Data-Efficient Feasible Region Identification for Engineering Design

Design Space Exploration is an important concept in engineering design in which the design space is being explored for suitable design candidates. Most of the time there are design specifications which limit the design space to a particular area of interest called the feasible region. To identify the feasible region(s), time-consuming simulations are run that characterize the design space and its constraints.
A cheaper and faster alternative to the expensive simulations is to use surrogate models which are data-efficient machine learning models. The training samples for the model can be chosen either without using any knowledge of the underlying design problem, but this could lead to many simulations that are of no use to the engineer because they are infeasible; or by using active learning.
Active learning is an iterative process in which the samples are chosen in an intelligent way, so that the required performance of the model is achieved with as few samples as possible. In addition, active learning can focus on simulating designs that are of interest for finding the feasible region.
However, in the case of a design process with many specifications, i.e., a highly constrained problem, the search for the feasible region can become even harder. When there are many constraints, the feasible region can be very small and even scattered across the design space. This makes it sometimes impossible to find the feasible region(s), even with an active learning strategy. Therefore, the aim of this work is to improve current state-of-the-art methods so that they can also be used successfully for feasible region identification in highly constrained design spaces.