Hull Form Optimization Based on Multi-fidelity Deep Neural Network

Objectives To improve the optimization efficiency and obtain better optimization results, different fidelity data are organically integrated, and multi-fidelity deep neural network is applied to hull optimization design. Methods A multi-fidelity deep neural network is constructed based on the idea of multi-source data fusion and transfer learning. Firstly, many low- fidelity sample data features evenly distributed in the design space are learned and predicted, and then the linear and nonlinear terms between the high-fidelity data are constructed by fusion learning with a small amount of high-fidelity data to obtain a high-precision surrogate model. Based on this method, the optimization design of the resistance of DTMB5415 ship is carried out. The free form deformation method and the shifting method are used to deform the DTMB5415 bow sonar dome and the whole ship, and a total of 50 sample points are generated. The potential flow and viscous flow are used to evaluate the resistance of the sample points, respectively. The potential flow calculation results are used as low-fidelity data, and the viscous flow calculation results are used as high-fidelity data. Through discussion, it is found that the surrogate model constructed by 50 sets of low-fidelity sample points and 15 sets of high-fidelity sample points have the highest accuracy. The optimal solution is obtained by genetic algorithm and compared with the optimal solution of Kriging model constructed by only 50 sets of high-fidelity data. Results Based on the multi-fidelity deep neural network method, the resistance of DTMB5415 is reduced by 6.73 %. Based on the Kriging model, the resistance of DTMB5415 is reduced by 5.59 %. Conclusions The multi-fidelity deep neural network surrogate model can take into account both efficiency and accuracy, which can be used to optimization. The optimized hull form obtained by it has more significant resistance optimization effect.