arxiv: 2604.22224 · v1 · submitted 2026-04-24 · 💻 cs.CE · cs.LG· physics.comp-ph

Recognition: unknown

AI-Driven Performance-to-Design Generation and Optimization of Marine Propellers

Leah Chen , Keni Chih-Hua Wu , Boon Tat Chia , Xiuqing Xing , Jian Cheng Wong

Authors on Pith no claims yet

Pith reviewed 2026-05-08 09:25 UTC · model grok-4.3

classification 💻 cs.CE cs.LGphysics.comp-ph

keywords marine propellergenerative AIdesign generationperformance predictionevolutionary optimizationphysics simulationdiffusion model

0 comments

The pith

AI generates marine propeller designs directly from performance targets using physics data and optimization.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a generative AI framework for creating marine propeller geometries directly from performance specifications like target thrust and power. It starts by generating a large dataset of over 20,000 propellers with their simulated open-water curves through physics-based modeling. The framework includes a conditional model to propose designs, a fast neural surrogate to predict performance, and evolutionary optimization to refine them under practical constraints. This setup produces plausible designs that meet the targets across various conditions while cutting down on the lengthy traditional design iterations.

Core claim

The proposed three-module framework, built on a physics-synthesized database of 20,000+ propeller geometries, enables direct performance-to-design generation of hydrodynamically plausible marine propellers that match prescribed targets, with latent diffusion models providing greater design diversity than variational autoencoders.

What carries the argument

A modular AI system with a Conditional Generation Model for proposing geometries from specs, a neural Performance Prediction Model as a fast surrogate, and an evolutionary optimization stage to enforce constraints like thrust requirements and geometric bounds.

If this is right

The framework generates designs that satisfy target thrust and power while respecting limits on blade-area ratio and thickness.
Latent diffusion-based generators produce more varied designs than conditional variational autoencoders under identical conditions.
Design-iteration time decreases substantially compared to expert-guided traditional methods.
Reliance on expensive high-fidelity simulations is reduced to only the final validation stage.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

Similar pipelines could be adapted for designing other ship components such as hulls or rudders.
Validating the generated designs through physical prototypes would test the assumption that simulations translate to reality.
Further integration of more detailed fluid dynamics could enhance the framework for high-speed or off-design conditions.

Load-bearing premise

The physics-based simulations provide open-water performance curves that are accurate enough proxies for actual propeller behavior in operation.

What would settle it

Conducting a physical experiment or sea trial on an AI-generated propeller to verify whether its measured performance matches the simulated targets and the design remains stable.

Figures

Figures reproduced from arXiv: 2604.22224 by Boon Tat Chia, Jian Cheng Wong, Keni Chih-Hua Wu, Leah Chen, Xiuqing Xing.

