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

arXiv:2602.14907 · physics.flu-dyn, cs.LG · Submitted 2026-02-16 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model".

Jane: Adjoint-based shape optimization of ship hulls using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model addresses the computational challenges associated with applying adjoint methods to complex,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So folks, we're diving into this paper today titled "Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model." Essentially, the authors are tackling how to optimize ship hulls while dealing with those super complex and time-dependent propulsion systems that make traditional adjoint methods really tough to use.

Jane: That sounds intense, Tom. The core idea seems to be using a machine learning model, specifically a Conditional Variational Autoencoder or CVAE, to create a surrogate for the propeller's flow field so they can use adjoint methods more effectively for hull resistance minimization.

Lu: Exactly! The paper points out that directly resolving the rotating propeller would force them into unsteady adjoint simulations, which is computationally demanding because you need to store the whole history of the flow field during that reverse temporal propagation. This makes it practically impossible for large-scale applications.

Meng: From an engineering standpoint, that's a huge hurdle; handling those massive time steps and storage requirements just doesn't work with current high-performance computers for real-time optimization. I wonder how practical this surrogate model actually is in terms of its accuracy under different operational speeds.

Lalam: I think the CVAE aspect is really interesting because they enrich the latent space with labels, like ten geometric parameters and one operating parameter capturing cruising speed, to parameterize the fields they are learning. This structured conditioning should make the surrogate much more reliable than a generic neural network.

Tom: Right, so they're using that structured data to build this surrogate model that emulates the time-averaged flow field, which is a significant step away from having to run those huge unsteady simulations every single time. Jane, can you explain what they claim this approach actually achieves in terms of the overall optimization process?

Jane: They show that by using this ML-assisted approach, which involves transferring velocities from the surrogate model onto a meta grid and then interpolating them to their CFD mesh, they can significantly speed up the entire optimization loop. The method allows for a substantial reduction in the required (pseudo) time steps needed to converge on a shape compared to starting from scratch.

Paper summary: Lu: It's about taking the complexity out of the primal problem initially, letting the ML handle that heavy lifting, and then using adjoint methods for the actual sensitivity analysis on the hull itself. This separation of concerns is smart for high-dimensional problems.

Meng: I’m focused on the practical application here. The paper compares a No-Propeller case against a Surrogate-Propeller case, and they found that omitting the propulsion system during optimization can actually be detrimental, showing a decrease in resistance of approximately two percent for the NP run versus about eight point two percent for the SP run. That difference is what really matters for designers on the ground.

Lalam: That result suggests that incorporating the propeller effects via this surrogate model isn't just a minor tweak; it actively contributes to substantial resistance reduction, which is a very tangible improvement for vessel design. It really highlights how these AI tools can move beyond just predicting data and start influencing the actual engineering decisions being made.

Tom: So we've seen that the framework works to get significant resistance reductions, but what are the broader implications for how we approach optimizing complex naval architectures in general? What does this mean for future research in ship design?

Jane: It suggests that machine learning surrogates aren't just useful tools; they become integral parts of the optimization pipeline when dealing with unsteady or time-dependent physical phenomena, like those caused by a VSP. It opens up the door for applying adjoint methods to more complex systems that were previously considered computationally infeasible due to their transient nature.

Lu: I see this as a path toward designing vessels where propulsion interaction is treated as an integrated, learned component rather than something you have to simulate separately and painstakingly reverse-integrate. This could lead to much more holistic design processes for self-propelled ships.

Meng: From a practical standpoint, this means that the computational barrier to high-fidelity shape optimization for vessels is lowered considerably, allowing designers to explore much wider design spaces without getting bogged down in massive simulation times. That translates directly into faster development cycles for new ship designs.

Lalam: If this technology matures, it has the potential to improve the efficiency of global shipping and reduce the environmental impact we discussed earlier by enabling rapid design iterations that incorporate propulsion effects accurately. It’s about making smarter decisions faster.

