익형의 충격파 예측을 위한 인공지능 모델 학습 및 설명가능한 인공지능 적용 (TRAIN OF ARTIFICIAL INTELLIGENCE MODEL AND APPLICATION OF EXPLAINABLE ARTIFICIAL INTELLIGENCE FOR SHOCK WAVE PREDICTION OF AIRFOIL)

서지정보

· 발행기관 : 한국전산유체공학회
· 수록지 정보 : 한국전산유체공학회지 / 28권 / 4호 / 16 ~ 26페이지
· 저자명 : 박현빈, 홍윤표, 강유업, 이상아

초록

Recently, there have been attempts to predict shock waves using high performance neural networks. Those studies have a limitation in that it is difficult to explain the internal operation process by using the deep learning model only as a black box model, making it difficult to understand. In this study, artificial intelligence was used to predict the type of the shock wave from the shape of the airfoil. To this end, a total of 500 airfoils were generated by changing eight variables using the Class/Shape-function Transformation(CST) parametrization method, and flow was solved using KFLOW. Using the analysis results, Multi-Layer Perceptron(MLP) was trained to predict the type of the shock wave from the airfoil geometry. Using eXplainable Artificial Intelligence(XAI) techniques Local Interpretable Model-agnostic Explanation(LIME), SHapley Additive exPlanation(SHAP), and Explainable Boosting Machine(EBM), the prediction results of the model were presented in an interpretable form. The analysis results obtained by applying LIME and SHAP to the MLP model were compared with the analysis results of EBM.. As a result of global and local explanation, it was confirmed that the influence of the variable in charge of the upper middle of the airfoil was the greatest on shock wave prediction.

영어초록

Recently, there have been attempts to predict shock waves using high performance neural networks. Those studies have a limitation in that it is difficult to explain the internal operation process by using the deep learning model only as a black box model, making it difficult to understand. In this study, artificial intelligence was used to predict the type of the shock wave from the shape of the airfoil. To this end, a total of 500 airfoils were generated by changing eight variables using the Class/Shape-function Transformation(CST) parametrization method, and flow was solved using KFLOW. Using the analysis results, Multi-Layer Perceptron(MLP) was trained to predict the type of the shock wave from the airfoil geometry. Using eXplainable Artificial Intelligence(XAI) techniques Local Interpretable Model-agnostic Explanation(LIME), SHapley Additive exPlanation(SHAP), and Explainable Boosting Machine(EBM), the prediction results of the model were presented in an interpretable form. The analysis results obtained by applying LIME and SHAP to the MLP model were compared with the analysis results of EBM.. As a result of global and local explanation, it was confirmed that the influence of the variable in charge of the upper middle of the airfoil was the greatest on shock wave prediction.

참고자료

· 없음