The effects of online-training artificial neural network mechanism and multi-stage parametric modeling method on simulation-based design system for ship optimization

To dynamically re-calibrate the weights of efficiency and precision in high-dimensional issues of ship resistance optimization, a hybrid framework containing both Computational Fluid Dynamics (CFD) solver and Artificial Neural Network (ANN) surrogate model, and driven by an online training Self-Adaptive Mechanism is proposed in this investigation. Initially, a classic Simulation-Based Design (SBD) system including an optimizer of particle swarm optimization (PSO) algorithm, a multi-level Free-Form-Deformation modeler and a Star CCM + solver was established; secondly, a fully-connected neural network surrogate model, whose input is a 4108 geometric feature tensor and output is total resistance coefficient CT, was constructed; finally, a self-adaptive mechanism switching between CFD solver and surrogate model by the dynamic error of inspected samples was plugged into the framework. According to the result, the hybrid framework optimized the CT of a scaled 5415 DTBM model with 6.683% at Fr = 0.28. The adaptive mechanism reduced computational cost from 1638 min to 456 min, which improved 72.16% optimization efficiency by activating the surrogate model instead CFD solver while the dynamic errors are below the criteria. The proposed self-adaptive mechanism could dynamically re-balance the efficiency and precision of the framework, and provide theoretical and technical support for ship design optimization.