**Figure 1.** Figure 1: FIGURE 1: (Left) diameter, (right) chord length, maximum thick view at source ↗

**Figure 2.** Figure 2: FIGURE 2: (Left) skew, (middle) rake, and (right) pitch [ view at source ↗

**Figure 3.** Figure 3: FIGURE 3: Open-water characteristics of two propellers. view at source ↗

**Figure 4.** Figure 4: FIGURE 4: Predicted vs. actual for (left) thrust coefficient view at source ↗

**Figure 6.** Figure 6: FIGURE 6: Generated designs from the cVAE model trained with view at source ↗

**Figure 7.** Figure 7: FIGURE 7: (Left) relative view at source ↗

**Figure 8.** Figure 8: FIGURE 8: Distribution of designs generated by cVAE. The “X” view at source ↗

**Figure 9.** Figure 9: FIGURE 9: Generated designs with different view at source ↗

**Figure 10.** Figure 10: FIGURE 10: (Left) relative view at source ↗

**Figure 11.** Figure 11: FIGURE 11: Generated design sample distribution from the latent view at source ↗

**Figure 12.** Figure 12: FIGURE 12: Comparison of generated designs from (top) cVAE view at source ↗

**Figure 13.** Figure 13: FIGURE 13: Out-of-distribution design reconstruction. Solid lines view at source ↗

read the original abstract

AI is increasingly used to accelerate engineering design by improving decision-making and shortening iteration cycles. Application to marine propeller design, however, remains challenging due to scarce training data and the lack of widely available pretrained models. We address this gap with a physics-based data generation pipeline and a generative-AI framework for direct performance-to-design generation tailored to marine propellers. First, we build a database of over 20,000 four- and five-bladed propeller geometries, each accompanied by simulated open-water performance curves. On top of this dataset, we develop a three-module design framework: (1) A Conditional Generation Model that proposes candidate geometries conditioned on design specifications such as target thrust, power, and diameter. (2) A Performance Prediction Model, implemented as a neural-network surrogate, that predicts thrust, torque, and efficiency in milliseconds, enabling rapid evaluation of generated designs. (3) A design refinement stage that applies evolutionary optimization to enforce practical constraints such as required thrust under power limits and bounds on blade-area ratio and thickness. Experimental results over a range of operating conditions show that the framework can generate hydrodynamically plausible propeller designs that match prescribed performance targets while substantially reducing design-iteration time relative to the traditional expert-guided refinement. Latent diffusion-based generator produces more diverse designs under the same conditions than the conditional variational autoencoder, suggesting a stronger capacity for design-space exploration with diffusion models. By coupling physics-based data synthesis with modular AI models, the proposed approach streamlines the propeller design cycle and reduces reliance on expensive high-fidelity simulations to final validation stages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper builds a 20k physics-simulated propeller dataset and shows a workable three-module pipeline with diffusion generation plus surrogate-assisted evolutionary refinement, but the performance-matching claims rest on unverified surrogate outputs.

read the letter

The main thing to know is that the authors created a large database of over 20,000 four- and five-bladed propeller geometries with simulated open-water curves, then wrapped it in a conditional latent diffusion generator, a millisecond neural surrogate, and evolutionary optimization to turn performance targets into candidate shapes. This is a direct application to marine propeller design rather than a general new method.

Referee Report

2 major / 2 minor

Summary. The paper presents a three-module AI framework for performance-to-design generation of marine propellers. It first constructs a dataset of over 20,000 four- and five-bladed geometries with physics-based open-water performance curves, then trains a conditional generative model (latent diffusion or CVAE) to propose candidate shapes from targets such as thrust, power, and diameter; a neural-network surrogate (Performance Prediction Model) for millisecond-scale thrust/torque/efficiency evaluation; and an evolutionary optimization stage to enforce geometric and performance constraints. The central claim is that the resulting designs are hydrodynamically plausible, match prescribed targets, and substantially reduce iteration time relative to expert-guided refinement, with diffusion models yielding greater design diversity.

Significance. If the surrogate-guided optimizations are shown to produce designs whose performance holds under the original physics simulator, the work could meaningfully accelerate marine propeller design cycles by shifting most iteration to fast ML surrogates while reserving high-fidelity simulation for final validation. The scale of the physics-generated dataset and the modular separation of generation, prediction, and refinement are clear strengths; the explicit comparison of diffusion versus VAE generators for exploration capacity is also useful. The absence of post-optimization validation against the data-generating model, however, limits the strength of the matching and plausibility claims.

major comments (2)

[Abstract and Experimental Results] Abstract and Experimental Results: The headline claim that generated designs 'match prescribed performance targets' and are 'hydrodynamically plausible' rests exclusively on predictions from the neural-network Performance Prediction Model. No quantitative error metrics (MAE, R², etc.) on held-out test geometries, no baseline comparisons against existing surrogate or optimization methods, and no post-optimization re-evaluation of final designs with the original physics-based simulation pipeline are reported. Because the evolutionary refinement loop uses only the surrogate, even modest surrogate bias on out-of-distribution blade shapes produced by the generator would invalidate the reported target matches.
[Design Refinement Stage] Design Refinement Stage: The evolutionary optimization enforces thrust/power targets and geometric bounds (blade-area ratio, thickness) via the surrogate; without a final verification step that re-simulates the optimized geometries in the original high-fidelity pipeline, the claim of reduced iteration time and target compliance cannot be confirmed under the data-generating model itself.

minor comments (2)

The abstract states that the framework 'substantially reduc[es] design-iteration time' but supplies no quantitative timing data, number of iterations, or comparison protocol against the traditional expert-guided process.
Training details for the neural-network surrogate and the two generative models (architecture, loss functions, hyperparameter selection, dataset split) are not summarized, making reproducibility of the reported performance and diversity results difficult.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed comments. We agree that the original manuscript would benefit from explicit quantitative validation of the Performance Prediction Model and post-optimization verification against the high-fidelity physics simulator. We have revised the manuscript to include these elements, which directly strengthen the claims regarding target matching and hydrodynamic plausibility.

read point-by-point responses

Referee: [Abstract and Experimental Results] Abstract and Experimental Results: The headline claim that generated designs 'match prescribed performance targets' and are 'hydrodynamically plausible' rests exclusively on predictions from the neural-network Performance Prediction Model. No quantitative error metrics (MAE, R², etc.) on held-out test geometries, no baseline comparisons against existing surrogate or optimization methods, and no post-optimization re-evaluation of final designs with the original physics-based simulation pipeline are reported. Because the evolutionary refinement loop uses only the surrogate, even modest surrogate bias on out-of-distribution blade shapes produced by the generator would invalidate the reported target matches.

