Problem

Traditional ship propeller design workflows often depend on experienced engineers, trial-and-error, or computationally expensive simulation-driven design (SDD) techniques. While SDD optimizes a propeller\'s geometry for predicted performance, it works in a \'forward\' manner—evaluating a design\'s performance. The emerging need is for \'inverse design,\' where a designer specifies desired performance characteristics, and the system generates suitable geometries. A significant hurdle for applying generative artificial intelligence (GenAI) in engineering design is the limited availability of high-quality, domain-specific data due to proprietary concerns, hindering the development of robust GenAI models for complex tasks like propeller design.

Method

The authors propose a generative artificial intelligence approach utilizing Conditional Flow Matching (CFM) for the inverse design of ship propellers. This method establishes a bidirectional relationship between propeller design parameters (e.g., number of blades, pitch, chord, camber) and simulated noise, conditioned on desired performance metrics such as maximal efficiency (η\*), optimal advance ratio (J\*), and thrust coefficient (kT\*). Training data is generated through numerical simulations using a vortex lattice method (VLM) via OpenProp software, with propeller geometries parameterized in CAESES. CFM, an efficient technique for training Neural Ordinary Differential Equations (NODEs), learns a time-dependent vector field by direct regression against target vector fields constructed on a per-sample basis, bypassing the computational intensity of traditional ODE solvers during training. Furthermore, a data augmentation strategy is introduced where pseudo-labels are derived from faster forward surrogate models. This augmentation helps to enhance model accuracy and stability, particularly when the original simulation data is scarce. The CFM model itself is implemented as an 8-layer feed-forward neural network.

Results

The trained Conditional Flow Matching (CFM) models successfully generated diverse ship propeller designs that accurately corresponded to a wide range of specified performance requirements. When tested against a dataset of 1000 target performance labels, the generated designs exhibited high accuracy for all three performance metrics (η\*, J\*, and kT\*). Specifically, for a fixed target performance of η\*=0.8, J\*=1.0, and kT\*=0.1, the model demonstrated its capability to produce a variety of geometrically distinct propeller designs, all of which achieved performance characteristics closely matching the targets. The study also highlighted the effectiveness of synthetic data augmentation using surrogate models; this strategy yielded significant improvements in accuracy for performance labels with more intricate parameter relationships, particularly when the initial amount of training data was low. For instance, `kT*` saw the greatest improvement through augmentation. The models were computationally efficient, with a CFM model training in about 15 minutes, significantly faster than the 6 hours required to generate the full training dataset via direct simulation.

Implications

This research demonstrates that generative Conditional Flow Matching models provide a powerful and efficient framework for inverse design in engineering, specifically for ship propellers. By enabling the generation of diverse design alternatives that precisely meet specified performance requirements, the approach significantly enhances the design workflow, moving beyond traditional forward simulation models. This newfound capability offers engineers greater flexibility to choose optimal designs based on additional criteria like manufacturing constraints. Furthermore, the successful implementation of data augmentation through surrogate models presents a critical solution for addressing data scarcity in specialized engineering domains, potentially accelerating the broader adoption of generative AI. Future research could expand on this by incorporating more complex propeller parameters and performance criteria, such as cavitation risk, and integrating higher-fidelity computational fluid dynamics (CFD) methods, which would further underscore the necessity and value of efficient generative design strategies.