Robust parameter design for rocket boosters using deep Gaussian process

Abstract

In aerospace engineering optimization, multiple correlated non-stationary responses (NSRs) are often produced. Ignoring the non-stationary characteristic of data may impact the prediction accuracy of response surface models and the reliability of optimization solutions. Traditional stationary Gaussian process model struggles to handle non-stationary data with limited samples. Recently, deep Gaussian process (DGP) has gained popularity in non-stationary models due to its ability to handle rapidly changing data. However, efficiently inferring the posterior distribution of DGP remains a significant challenge. This paper uses DGP surrogate model based on Bayesian inference to simulate non-stationary response surfaces and proposes the Chain Rule Hamiltonian Monte Carlo (CRHMC) algorithm to efficiently and accurately sample the posterior distribution. Additionally, a new multivariate quality loss function (MQLF) is introduced to optimize multiple correlated non-stationary responses and ensure the robustness of optimal design parameters. Numerical examples and a real rocket booster engineering case demonstrate the superiority of the proposed method in fast computation and robust optimization against other competing methods.