Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model

This paper presents a machine learning-assisted adjoint-based shape optimization framework that utilizes a Conditional Variational Autoencoder to surrogate a complex Voith Schneider Propeller, enabling efficient ship hull designs that achieve over an 8% resistance reduction while avoiding the prohibitive computational costs of full transient propulsion simulations.

The Gist

Imagine you are trying to design the perfect shape for a boat to make it glide through water as easily as possible, saving fuel and reducing pollution. This is the job of naval architects.

Usually, they use powerful computer simulations to test thousands of different hull shapes. However, there's a catch: the engine (propeller) changes everything.

The Problem: The "Time-Travel" Nightmare

In this paper, the researchers are dealing with a very special type of propeller called a Voith Schneider Propeller (VSP). Unlike a normal boat propeller that just spins like a fan, this one has blades that spin and tilt up and down simultaneously. It's like a helicopter blade trying to swim.

This creates a chaotic, swirling mess of water that changes every split second. To simulate this accurately, computers have to:

1. Take tiny snapshots of the water flow thousands of times per second.
2. Remember every single snapshot to work backward and figure out how to improve the boat shape.

The Analogy: Imagine trying to solve a puzzle where you have to remember every single frame of a 3-hour movie to figure out how to change the ending. You'd need a supercomputer with a massive hard drive, and it would take forever. For ship designers, this is too expensive and slow to do for every new boat shape they want to test.

The Solution: The "Magic Crystal Ball" (AI Surrogate)

To solve this, the authors built a Machine Learning "Surrogate Model." Think of this as a crystal ball or a highly trained intern.

Instead of simulating the complex, spinning propeller every time they test a new boat shape, they train an AI (specifically a Conditional Variational Autoencoder, or CVAE) on thousands of past simulations.

• The Training: They showed the AI 180 different boat shapes and speeds, along with the resulting water swirls.
• The Magic: Once trained, the AI can look at a new boat shape and instantly guess what the water swirls will look like, without doing the heavy math. It's like the AI has "seen" enough movies to predict the ending without watching the whole thing.

How They Used It: The "Ghost Propeller"

The researchers didn't just let the AI guess; they integrated it into the ship design software.

1. They removed the physical propeller from the computer model (to save time).
2. They told the computer: "Hey, pretend there's a propeller here, and use the AI's prediction of how the water should move."
3. The computer then used a mathematical trick (called Adjoint Optimization) to nudge the boat's hull shape, inch by inch, to find the perfect design that minimizes drag.

The Big Reveal: Ignoring the Engine is a Trap

The most important finding of this paper is a warning: If you design a boat without considering the propeller, you might make it worse.

• The "No-Propeller" Test: When they optimized the boat shape ignoring the engine, the computer said, "Great! We reduced drag by 2%!" But when they actually put the real propeller back in and tested it, the drag increased by 3%. It was like tuning a car engine while ignoring the tires; the car ended up slower.
• The "AI-Propeller" Test: When they used their AI model to include the propeller's effect during the design phase, the computer found a shape that reduced drag by over 8%. When they tested this with the real propeller, it worked exactly as predicted.

The Takeaway

This paper shows that by using a smart AI "crystal ball" to mimic complex engines, engineers can design better ships much faster and cheaper. It proves that you can't just design the body of the car without thinking about the engine; they work together, and if you ignore one, the whole system fails.

In short: They taught a computer to "dream" about how a tricky propeller moves water, used those dreams to design a better boat, and proved that ignoring the engine leads to a slower ship.

Technical Summary

1. Problem Statement

The paper addresses the challenge of optimizing ship hull shapes to minimize hydrodynamic resistance while accounting for complex propulsion systems, specifically the Voith Schneider Propeller (VSP).

• The Challenge: Traditional adjoint-based shape optimization for vessels with VSPs is computationally prohibitive. VSPs generate unsteady, spatially complex flows with strong vortex shedding and blade-vortex interactions.
• Computational Bottlenecks:
  1. Time-Dependence: Resolving the rotating propeller requires transient simulations with small time steps.
  2. Adjoint Backward Integration: Adjoint methods require storing the entire history of the primal flow field to propagate gradients backward in time. For large-scale industrial simulations (millions of grid points, thousands of time steps), this creates insurmountable storage and memory demands.
  3. Inapplicability of Simplifications: Standard approximations like Multiple Reference Frames (MRF) or actuator discs are unsuitable for VSPs due to their complex kinematics (simultaneous rotation and pitching) and the perpendicular alignment of thrust and rotation.
• Consequence: Ignoring the propulsion system in optimization often leads to hull designs that perform worse than the initial shape when the propeller is actually installed.

2. Methodology

The authors propose a Machine Learning (ML)-assisted optimization framework that replaces the geometrically resolved, time-resolved propeller with a Conditional Variational Autoencoder (CVAE) surrogate model.