Paper summary: Tom: That's a powerful summary of where this research is heading—moving complex physics into the realm of efficient, ML-assisted optimization, and I think that's something everyone in the industry needs to pay close attention to as they look toward next-generation ship designs.

Jane: It really does feel like we are moving from simulating everything from scratch to intelligently approximating the hardest parts of the simulation using learned models. That shift in methodology is key here, and it’s what makes this paper so compelling for anyone interested in computational fluid dynamics applications.

Lu: The integration of a CVAE specifically conditioned on operating parameters shows a sophisticated way to map the input conditions—like speed—directly into the learned flow representation, which is quite advanced. This level of control over the surrogate model's behavior is what makes this approach viable for real-world use.

Meng: I just want to make sure we emphasize that while it saves computation time, the accuracy still relies heavily on how well that initial training data represents all possible flow regimes, which is always a factor in AI modeling. It’s a trade-off between speed and perfect fidelity.

Lalam: And I think the cultural impact of this kind of work is that it shows how deeply integrated AI can become in traditionally very physical engineering disciplines like naval architecture, making advanced simulation accessible to a broader range of designers. It broadens what we consider possible for optimization problems.

Tom: Absolutely. So the main point of this paper, "Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model," is that they successfully use an ML model to bypass the computational nightmare of unsteady adjoint simulations when optimizing hulls with complex propellers.

Jane: And what they claim is that this method results in shapes that show more than an eight percent reduction in resistance compared to the initial shape, which is a solid performance metric for any optimization study. That comparison really shows the practical benefit of using this new framework.

Lu: The methodology hinges on training the CVAE on time-averaged unsteady CFD data to emulate the propeller's flow field, and then feeding those ML velocities into a meta grid that connects smoothly to the main CFD mesh for optimization. It’s a very clever way to bridge the gap between the learned model and the actual simulation environment.

Paper summary: Meng: I'm looking at how robust this is when we move beyond just cruising speed, because that operating parameter is crucial for real-world performance in different sea states. We need to know if the learned model generalizes well outside the specific conditions it was trained on.

Lalam: And from a cultural perspective, this kind of work pushes the boundaries of what we think is feasible in traditional fluid dynamics, demonstrating that complex physical systems can be effectively modeled and optimized using modern AI techniques. It shows that collaboration between disciplines like naval architecture and advanced ML research is yielding tangible results.

Tom: So to wrap up this discussion on the "Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model," we've seen how they tackle the computational constraints imposed by complex propulsion systems by using a CVAE as an intelligent surrogate.

Jane: And the main implication is that this provides a pathway to apply adjoint methods to vessel resistance optimization in situations where unsteady simulations are otherwise too expensive or time-consuming for industry use. It’s a practical tool that addresses those high-dimensional design problems we face today.

Lu: The future potential lies in extending this concept to other complex, time-resolved interactions in marine engineering where surrogate models can replace expensive transient physics calculations. Imagine applying this to propulsion system optimization itself rather than just hull shape, which is a natural next step.

Meng: Practically speaking, it means we can start exploring much more ambitious and detailed designs sooner because the computational cost of checking those designs doesn't skyrocket with every minor change anymore. That accessibility to high-fidelity design is a big practical win for the industry.

Lalam: I think the biggest cultural impact is showing that integrating sophisticated AI models into core engineering workflows isn't just theoretical; it’s delivering concrete, measurable improvements in efficiency and design speed. It sets a precedent for how other physical sciences can leverage these tools to solve real-world problems.

Tom: That’s all the time we have for this deep dive into the "Adjoint-based shape optimization of a ship hull using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model," folks. We've covered how they tackle those nasty computational hurdles and what that means for future design work.

Conclusion: Tom: So we've been looking at how this paper tackles the challenge of optimizing ship hulls when they have complex propulsion systems, and now we're getting to the conclusion about that CVAE-assisted adjoint method.

Jane: It really boils down to how they managed to use machine learning to handle those tricky time-resolved flow fields so engineers can actually minimize resistance without needing impossible simulations.

