Rapid optimization of rocket projectile tailfins based on multi-fidelity neural networks with transfer learning

Abstract

The aerodynamic characteristics of a rocket projectile have a direct impact on its flight performance, with tailfins playing a crucial role. Optimizing the tailfin shape is therefore a key focus in rocket projectile design. However, traditional optimization methods based on computational fluid dynamics (CFD) are time-consuming and inefficient. To address this, the study introduces a multi-fidelity fully connected neural network model based on transfer learning (MFFCN-TL) and applies it to the aerodynamic optimization of a 122 mm rocket projectile. The MFFCN-TL model was pre-trained on 1100 low-fidelity CFD samples and fine-tuned with 80 high-fidelity samples to predict the lift coefficient, drag coefficient, and pressure center, with Monte Carlo Dropout used to quantify uncertainty. The prediction performance of MFFCN-TL was first evaluated against several surrogate models, followed by single-objective and multi-objective optimizations of the tailfin design. In single-objective optimization, the goal was to maximize the lift-to-drag ratio by enhancing Bayesian optimization with a heteroscedastic Gaussian process and an improved upper confidence bound acquisition function, enabling the use of model uncertainty to guide the optimization process and balance exploration and precision. Multi-objective optimization employed multi-objective Bayesian optimization (MOBO) and the non-dominated sorting genetic algorithm II (NSGA-II) to balance the lift-to-drag ratio and drag coefficient, offering comprehensive design trade-offs. The results demonstrate that MFFCN-TL achieved determination coefficients of 0.9224, 0.9912, and 0.9688 for the drag coefficient, lift coefficient, and pressure center, respectively, surpassing other models. This method increased the lift-to-drag ratio by 37%, delivering optimization results comparable to CFD-based methods, while shortening the optimization time to 17.5 days from 26.5 days required by CFD. Multi-objective optimization produced a Pareto front and were validated by CFD simulations, confirming the reliability of the model. This approach provides an efficient and reliable solution for rapid tailfin design, significantly reducing computational cost while maintaining high accuracy.