A. Data Generation (Offline Phase)

• Dataset: 180 transient CFD simulations of a Service Operation Vessel (SOV) equipped with two VSPs.
• Parameters: Simulations varied 10 geometric parameters (stern deadrise, inclination, angles, headbox dimensions) and 3 cruising speeds.
• Processing: Unsteady RANS solutions were time-averaged over one propeller revolution. The unstructured CFD data was mapped onto a structured cylindrical meta-grid surrounding the propeller to create a consistent input for the ML model.

B. The Surrogate Model (CVAE)

• Architecture: A Conditional Variational Autoencoder was employed to learn the mapping between input conditions (geometry + speed) and the resulting time-averaged 3D velocity field.
• Key Components:
  • Conditioning: The model takes 11 "labels" (10 geometric + 1 speed) to parameterize the flow field generation.
  • Encoder/Decoder: Utilizes Convolutional Neural Networks (CNNs) with Residual Blocks and Self-Attention mechanisms to capture local gradients and long-range flow dependencies (e.g., vortex shedding).
  • Loss Function: A composite loss combining Mean Squared Error (MSE) for reconstruction accuracy and Kullback–Leibler (KL) divergence for latent space regularization.
• Output: The model predicts the time-averaged velocity field ($\tilde{\phi}$) on the meta-grid.

C. Integration into CFD (Online Phase)

• Non-Resolved Propeller: The optimization CFD mesh does not resolve the propeller geometry.
• Implicit Forcing: The predicted velocity field from the CVAE is interpolated onto the CFD mesh. Instead of overwriting the solution (which would violate continuity), the authors use an implicit solution forcing approach. A source term is added to the momentum equations:
  $$A_P \phi_P + \sum A_N \phi_N = Q_P + \hat{\alpha} A_P (\tilde{\phi} - \phi)_P$$
  Where $\hat{\alpha}$ is a dimensionless forcing coefficient. This blends the ML-predicted velocity with the CFD solution, ensuring robustness and allowing the solver to satisfy continuity constraints locally.

D. Optimization Algorithm

• Adjoint System: A continuous adjoint solver computes gradients of the total resistance with respect to the hull shape.
• Mesh Deformation: The Steklov–Poincaré method is used to generate admissible mesh deformations while conserving hull volume.
• Strategy: The surrogate model is queried once to generate the initial velocity field. The optimization loop then iterates using the adjoint solver and mesh deformation without re-querying the ML model at every step (though the authors note future work will update the ML input dynamically).

3. Key Contributions

1. Novel Surrogate Framework: First application of a CVAE with self-attention and residual blocks to model the complex, time-averaged flow field of a Voith Schneider Propeller for shape optimization.
2. Implicit Forcing Integration: Development of a robust method to integrate ML-predicted velocities into CFD solvers via implicit source terms, avoiding continuity violations and ensuring numerical stability.
3. Demonstration of Propulsion Importance: Rigorous proof that neglecting propulsion in hull optimization yields detrimental results, whereas including it via a surrogate model yields significant gains.
4. Computational Efficiency: The method bypasses the need for expensive transient adjoint simulations and massive storage requirements associated with resolving the propeller geometry directly.

4. Results

• Surrogate Accuracy: The CVAE achieved high fidelity in reconstructing flow fields. The maximum relative error in velocity magnitude was approximately 1–2% of ship speed, with local peaks around 10% in high-gradient regions.
• Optimization Performance:
  • Case 1 (No Propeller - NP): Optimizing without propulsion influence resulted in a predicted 2% resistance reduction. However, when validated with a fully resolved propeller simulation, the resistance increased by ~2.3%.
  • Case 2 (Surrogate Propeller - SP): Optimizing with the CVAE surrogate resulted in a predicted 8.2% resistance reduction. Validation with a fully resolved propeller simulation confirmed a 9.29% reduction.
• Geometric Changes: The SP optimization created a "tunnel-like" structure behind the headbox to smooth flow transition to the transom, a feature absent in the NP optimization.

5. Significance

This work demonstrates a viable pathway for industrial-scale adjoint shape optimization of complex marine vessels.

• Bridging the Gap: It successfully bridges the gap between high-fidelity physics (CFD) and data-driven efficiency (ML), enabling the optimization of hull-propeller interactions that were previously too computationally expensive.
• Design Reliability: It highlights the critical risk of "false positives" in hull design when propulsion is ignored; a shape optimized for a bare hull can be hydrodynamically inferior when a propeller is added.
• Future Potential: The framework is extensible to two-phase flows (free surface) and self-propulsion conditions, offering a scalable solution for next-generation ship design.

Conclusion: The proposed CVAE-assisted framework reduces computational costs significantly while maintaining the physical accuracy required to optimize hull-propeller interactions, achieving a verified resistance reduction of over 8% compared to the baseline.