates” in “Mechanical Engineering”