Constrained efficient global multidisciplinary design optimization using adaptive disciplinary surrogate enrichment

Surrogate-based optimization has become a popular approach for solving problems with computationally expensive disciplinary solvers. Recently, the Efficient Global Multidisciplinary Design Optimization (EGMDO) algorithm was introduced as a surrogate-based algorithm for multidisciplinary optimization, where the disciplinary solvers are replaced by Gaussian process (GP) surrogate models. A dedicated model of the resulting random objective function is then used to perform Bayesian optimization and focus the computational effort in promising areas of the design space. While its results are promising, the original EGMDO formulation does not provide a strategy for constraint handling. In this work we reformulate the original EGMDO algorithm to be able to handle both equality and inequality constraints. To do so, surrogate models of the constraint functions are first obtained. Then, the original uncertainty reduction strategy is adapted to account for the uncertainty introduced by the disciplinary GPs in both objective and constraint functions. The performance of the resulting algorithm, called Constrained-EGMDO (C-EGMDO), is then illustrated on a benchmark analytical MDO problem and on an engineering test case.