Authors: We agree that reporting surrogate accuracy metrics and performing post-optimization high-fidelity validation are necessary to fully substantiate the claims and mitigate concerns about potential bias. In the revised manuscript we have added a new subsection detailing the Performance Prediction Model's accuracy on a held-out test set of 2,000 geometries, including MAE, RMSE, and R² for thrust, torque, and efficiency. We have also re-simulated 50 representative optimized designs with the original physics-based pipeline; the results show average discrepancies below 5% in thrust and efficiency, confirming that target matches hold under the data-generating model. Finally, we have included a baseline comparison against a conventional evolutionary optimization procedure (without the generative stage) to quantify the iteration-time savings of the proposed framework. revision: yes
Referee: [Design Refinement Stage] Design Refinement Stage: The evolutionary optimization enforces thrust/power targets and geometric bounds (blade-area ratio, thickness) via the surrogate; without a final verification step that re-simulates the optimized geometries in the original high-fidelity pipeline, the claim of reduced iteration time and target compliance cannot be confirmed under the data-generating model itself.

Authors: We acknowledge that the original submission lacked an explicit final verification step with the high-fidelity simulator. The revised manuscript now incorporates such a verification: after surrogate-guided evolutionary optimization, a subset of the resulting geometries is re-evaluated in the original physics pipeline. These checks confirm that the designs satisfy the prescribed thrust and power targets within engineering tolerances and that the geometric constraints are respected. This addition supports the claim of substantially reduced iteration time, as the majority of evaluations occur via the millisecond-scale surrogate while high-fidelity simulation is reserved for final validation. revision: yes

Circularity Check

0 steps flagged

No circularity; framework uses external physics simulations and standard ML training

full rationale

The paper constructs a dataset via independent physics-based simulations (20k geometries with open-water curves), then trains a conditional generator and a separate neural surrogate for performance prediction. Evolutionary optimization enforces targets using the surrogate, but this is standard surrogate-assisted optimization rather than any result being defined in terms of itself. No equations or claims reduce a prediction to its own fitted inputs by construction, and no self-citation chain bears the central load. The pipeline remains self-contained against the external simulation benchmark.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on the assumption that the 20,000 simulated geometries adequately sample the relevant design space and that standard neural-network training will generalize to unseen operating conditions; no explicit free parameters or invented physical entities are introduced beyond conventional ML hyperparameters.

free parameters (1)

neural network hyperparameters
Architecture depth, learning rate, and latent dimension choices are fitted during training but not enumerated in the abstract.

axioms (2)

domain assumption Simulated open-water curves are faithful enough for training and evaluation
Invoked when the authors treat the generated database as ground truth for both generator training and surrogate supervision.
domain assumption Generated blade geometries remain within manufacturable and structurally feasible bounds
Implicit in the claim that designs are hydrodynamically plausible without additional manufacturing or stress checks.

pith-pipeline@v0.9.0 · 5598 in / 1489 out tokens · 17637 ms · 2026-05-08T09:25:29.866363+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

16 extracted references

[1]

The marine propeller design spiral

Praefke, Eckhard. “The marine propeller design spiral.” Second International Symposium on Marine Propulsors. June. Hamburg. Germany. 2011

2011
[2]

Data-driven de- sign exploration method using conditional variational au- toencoder for airfoil design

Yonekura, Kazuo and Suzuki, Katsuyuki. “Data-driven de- sign exploration method using conditional variational au- toencoder for airfoil design.”Structural and Multidisci- plinary OptimizationVol. 64 No. 2 (2021): pp. 613–624

2021
[3]

Pre- liminary Study of Airfoil Design Synthesis Using a Condi- tional Diffusion Model and Smoothing Method

Yonekura,Kazuo,Oshima,YutaandAichi,Masaatsu.“Pre- liminary Study of Airfoil Design Synthesis Using a Condi- tional Diffusion Model and Smoothing Method.”Compu- tationVol. 12 No. 11 (2024): p. 227. 9 TABLE 6: Summary of expert or reference propeller designs by vessel type. Hull type Count Hull length (m) No. shafts Engine power (kW) RPM Prop. diam. (mm) No...

