Scientists are increasingly focused on optimizing hull designs to reduce drag and improve efficiency, a complex task especially when considering advanced propulsion systems. Moloud Arian Maram from JM Voith SE & Co. KG, Georgios Bletsos and Thomas Rung from the Institute of Fluid Dynamics and Ship Theory at the Hamburg University of Technology, Thanh Tung Nguyen and Michael Palm from JM Voith SE & Co. KG, and Ahmed Hassan from the Institute of Fluid Dynamics and Ship Theory collaborated with colleagues to develop a new optimization framework that addresses these challenges. In their study, a conditional variational autoencoder (CVAE)-based surrogate model was introduced to represent the time-averaged flow field generated by a Voith-Schneider propeller, significantly reducing the computational demands previously associated with adjoint-based shape optimization. This collaboration demonstrates that accurate modeling of propulsion systems is critical to effective hull design, reducing drag by more than 8% compared to designs that ignore propeller effects, providing a path to more efficient and sustainable maritime transport.
Scientists are grappling with the enormous computational burden of designing more efficient ships. Optimizing the hull shape of ships with complex propulsion systems has traditionally required prohibitive computational power and storage. New machine learning techniques are expected to accelerate this process and significantly reduce air resistance and fuel consumption.
Scientists have developed a new optimization framework for hull design that leverages machine learning to overcome long-standing computational hurdles. This research addresses a critical need in shipbuilding to minimize vessel drag while accounting for the complex interactions of advanced propulsion systems. This surrogate model exactly replicates the time-averaged flow field produced by a Voith-Schneider propeller, effectively replacing the detailed propeller geometry and its time-dependent behavior with a data-driven approximation.
Initial tests have demonstrated that this approach significantly reduces computational costs while maintaining the accuracy required for reliable hull optimization. The team’s research revealed that ignoring the effects of the propulsion system during hull design can lead to suboptimal results and increased drag compared to the initial hull shape.
In contrast, the proposed method integrating CVAE-based propulsion alternatives achieves more than 8% drag reduction. This advancement is expected to enable more efficient ship designs and contribute to reducing the environmental impact of maritime transport by reducing fuel consumption and associated emissions.
Minimizing hull resistance by optimizing the shape with propeller in mind
Optimization studies utilizing the proposed framework achieved more than 8% reduction in total ship drag compared to the baseline hull geometry. The surrogate model accurately reproduces the time-averaged flow field induced by the Voith-Schneider propeller and replaces the need to geometrically and temporally resolve the propeller itself.
Primal flow validation confirms that this data-driven approximation can achieve significant computational savings while maintaining the required accuracy. Specifically, this study demonstrates that ignoring propulsion system effects during hull optimization can lead to designs that exhibit increased drag when evaluated under realistic propulsion conditions.
The initial geometry, which was optimized without considering propeller effects, showed a decrease in performance when propulsion was reintroduced. Geometries generated using AI-assisted techniques consistently outperformed these, achieving the aforementioned 8% drag reduction. Residual blocks and self-attention mechanisms were incorporated to further improve the accuracy of the surrogate model’s output. Data transfer between the machine learning model and the optimization study is constrained to flow velocity by utilizing a meta-grid that extends beyond the propeller sweep area. This machine learning approach avoids the need for long time-resolved simulations by learning how to reproduce the time-averaged flow field produced by a Voith-Schneider propeller, a type of marine propulsion system with rotating and pitching blades.
Rather than directly simulating the propeller geometry and motion at each time step, the surrogate model provides a data-driven approximation, significantly reducing computational demand. The methodology begins with a detailed simulation of a Voith-Schneider propeller operating within a representative flow environment. The encoder network compresses complex flow data into a low-dimensional latent space, and the decoder reconstructs the flow field from this compressed representation.
To verify the accuracy of the surrogate model, a pristine flow validation study was conducted to compare the flow field predicted by the surrogate model with the flow field obtained from a complete geometrically resolved CFD simulation. This ensured that the data-driven approximation maintained sufficient fidelity to be used in subsequent optimization processes. The advantage of this approach lies in the ability to replace the computationally expensive and time-dependent propeller model with a fast static representation, allowing optimization studies that would otherwise be impractical.
Furthermore, this study intentionally avoids replacing the propeller with body forces or uniform inflow, recognizing that such simplifications may lead to suboptimal hull design. Instead, surrogate models capture the subtle interactions between the propeller and the hull, allowing for a more accurate evaluation of drag and a more effective optimization process. This careful choice of methodology ensures that the resulting hull shape truly benefits from the integration of the propulsion system.
Machine learning streamlines Voith-Schneider propeller modeling to enhance ship drag predictions
The constant pursuit of hydrodynamic efficiency in ship design has long been constrained by computational bottlenecks. For decades, naval architects have relied on painstakingly detailed simulations to refine ship shapes, but the addition of complex propulsion systems essential for realistic performance predictions has pushed those simulations to breaking point.
The challenge is not just one of processing power, but the exponential growth of data and the time required to accurately model transient effects. This research offers a clever workaround, employing machine learning not as a replacement for physics, but as an intelligent proxy. By training a variational autoencoder on the flow field generated by a Voith-Schneider propeller, researchers created a “digital twin” that can mimic the effects of a propeller at a fraction of the computational cost.
This is not at the expense of accuracy. The results show a significant reduction in resistance by more than 8% without compromising fidelity. More importantly, it highlights the danger of oversimplifying the problem by ignoring propellers completely. This is a common practice that can lead to suboptimal designs. The real potential lies in not only optimizing individual hull shapes, but creating a dynamic design space in which countless variations can be rapidly evaluated, paving the way for a new era of bespoke ship designs tailored to specific operational needs and environmental constraints.
👉 More information
🗞 Adjoint-based shape optimization, machine learning-based surrogate models, conditional variational autoencoders (CVAE), Voith-Schneider propulsion (VSP), self-propelled ships, propulsion models, hull optimization
🧠ArXiv: https://arxiv.org/abs/2602.14907
