Naval architects design hull shapes from precedent and hydrodynamics, validated against experiment, and accuracy is not optional: small geometry changes alter a vessel's hydrodynamic performance drastically. Evaluating resistance, sea-keeping and green water pushes engineers towards simulation-driven design, where significant bottlenecks still remain today.
Key takeaways
Marine engineering blends traditional design practice with modern simulation and experimental validation.
Slight geometry modifications can drastically change overall hydrodynamic performance, so precision matters.
Resistance, sea-keeping and green water are the performance indices that drive hull evaluation.
Simulation-driven design is where the field is converging, though bottlenecks remain in the workflow.
Marine engineering is a complex blend of traditional design best practices with modern knowledge. Naval architects can nowadays carefully design hull shapes based on previous examples and the science of hydrodynamics, leveraging modern simulation software, combined and validated with experimental results. Accuracy is not an option for the naval engineer: slight modifications in the geometry can have a drastic impact on the overall hydrodynamic performance of the vessel!
To evaluate the vessel's hydrodynamic resistance and various performance indices such as sea-keeping and green water, engineers are converging towards a simulation-driven design approach, but some major bottlenecks remain. CFD simulations are time-consuming and expensive, and the design space exploration remains limited to just a few parameters.
Neural Concept, leveraging Geometric Deep Learning models, allows to overcome the above limitations and to radically improve the design process of organizations, also in the marine sector.
In this use case, we show how one of our partners, CAD-IT Korea, within the first month of using the Neural Concept platform, could build a model which predicts the hydrodynamic performances of boat hulls, directly from the CAD geometry, within a few ms.
The model was trained using 150 different geometries, directly from the CAD format, together with the corresponding STAR-CCM+ simulations. These hull geometries correspond to typical variations that can be expected for a standard tanker ship. The aim of the training process was to learn how to predict the pressure distribution on the hull, to prevent any undesired flow behavior; also, Deep Learning process involved the prediction of overall performance parameters of the hull: the shear force coefficient, and the pressure force coefficient.
After the training phase, the Deep Learning model was put in production. The results on new, unseen shapes, are given below:
Table 1: Accuracy of the Neural Concept model on 37 unseen geometries for scalar predictions, comparison with STAR-CCM+ simulationsFigure 1: Comparison of Neural Concept prediction with a STAR-CCM+ simulation for the pressure distribution on a new geometryFigure 2: Comparison of predictions with a STAR-CCM+ simulation for the pressure distribution on a new geometry