2024
[4]

An inverse design method for supercritical airfoil based on conditional generative models

Jing,WANG,Runze,LI,Cheng,HE,Haixin,CHEN,Chen, ZHAI et al. “An inverse design method for supercritical airfoil based on conditional generative models.”Chinese Journal of AeronauticsVol. 35 No. 3 (2022): pp. 62–74

2022
[5]

On surrogate methods in propeller optimisation

Vesting, Florian and Bensow, Rickard E. “On surrogate methods in propeller optimisation.”Ocean engineering Vol. 88 (2014): pp. 214–227

2014
[6]

Evolutionary optimisation for reduc- tionofthelow-frequencydiscrete-spectrumforceofmarine propellerbasedonadata-drivensurrogatemodel

Jiang, Jing-Wei, Yang, Yang, Ren, Tong-Wei, Wang, Fei and Huang, Wei-Xi. “Evolutionary optimisation for reduc- tionofthelow-frequencydiscrete-spectrumforceofmarine propellerbasedonadata-drivensurrogatemodel.”Journal of Marine Science and EngineeringVol. 9 No. 1 (2020): p. 18

2020
[7]

A unified cross-series marine propeller design method based on machine learning

Tadros, Mina, Shi, Weichao, Xu, Yunxin and Song, Yang. “A unified cross-series marine propeller design method based on machine learning.”Ocean EngineeringVol. 314 (2024): p. 119691

2024
[8]

An improved parametric model for marine pro- peller: iPM4MP

Zheng, Long, Chen, Shunhuai, Ji, Shengchen and Chen, Zhenlin. “An improved parametric model for marine pro- peller: iPM4MP.”Ocean EngineeringVol. 287 (2023): p. 115712

2023
[9]

Machine learning for naval architecture, ocean and marine engineering

Panda, JP. “Machine learning for naval architecture, ocean and marine engineering.”Journal of Marine Science and TechnologyVol. 28 No. 1 (2023): pp. 1–26

2023
[10]

Iden- tification of the geometric design parameters of propeller blades from 3D scanning

Njaastad, Eirik B, Steen, Sverre and Egeland, Olav. “Iden- tification of the geometric design parameters of propeller blades from 3D scanning.”Journal of Marine Science and TechnologyVol. 27 No. 2 (2022): pp. 887–906

2022
[11]

Butterworth-Heinemann (2018)

Carlton, John.Marine propellers and propulsion. Butterworth-Heinemann (2018)

2018
[12]

High-resolution image synthesis with latent diffusion models

Rombach, Robin, Blattmann, Andreas, Lorenz, Dominik, Esser, Patrick and Ommer, Björn. “High-resolution image synthesis with latent diffusion models.”Proceedings of the IEEE/CVF conference on computer vision and pattern recognition: pp. 10684–10695. 2022

2022
[13]

TheCMAevolutionstrategy: acompar- ingreview

Hansen,Nikolaus.“TheCMAevolutionstrategy: acompar- ingreview.”Towardsanewevolutionarycomputation: Ad- vances in the estimation of distribution algorithms(2006): pp. 75–102

2006
[14]

ShipHullGAN:Agenericpara- metric modeller for ship hull design using deep convolu- tional generative model

Khan,Shahroz,Goucher-Lambert,Kosa,Kostas,Konstanti- nosandKaklis,Panagiotis.“ShipHullGAN:Agenericpara- metric modeller for ship hull design using deep convolu- tional generative model.”Computer Methods in Applied Mechanics and EngineeringVol. 411 (2023): p. 116051

2023
[15]

Multi-objective optimization design method of ma- rine propeller based on fluid-structure interaction

Guan, Guan, Zhang, Xiangyu, Wang, Panpan and Yang, Qu. “Multi-objective optimization design method of ma- rine propeller based on fluid-structure interaction.”Ocean EngineeringVol. 252 (2022): p. 111222

2022
[16]

RulesfortheClassificationofSteelShips (NR 467)

Veritas,Bureau. “RulesfortheClassificationofSteelShips (NR 467).”France: Bureau Veritas (BV)(2015). APPENDIX A. EXPERT AND REFERENCE PROPELLER DESIGNS The expert and reference propeller designs used to derive the latent space (Section 3) comprise 30 propeller–vessel sam- ples that were designed by marine propeller experts in response to client requests wi...

2015