Reproduce “Ship-D: Ship Hull Dataset for Design Optimization using Machine Learning” on OpenResearch

About this paper

Noah J. Bagazinski, Faez Ahmed

Researchers at MIT developed SHIP-D, a publicly available dataset comprising 30,000 parameterized ship hulls with corresponding hydrodynamic drag measurements, to address the data scarcity hindering machine learning applications in ship design. This work introduced a novel 45-term parameterization for diverse hull forms and a residual neural network capable of predicting 32 wave drag coefficients with an R-squared of 0.969, leading to up to 60.7% drag reduction in optimized hull designs.

Key insights

• The proposed 45-term parameterization can accurately reconstruct diverse existing ship hulls, with normalized RMS Chamfer distances ranging from 0.0802% to 0.559% of the hull's length overall.
• The generated SHIP-D dataset effectively samples the feasible design space, providing a comprehensive resource for training machine learning models.
• A ResNet surrogate model can predict 32 wave drag coefficients with high accuracy (R-squared of 0.969) and achieves a 480x speedup compared to direct Michell Integral calculations, enabling rapid design optimization.

Summary by alphaXiv.