On the Use of Machine Learning to Classify Laminar Separation Bubbles

Abstract In this work, large-eddy simulations (LES) are used to study a laminar separation bubble (LSB) that develops along the suction surface of a new high-lift, high-work low-pressure turbine blade at low Reynolds numbers. It is shown that the LSB bursts (i.e., transitions from short to long) over a critical range of Reynolds numbers and the effect of bursting on transition and vortex shedding is examined. The results of this work make clear that long LSBs are not just longer versions of short LSBs; they are phenomena unto themselves. Typically, pressure is used to differentiate between short and log LSBs. As a rule of thumb, an LSB is said to be long if it has a “large” effect on the pressure distribution and short if it has a “small” effect on the pressure distribution — “large” and “small” being subjective terms. This work proposes using machine learning to make this determination more objective. A machine learning model is trained to differentiate between long and short LSBs based on pressure. The final model is demonstrated to perform well; not only does the model successfully classify the LSB over a range of Reynolds numbers, but it removes much of the ambiguity in doing so. Differentiating between long and short LSBs can be quite nuanced, especially for well-behaved airfoils at low Reynolds numbers. While the final machine learning model is not expected to be fully general, it is expected that, given sufficient training data, a more general model could be developed using the approach outlined in this work.