[论文解读] Generative Design of Ship Propellers using Conditional Flow Matching

Patrick Kruger, Rafael Diaz+3 more|arXiv (Cornell University)|Jan 29, 2026

Cavitation Phenomena in Pumps被引用 0

一句话总结

本文提出一个条件流匹配（CFM）生成模型，用以生成满足指定性能目标的船用螺桨几何形状，在具有数据增强的代理模型的参数化螺桨数据集上进行了验证。

ABSTRACT

In this paper, we explore the use of generative artificial intelligence (GenAI) for ship propeller design. While traditional forward machine learning models predict the performance of mechanical components based on given design parameters, GenAI models aim to generate designs that achieve specified performance targets. In particular, we employ conditional flow matching to establish a bidirectional mapping between design parameters and simulated noise that is conditioned on performance labels. This approach enables the generation of multiple valid designs corresponding to the same performance targets by sampling over the noise vector. To support model training, we generate data using a vortex lattice method for numerical simulation and analyze the trade-off between model accuracy and the amount of available data. We further propose data augmentation using pseudo-labels derived from less data-intensive forward surrogate models, which can often improve overall model performance. Finally, we present examples of distinct propeller geometries that exhibit nearly identical performance characteristics, illustrating the versatility and potential of GenAI in engineering design.

研究动机与目标

• Develop a generative model that can propose propeller designs achieving predefined performance targets.
• Leverage a fully parametric propeller dataset and low-fidelity simulations to train the model.
• Evaluate accuracy and diversity of generated designs and investigate data augmentation with surrogate models.
• Explore the effect of reduced training data and surrogate-based augmentation on model performance.

提出的方法

• Use a flow-based generative framework, specifically Conditional Flow Matching (CFM), to model p(p|l) where p are propeller design parameters and l are performance labels.
• Represent propeller geometry with six design variables and derive performance labels via a vortex lattice method-based OpenProp simulation.
• Train a neural ODE-based flow with a conditional vector field to transport samples from a simple source distribution to the target design distribution.
• Incorporate a surrogate model ensemble to enable data augmentation and speed up validation of label predictions.
• Evaluate generated designs by running the original simulation to obtain true performance labels and compare to targets.

实验结果

研究问题

• RQ1 Can Conditional Flow Matching generate propeller designs that meet given performance targets (η*, J*, kT*)?
• RQ2 How accurate are generated designs when evaluated with the original OpenProp/CFD workflow?
• RQ3 Do generated designs exhibit meaningful geometrical diversity for a fixed target performance?
• RQ4 Does data augmentation with surrogate models improve model performance, especially with limited training data?
• RQ5 What is the trade-off between data amount and model accuracy in this inverse-design setting?

主要发现

• CFM can generate multiple diverse propeller designs that achieve specified target performance labels.
• Generated designs show high accuracy when compared to target labels across η*, J*, and kT* after evaluation with the simulation workflow.
• The method yields a meaningful variety of geometries even for the same performance target, including designs with different numbers of blades.
• Surrogate-model-based data augmentation improves accuracy for complex targets (notably kT*) when training data is limited, with varying effects across labels.
• Data augmentation benefits are most pronounced for complex labels and small initial datasets, while less impactful for simpler labels with strong design correlations.