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

This paper proposes a modular AI framework that utilizes a physics-based dataset of 20,000 propeller geometries to enable rapid, direct performance-to-design generation and optimization through conditional generative models and neural-network surrogates.

The Gist

Imagine you are a chef trying to create the perfect recipe for a new type of cake. Traditionally, you would spend months trial-and-erroring: you bake a cake, taste it, realize it’s too dry, change the flour, bake it again, realize it’s too sweet, change the sugar, and repeat. This is exactly how marine engineers design propellers for ships—it’s a slow, expensive "design spiral" of building a shape, running a massive computer simulation to see how it works in water, and tweaking it over and over.

This paper introduces an AI-powered "Magic Recipe Generator" for propellers. Instead of starting with a shape and seeing what happens, engineers can now start with the result they want and have the AI "dream up" the shape instantly.

Here is how their three-step system works, explained through the analogy of a Professional Architect and a 3D Printer:

1. The "Super-Library" (Data Generation)

Before the AI could learn, it needed a massive textbook. The researchers didn't just use old books; they built a "digital laboratory." They used physics software to "bake" over 20,000 different propeller shapes and recorded exactly how each one performed in the water.

• The Analogy: It’s like giving a student architect 20,000 blueprints of every possible building ever made, along with notes on how much wind each building can withstand.

2. The "Dreamer" and the "Critic" (The AI Framework)

The heart of the paper is a two-part AI system:

• The Dreamer (Generative Models): This is the part that takes a request—like "I need a propeller that is 2 meters wide and provides massive thrust for a heavy cargo ship"—and instantly sketches out several possible shapes.
  • The researchers tested two types of "Dreamers." One (called cVAE) is like a Conservative Artist: it stays very close to what it has seen before, producing safe, reliable, but similar designs. The other (called Latent Diffusion) is like an Avant-Garde Artist: it explores wilder, more creative shapes that might work in ways humans haven't thought of yet.
• The Critic (Surrogate Model): Because the "Dreamer" might occasionally sketch something impossible (like a propeller with a hole in the middle), they built a "Critic." This is a lightning-fast AI that looks at the sketch and says, "In a millisecond, I can tell you that shape will be 85% efficient." It replaces the slow, days-long simulations with an instant "thumbs up" or "thumbs down."

3. The "Polisher" (Evolutionary Optimization)

Even a great sketch needs fine-tuning. The final stage takes the AI's best "dreams" and runs a quick optimization process to make sure they meet strict real-world rules—like making sure the blades aren't too thin to break or too wide to fit in the ship's engine.

• The Analogy: It’s like taking a beautiful architectural sketch and running it through a computer to ensure the plumbing works and the walls are thick enough to hold up the roof.

Why does this matter?

In the old way, designing a propeller could take months. With this AI framework, engineers can explore thousands of different "what if" scenarios in a fraction of the time.

The Bottom Line: This paper moves propeller design from a slow process of "Guess, Test, Repeat" to a fast process of "Tell the AI what you want, and let it show you the possibilities."

Technical Summary

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

1. Problem Statement

Traditional marine propeller design follows a "design spiral"—an iterative, manual process of geometric modification, high-fidelity simulation (e.g., CFD), and expert refinement. This cycle is computationally expensive and time-consuming, often taking months. While AI has successfully revolutionized 2D airfoil design (inverse design), applying it to 3D marine propellers is hindered by:

• Data Scarcity: A lack of large, standardized, high-fidelity datasets.
• Complexity: The high-dimensional nature of 3D geometries and the non-linear relationship between geometry and hydrodynamic performance.
• Methodological Gap: Most existing marine AI research focuses on forward modeling (predicting performance from geometry) rather than inverse modeling (generating geometry from performance targets).

2. Methodology

The authors propose a modular, three-stage AI-driven framework that shifts the paradigm from "geometry $\rightarrow$ performance $\rightarrow$ optimization" to a direct "performance $\rightarrow$ design" approach.

A. Physics-Based Data Generation
To solve the data scarcity problem, the authors built a synthetic database of over 20,000 four- and five-bladed propeller geometries.

• Representation: Geometries are parameterized using a 162-dimensional vector consisting of six radial distributions (chord length, skew, max thickness, rake, pitch, and max camber) sampled at 27 positions, plus global parameters (diameter and blade count).
• Synthesis: They used Principal Component Analysis (PCA) on expert designs to create a latent space, from which new geometries were sampled and evaluated using the PropElements hydrodynamic solver to generate performance curves ($K_T, K_Q, \eta$).

B. The Three-Module Framework

1. Conditional Generative Models (Inverse Design): Two architectures were developed to map performance specifications ($J, K_T, \eta, D, B$) to candidate geometries:
   • Conditional Variational Autoencoder (cVAE): Learns a latent distribution to reconstruct designs based on conditions.
   • Latent Diffusion Model (LDM): A two-stage approach where a VAE first compresses the geometry into a latent space, and a diffusion model iteratively denoises latent codes conditioned on performance targets.
2. Performance Prediction Surrogate Model: A multilayer perceptron (MLP) neural network trained to predict $K_T, K_Q,$ and $\eta$ in milliseconds, acting as a high-speed evaluator for the generated designs.
3. Design Refinement (Evolutionary Optimization): To ensure engineering viability, the framework uses the CMA-ES (Covariance-Matrix Adaptation Evolution Strategy) algorithm. This stage refines the AI-generated "concepts" to satisfy hard constraints like blade-area ratio (BAR), minimum thickness, and power limits.

3. Key Contributions

• New Dataset: A reusable, large-scale physics-based dataset of 20,000+ propellers with full open-water performance curves.
• End-to-End Inverse Design: A framework that bypasses the traditional iterative loop by generating geometries directly from performance requirements.
• Hybrid Optimization: Coupling generative AI with evolutionary algorithms to bridge the gap between "plausible AI concepts" and "manufacturable engineering designs."

4. Results and Discussion

• Surrogate Accuracy: The MLP surrogate model demonstrated extremely high accuracy, with $R^2$ values exceeding 0.99 for $K_T$ and $K_Q$.
• cVAE vs. LDM:
  • cVAE: Provided tighter control and lower error in matching specific performance targets but produced more conservative, uniform designs (interpolation).
  • LDM: Produced significantly more diverse designs (higher Spread Coefficient) and explored regions beyond the expert design manifold (extrapolation) while maintaining hydrodynamic plausibility.
• Robustness & Limitations: The models performed best in high-efficiency regimes but showed higher errors in extreme, low-efficiency, or high-load regions due to data sparsity. The VAE component showed a tendency to "pull" out-of-distribution (OOD) designs back toward the training manifold, indicating a limit to the model's ability to generalize to entirely new design philosophies.

5. Significance

This work represents a significant step toward the automation of naval architecture. By replacing slow CFD-driven iterations with millisecond-scale AI predictions and generative design, the framework drastically reduces the design cycle time. It provides a roadmap for using generative AI not just as a decision-support tool, but as a primary engine for exploring new, high-performance propeller geometries that human designers might not intuitively conceive.