Further advancing key technologies requires the optimization of increasingly
complex systems with strongly interacting parameters—like efficiency
optimization in engine development for optimizing the use of energy. Systematic
optimization approaches based on metamodels, so-called Metamodel-Based Design
Optimization (MBDO), present one key solution to these demanding problems.
Recent advanced methods either focus on Single-Objective Optimization (SoO) on
local metamodels with online adaptivity or Multiobjective Optimization (MoO) on
global metamodels with only limited adaptivity. In the scope of this work, a
fully online adaptive (“in the loop”) optimization approach, capable of both SoO
and MoO, is developed which automatically approximates the global system
response and determines the (Pareto) optimum. A combination of a new Design of
Experiment (DoE) method for sampling points, Neural Networks as
metamodel/Response Surface Model (RSM), and a Genetic Algorithm (GA) for global
optimization performed on the RSM enables very high flexibility. Key features of
the presented MBDO methodology are as follows: A new fully online, adaptive
approach working in iterative loops combined with successive refinements of the
RSM; Two novel boundary treatment approaches for handling arbitrarily complex
constraints; A novel approach to automatically adapt the number of neurons of
the Neural Network to the system complexity; An innovative uncertainty-based DoE
method to maximize information gain for each new sample point; Comprehensive
additional sampling strategies. Detailed benchmarks to popular DoE methods and
MBDO approaches from the literature are conducted. The benchmarks show
comparable to slightly better performance to current state-of-the-art SoO MBDO
approaches with the significant benefit that a global RSM is obtained, providing
valuable insight into the system behavior. Compared to state-of-the-art MoO MBDO
approaches, the benchmark highlights a considerably better performance in terms
of the needed number of samples (i.e., simulations or experiments),
significantly fewer resources required, and high accuracy approximation of the
Pareto front.