Fusion of ML with numerical simulation for optimized propeller design

In computer-aided engineering design, the goal is to find an optimal design solution for given specific requirements using a numerical simulator and an optimization method. This study focuses on design problems with low computational cost but high-dimensional design spaces, where traditional forward surrogate-based optimization that reduces computational cost of forward evaluation is ineffective and doesn't improve the design search challenge that is presented by the high dimensionality of problem. We propose to solve it using machine learning models as surrogates that tries to model the inverse mapping using historical collected data and it can be queried to produce good set of designs that can be used as initial seed design for the next stage of optimization using traditional method. This potentially reduces the time to find the optimal design . We call this approach, as Surrogate Assisted Optimization (SAO) , combines machine learning surrogates with conventional optimization methods. Empirical evaluations on propeller design problems show that SAO can achieve better designs on a given budget of evaluation in comparison to traditional optimization method.