Multi-Fidelity Surrogate Models for VPP Aerodynamic Input Data

Abstract

Predicting the performance of a sail design is important for optimising the performance of a yacht, and Velocity Prediction Programs (VPPs) are commonly used for this purpose. The aerodynamic force data for a VPP is often calculated using Computational Fluid Dynamics (CFD) models, but these can be computationally expensive. A full VPP analysis for sail design is therefore usually restricted to high-budget design projects or research activities and is not practical for many industry projects. This work presents a method to reduce the computational cost of creating lift and drag force coefficient curves for input into a VPP using both multi-fidelity kriging surrogate modelling and data from existing sail designs. This method is shown to reduce the number of CFD simulations required for a desired accuracy when compared to a single-fidelity model. A maximum reduction in the required computational effort of 57% was achieved for model-scale symmetric spinnaker sails. For the same number of simulations, the accuracy of the model predictions was improved by up to 72% for scale-symmetric spinnaker sails, and 90% for asymmetric spinnakers.

Selection of Existing Sail Designs for Multi-Fidelity Surrogate Models

“…These can be used to improve the accuracy of the surrogate model, or to allow the use of fewer high-fidelity points for the same model accuracy. Further details on kriging and multi-fidelity kriging can be found in Forrester et al (2007; and Peart et al (2021).…”

“…This results in fewer high-fidelity points being needed. The difference in fidelity can be as a result of having different mesh resolutions (Serani et al, 2019), different solvers (Pellegrini et al, 2016), different models of the physics (Sacher et al, 2021), a combination of experimental and numerical methods (Kuya et al, 2012), or using data points from previous simulations of similar designs (Peart et al, 2021). Multifidelity methods have the potential to further reduce the computational cost of evaluating a function, or improve the accuracy for the same computational cost, when compared to conventional surrogate modelling methods (de Baar and Roberts, 2015;Peart et al, 2021).…”

“…The overall goal of this work is to construct accurate multi-fidelity kriging surrogate models to predict the force coefficient curves of new sail designs over a range of Apparent Wind Angles (AWAs), which could be used as the aerodynamic input to a VPP. The multi-fidelity model methodology used is as presented in Peart et al (2021). The high-fidelity data comes from simulations of the performance of a new sail design, which is the sail for which the lift and drag coefficient curves are desired for performance analysis.…”

“…It is therefore important that the trends of the low-fidelity and high-fidelity curves are similar, so that the final multi-fidelity curve matches the true curve as closely as possible. When these trends are not similar, as demonstrated by Peart et al (2021) for symmetric spinnaker sails, using a multi-fidelity model can have an increased error compared to a single-fidelity model trained on just the high-fidelity data points.…”

Combining Analytical Models and Mesh Morphing Based Optimization Techniques for the Design of Flying Multihulls Appendages

“…The use of numerical data can be expensive due to the cost of computations. In (Peart, et al, 2021) several approaches to reduce the calculation requirements are suggested. In (Cella, et al, 2016) it is possible to find an example on how sail trim optimization procedures and analytical VPPs can be coupled.…”