White papers

Deep Learning Physics for the Hydrodynamics of Trading Vessels



Estimated time of reading

Jonas Verrière
November 29, 2022
Jonas is a former simulation expert, holding a PhD in Computational Fluid Dynamics, with 10years + experience in the automotive and aerospace industry, working several years as an aerodynamicist and CFD engineer at Airbus, and later as industry process consultant in the simulation software industry. He joined Extrality as a technical consultant for presales and customer success.
In this study, we are using Deep Learning Physics (DLP) to build a generic AI-model for hydrodynamic studies of tankers.‍ First we will introduce the Deep Learning Physics technology, then we will describe the dataset used to build a tanker boat DLP-model and finally assess this model vs CFD.


Efficiency of ship hulls is key for all trade boats like, bulk carriers, tankers, container vessels…. and rising environmental concerns and regulations, it is even more critical and requires an accurate estimation of the hull resistance always earlier in the design process to avoid late stage re-design or better optimized hulls. This resistance assessment usually starts with empirical models like the Holtrop (Holtrop, 1984) method to dimension the hull parameters during the initial design. Then later in the design cycle the shape optimization is carried out using Computational Fluid Dynamics (CFD) solvers to confirm and refine the resistance of the hull before proceeding to towing tank tests. While empirical models are extremely fast but not very accurate, the main benefits of the CFD solvers are their good accuracy, the large insights helping the designer provided with the 3D high-fidelity volume and surface fields. Their drawback lies in the complexity of enforcing CFD computations and their long turnaround time that prevents their extensive usage in the early design stages.

The ambition of the Deep Learning Physics developed by Extrality is to offer AI-models exhibiting the accuracy and level of insights of CFD solvers but with the simplicity and speed allowing real Simulation-Driven Design.

Deep Learning Physics

Extrality has developed the Deep Learning Physics technology (4 patents) at the forefront of state-of-the-art geometric deep learning research, leveraging on past data to deliver fast 3D high-fidelity predictions. The Extrality technology offers a unique multiscale implicit graph neural network architecture combined with physical priors to deliver AI-models for any type of physics. The Extrality platform enables anyone to create their own AI-models, to store them in a library and to run the suited AI-model to get high-fidelity predictions for a specific workflow. This platform is accessible through both a web application allowing agile work and a Software Development Kit (SDK) allowing process automation.

Dataset description

The total dataset consists of 288 CFD simulation samples with parametric and non-parametric variations of the hull and 1 parameter driving the operating condition, the speed of advance of the boat. The size of this dataset is significant and allows to cover a large variability of hull designs and operating conditions, as well as large amplitude of boat sizes.  It must be kept in mind though, that such a large dataset is not mandatory to take advantage of Deep Learning Physics, and a few dozens of samples can already lead to valuable DLP models.

The simulations were organized into three sets:

- The train set is used to train the model and contains 244 simulations. These data are seen during the training.

- The validation set is used to tune the model and contains 20 simulations. These simulations are not seen during the training but as they are used to finetune meta parameters they can bias our performance estimation. These data are often included in the train set.

- The test set is used to test the trained model. It is used to give an unbiased estimation of the model performances on new cases. It contains 20 simulations.

In order to give more insights about the true boundary of the design space for high dimensional input models (design features are learnt from million degrees of freedom of sample surface meshes), here is a table summarizing the extrema of some global features in the table below.

The corresponding Froude Number is varying from 0.13 to 0.21

This large database is then used to train an AI-model with Deep Learning Physics on the Extrality platform. 

DLP-model performances

20 simulations of the dataset are kept to assess the AI-model performances on some unseen configurations. This evaluation is made by analyzing time metrics and comparing the AI-predictions of the test set with the corresponding CFD results. 

Turn-around time 

One big advantage is the very minor pre-processing time needed to run a prediction with the model, since no specific mesh refinement or solver setup is required. A simple tesselation of your boat in a STL format is to be uploaded on the platform before starting the prediction. This can represent from minutes to hours of savings depending on your process automation and design complexity.

Each prediction launched with this model is equivalent in quality to a CFD performed over 128 CPUs during 2 hours. A prediction on a new geometry with the Extrality platform takes around 20 seconds, including the upload phase.

Global coefficients

The first comparison is carried out on the components of forces integrated on the hull surface, showing a mean relative error below 5% on the resistance force component (ForceX) and below 0.2% on the other force components. 

Surface results

As we can see the Extrality model is capable of correctly predicting the surface forces. We also show below an example of surface comparison on one of the configurations in the testset, never seen by the model during the training. 

Qualitative comparison of the iso-static pressure surface distribution
Qualitative comparison of the viscous stress surface distribution
Qualitative comparison of the volume fraction water surface distribution

This qualitative assessment is showing a very good representation of the pressure and wall-shear stress distribution as well as the free surface line on the hull. This can be quantitatively verified by computing the absolute error on each cell of the hull mesh. In this case the mean absolute error is lower than 15N.m-2.

The DLP-model also provides access to the full volume and some comparisons can be made within different slices, as below in an horizontal plane, showing as well a good agreement between CFD and Extrality’s model.

Slice with normal=z, origin=[0, 0, 0] and colored by Vz

Resistance force evolution along the hull

As a consequence of the correct surface distribution, the global resistance curve evolution on the advancement axis is showing a very good match with the CFD one.

Quantitative assessment of the resistance force evolution along the hull. Blue: Extrality. Orange:CFD

DLP-predictions on the Extrality platform

This tanker model is then available in a dedicated workspace of the Extrality platform and extremely easy to use to assess the hydrodynamic performance of new hull designs. You simply have to upload the STL file of your hull, select the advancement speed you are interested in and get your resistance curve in less than a minute. 

In case you are requested to study the hull efficiency at a new operating condition, add a new speed and benchmark all the existing designs in one click as presented in this video. This makes it very straightforward and fast to carry out hull optimization. 

This makes it very straightforward and fast to carry out hull optimization, as presented in this study.


Holtrop, J., 1984. A Statistical Re-Analysis of Resistance and Propulsion Data. International Shipbuilding Progress 28 (363), pp. 272-276, Delft University Press.