DETECTION OF FLOW SEPARATION USING ISING-BASED CLUSTERING

Turbulent flow separation over lifting surfaces is critical for many applications, such as high-lift systems, aircraft noise, wind turbines and turbomachinery flows. Awareness of the occurrence of flow separation could enable active measures to be taken to prevent phenomena associated with flow separation, such as additional noise, decreased lift and structural vibration. In this work, several machine learning methods for the detection of flow separation based on measured surface pressure fluctuations are presented and compared. The models were trained with experimental data covering symmetrical and cambered airfoils, a range of angles of attack from 0° to 23° and Reynolds numbers between Rec=0.8·10^6 and Rec=4.5·10^6. For the supervised training of the machine learning model, the data was labeled according to the status of the flow, attached or separated. This was achieved by evaluating measurement results such as the static pressure distribution along the chord, the lift coefficient and the power spectral density of the surface pressure fluctuations. Three different machine learning techniques were utilized, a multilayer perceptron, a support vector machine, and logistic regression approach. For each of them the hyperparameters were fine-tuned. The support vector machine technique performed best in terms of prediction accuracy. Overall, promising results are obtained: An accuracy of 0.99 and a Matthews correlation coefficient of 0.98 were achieved, significantly outperforming the other machine learning techniques. These results demonstrate the model’s high efficacy in the detection of flow separation. The majority of incorrectly predicted cases are configurations close to the onset of flow separation. Furthermore, the transitions from attached flow to flow separation happens gradually which implies that those cases could be characterized as a mixture of both states, which offers room for further improvement of the machine learning model. The sensitivity of results on database parameters is discussed along with the underlying flow physics and input data quality. First efforts focus on enhancing the model’s real-time capabilities by reducing computational overhead and incorporating multiclass classification to assess varying degrees of flow separation.