Lu: I think the title itself is very descriptive; it tells us exactly what's happening—using an adjoint method combined with a Conditional Variational Autoencoder for propulsion surrogate modeling.

Meng: From my side, it’s the practical outcome that matters most; they found ways to make those massive computational tasks manageable for real design work.

Lalam: I see this as a major step because it shows how we can use advanced AI to bridge the gap between theoretical fluid dynamics and actual, high-stakes engineering challenges in maritime design.

Tom: Exactly, Jane; they are showing that you can apply powerful adjoint techniques to problems that used to be completely out of reach because of the time-dependency of things like a propeller.

Jane: It's about taking something that requires storing huge amounts of history and instead using a learned model to approximate the time-averaged effect, which simplifies the math immensely.

Lu: The authors did a brilliant job conditioning their CVAE with specific labels for geometric and operating parameters; it’s really clever how they structured the learning process that way.

Meng: For practical application, this means designers can explore many more hull shapes much faster than before because they aren't waiting days for a single simulation run to finish.

Lalam: This kind of progress in integrating generative AI with traditional optimization workflows is setting a new standard for how we approach complex physical systems in design and manufacturing.

Tom: So, if we look at the authors, they clearly had a deep understanding of both the fluid mechanics and the machine learning side needed to put this whole framework together.

Jane: They’ve managed to weave together these very different areas—CFD, adjoint methods, and generative models—into one cohesive optimization strategy.

Lu: It’s really inspiring because it shows that AI isn't just for prediction anymore; it’s becoming a functional component in the optimization pipeline itself.

Meng: I'm interested in how this methodology scales; can we apply this same concept to optimizing the entire propulsion system rather than just its interaction with the hull?

Lalam: That’s a fantastic question, Meng, because that would be the natural next frontier for AI integration in marine engineering research.

Tom: We have to keep thinking about those future possibilities as we look at what this paper actually achieves in terms of reducing design cycles and improving resistance estimates.

Institute for Fluid Dynamics and Ship Theory, Hamburg University of Technology

physics.flu-dyn, cs.LG

Submitted: 2026-02-16

Updated: 2026-09-28

Comments: Accepted for publication in Computers & Fluids. 54 pages, 21 figures

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 73/100

The gist: Adjoint-based shape optimization of ship hulls using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model addresses the computational challenges associated with applying

Key concepts

Adjoint-based Shape Optimization
This is a mathematical technique used to efficiently find the best shape for a ship hull by calculating how small changes in the shape affect the total resistance. It uses 'adjoint' equations, which are essentially reverse sensitivity calculations, to determine the optimal direction for improvement without needing many expensive full simulations.
Conditional Variational Autoencoder (CVAE)
A type of machine learning model trained on time-averaged flow data from a propeller. The CVAE learns to generate or approximate the complex unsteady flow field induced by the VSP based on specific inputs, such as geometric parameters and cruising speed labels. This surrogate model replaces slow, detailed CFD simulations.
Meta Grid and Interpolation
A grid used to bridge the gap between the simplified ML surrogate model and the actual CFD simulation mesh. Velocities from the CVAE are calculated at points on this meta-grid, which are then mapped onto the main optimization mesh using nearest-neighbor interpolation. This ensures that ML flow data is correctly integrated into the larger simulation.
Frozen Turbulence Assumption
A simplification made in deriving the adjoint equations to make them solvable. It assumes that turbulence remains relatively constant during the sensitivity analysis, which helps simplify the complex fluid dynamics equations needed for optimization, allowing for faster computation.

Terminology

Summary

Adjoint-based shape optimization of ship hulls using a Conditional Variational Autoencoder (CVAE) assisted propulsion surrogate model addresses the computational challenges associated with applying adjoint methods to complex, time-resolved propulsion systems by employing a machine learning surrogate model to replicate the time-averaged flow field. This framework enables significant computational savings while maintaining necessary accuracy for optimizing vessel resistance.

