CFD-ML: Stream-based active learning of computational fluid dynamics simulations for efficient product design

Computational fluid dynamics (CFD) has been extensively used as a simulation tool for product development in various industrial fields. Engineers sequentially query the CFD simulator to evaluate their design instances, during which they improve the new designs based on previous evaluations. The high cost of performing CFD simulations for numerous design instances is a practical challenge. To reduce this cost, machine learning (ML) approaches have been employed to approximate CFD simulations. Although ML enables the fast approximation of CFD, it can suffer from low accuracy when making predictions for design instances that significantly deviate from the training dataset. In this study, we propose a CFD-ML combined system based on stream-based active learning to utilize the CFD simulator cost-efficiently. The proposed method has two main objectives: reducing the number of CFD simulations and ensuring high accuracy of the ML approximations. When a design instance is queried, the CFD-ML system interchangeably uses the CFD simulator and the ML model depending on the predictive uncertainty of the ML model. If the uncertainty of the ML model is high, the CFD simulator is used to obtain an evaluation result, which is subsequently used to enhance the ML model. Conversely, if the uncertainty is low, the ML model is used to obtain an approximated evaluation result. The CFD-ML system reduces computational costs compared to exclusive reliance on the CFD simulator and yields more accurate evaluations compared to exclusive reliance on the ML model. We demonstrated the effectiveness of the proposed method through a case study on a centrifugal fan development task.