A Distributed Method for Solving Large-Scale Multidisciplinary Optimization Problems

In this paper, we propose an iterative and distributed method for solving multidisciplinary design optimization problems using a block-coordinate-descent algorithm. The method is motivated by scenarios in which the monolithic formulation of the optimization problem either fails to reliably converge, or cannot be used at all due to implementation challenges. In these situations, we decompose the monolithic problem into multiple optimization sub-problems, each smaller and easier to solve than the original. Our work has three major contributions. First, we introduce a novel distributed architecture for multidisciplinary optimization problems that is solved using block coordinate descent. Second, we present a theoretical result that guarantees global convergence to a stationary point given an unconstrained optimization problem. Third, we demonstrate this theoretical result by solving distributed versions of a control co-design problem and two aero-structural optimization problems. In each of these examples we recover the monolithic solution, and, for the aero-structural problems, we show that distributed warm-starting improves the reliability of convergence from several random initializations.