The gist: A machine learning-assisted optimization framework employs a Conditional Variational Autoencoder (CVAE)-based surrogate model of the propulsion system to replicate the time-averaged flow field induced by a Voith Schneider Propeller, achieving significant computational savings while maintaining the necessary accuracy of the resolved propeller.

Challenges in Adjoint Shape Optimization

Adjoint-based shape optimization for ship hulls faces significant challenges when vessels employ complex propulsion systems like a Voith Schneider Propeller (VSP). These challenges arise from the need for transient simulations over long periods with small time steps, which makes reverse temporal propagation of primal and adjoint solutions demanding in terms of storage and computing power. Specifically, a direct resolution of the rotating propeller would require unsteady adjoint simulations, which are difficult due to the adjoint backward integration in time and the corresponding need to store the complete history of the primal flow field. Furthermore, there is a critical aspect related to the broad spectrum of relevant time scales, where the time scale of the ship motion is several orders of magnitude larger than the time scale of the flow around the propeller, rendering standard methods like Multiple Reference Frames (MRFs) inapplicable to VSP kinematics.

ML-Based Surrogate Model Development

The proposed solution involves integrating a machine learning (ML)-based surrogate model into a continuous adjoint-based shape optimization process. Instead of geometrically resolving the unsteady interaction, a generative neural network is trained on time-averaged unsteady CFD data to emulate the propeller-induced (time-averaged) flow field. The specific ML approach used is a Conditional Variational Autoencoder (CVAE). This CVAE features a latency space enriched with so-called labels, which are used to parameterize the fields to be learned. In this work, 11 labels—ten geometric parameters and one operating parameter capturing cruising speed—are employed. The training objective balances reconstruction accuracy and latent-space regularization through a composite loss function:

(1) Reconstruction Loss:

MSE = 1/N Σ X i N (x i - xˆ i) squared.

Integration into CFD Simulation

The transfer between the learned data and the optimization study is managed via a meta grid. This meta-grid extends over a circular cylindrical domain around the VSP, and velocities are derived from the ML-based surrogate model at discrete meta-mesh points. These ML velocities are then interpolated to the CFD mesh used for optimization studies using nearest-neighbor interpolation, defined by Equation (7). To incorporate these ML velocities into the momentum equations, an implicit enforcement approach is used via a source term:

AP ϕ P + X N(P) AN ϕ N = QP + α h ρV (ϕ˜ − ϕ) i P.

Adjoint-Based Shape Optimization Strategy

The optimization procedure follows an adjoint system where the objective function to be minimized is the total resistance of the underwater hull, defined by Equation (10). The adjoint field equations are derived assuming a frozen turbulence assumption, leading to:

Rˆp = -∂vˆ i/∂x i = 0 (11)

Rˆv i = ρv j∂v j/∂x i - ρv j∂v i/∂x j - ∂/(∂x j) 2(µeffS ij − pδˆij) = 0 (12)

The physical sensitivity is computed using Equation (14):

s = -∫ΓO µeff ∂vi/∂x j ∂vˆi/∂x k dΓ.

This physical sensitivity is then used to compute an admissible deformation field for every computational node of the mesh, employing the Steklov-Poincaré method to obtain the deformation field. The optimization algorithm utilizes an Armijo backtracking line search within a loop that solves primal and adjoint problems sequentially, resulting in a significant speed-up because the required (pseudo) time steps for convergence of the primal problem are considerably reduced in comparison to the initial shape.

Optimization Results and Validation

The study compares two cases: the No-Propeller (NP) case where propulsion is neglected, and the Surrogate-Propeller (SP) case where a forcing coefficient of αˆ = 0.5 is used. The results demonstrate that omitting the propulsion system during optimization can be detrimental, as the NP run reported a decrease of approximately 2% in resistance, while the SP case achieved a decrease of approximately 8.2%.

