전이 학습 기반 다중충실도 모델을 이용한 의류 건조기용 다익 원심팬의 최적 설계

This study proposes a deep learning-based multi-fidelity optimization design process to improve the aerodynamic performance of a clothes dryer fan system. The optimization targets are the multi-blade centrifugal fan and scroll, which have a dominant influence on drying performance. The design space was expanded by adding the scroll cutoff angle and operating pressure to the impeller inlet and outlet angles. To minimize data generation costs within this expanded design space, transfer learning was employed. First, flow results obtained through 2D Computational Fluid Dynamics (CFD) simulations were used to pre-train the model as low-fidelity data. Subsequently, transfer learning was performed on datasets containing the new variables. Then, an optimal Deep Neural Network (DNN) surrogate model was constructed using Automated Machine Learning (Auto-ML). The results indicate that the proposed model achieved a 35 % reduction in the required training data while maintaining prediction accuracy comparable to conventional multi-fidelity models. Validation using the derived optimal design parameters confirmed that the flow rate improved by approximately 24 % compared to the baseline model. Furthermore, verification via 3D CFD revealed a prediction error of approximately 1.4 %. In conclusion, this study demonstrates that the proposed design process is effective for the aerodynamic optimization of fan systems while reducing data generation costs.