Improvements for AI systems

Based on a rigorous review of the provided scientific paper, here are specific improvements that can be made to existing AI systems, and what those improved systems could achieve:

) Improvements for Existing AI Systems & Potential Capabilities

  1. Integration of Physics-Informed Constraints (PINNs) within CVAE Architectures

  2. Enhancement of Conditional Flow Field Generation via Self-Attention Mechanisms

  3. Implementation of Implicit Forcing Schemes for Robust, Continuity-Compatible Surrogate Models

  4. Development of a Closed-Loop, Parameter-Free Adjoint Optimization Framework

) Detailed Capabilities Enabled by These Improvements:

  1. Integration of Physics Constraints (PINNs) within CVAE Architectures

  2. The improved AI system could be trained not just on flow data but also on the governing Navier-Stokes equations (Eqs. 4, 5, 6).

  3. This would allow the surrogate model to generate physically plausible velocity fields even in regions where training data is sparse or nonexistent, significantly reducing reliance on massive datasets for specific flow regimes (like extreme maneuvers or novel hull geometries).

  4. The resulting AI could perform Physics-Guided Design Exploration, allowing engineers to explore design spaces that satisfy fundamental fluid dynamics laws (e.g., mass and momentum conservation) directly during the surrogate model inference, rather than relying solely on post-hoc constraint checking.

  5. Enhancement of Conditional Flow Field Generation via Self-Attention Mechanisms

  6. The CVAE would utilize self-attention layers (as described in Section 2.1) to learn long-range spatial dependencies inherent in complex flows like vortex shedding and flow separation, which are critical for propulsion systems like the Voith Schneider Propeller (VSP).

  7. This enables the AI to generate highly accurate, spatially coherent velocity fields that capture intricate local phenomena (e.g., strong blade-vortex interactions) with minimal computational overhead compared to full CFD simulations.

  8. The improved system could predict localized flow features—such as pressure fluctuations or shear stress maps near the hull surface—with high fidelity, which are essential for accurate sensitivity analysis in shape optimization.

  9. Implementation of Implicit Forcing Schemes for Robust, Continuity-Compatible Surrogate Models

  10. The system would replace explicit velocity overwriting with an implicit forcing approach (Eq. 9), blending the surrogate model's prediction with the CFD solution via a source term that respects momentum conservation principles at each cell.

  11. This ensures that the resulting surrogate model output is not only accurate in terms of flow features but is also mathematically compatible with the underlying CFD solver, preventing numerical instability and divergence during subsequent primal simulations (shape optimization).

  12. The AI could support Implicit-Coupled Optimization, where the surrogate model acts as a stable interface between the fast adjoint process and the slower primal solver, allowing for significantly faster convergence rates in high-dimensional shape optimization problems without requiring prohibitively expensive transient simulations at every iteration.

Abstract

Adjoint-based shape optimization of ship hulls is a powerful tool for addressing high-dimensional design problems in naval architecture, particularly in minimizing the ship resistance. However, its application to vessels that employ complex propulsion systems introduces significant challenges. They arise from the need for transient simulations extending over long periods of time with small time steps and from the reverse temporal propagation of the primal and adjoint solutions. These challenges place considerable demands on the required storage and computing power, which significantly hamper the use of adjoint methods in the industry. To address this issue, we propose a machine learning-assisted optimization framework that employs a Conditional Variational Autoencoder-based surrogate model of the propulsion system. The surrogate model replicates the time-averaged flow field induced by a Voith Schneider Propeller and replaces the geometrically and time-resolved propeller with a data-driven approximation. Primal flow verification examples demonstrate that the surrogate model achieves significant computational savings while maintaining the necessary accuracy of the resolved propeller. Optimization studies demonstrate that neglecting the propulsion system can result in hull designs whose performance is inferior to that of the initial shape when subsequently validated using a numerically resolved propulsor. In contrast, the proposed method produces shapes that actually achieve more than an 8% reduction in resistance.

Sources

Related papers