Transfer-learning prediction of transition location under cross-Mach number conditions

To predict the boundary-layer transition location over a flat plate across varying Mach numbers, an efficient method was developed for small-sample settings.

Flow-field disturbance datasets across multiple Mach numbers were generated using the NPSE (nonlinear parabolized stability equations), with Ma = 0.01 designated as the source domain and Ma = 0.1, 0.2, 0.4, 0.8, and 1.6 as target domains. The influence of Mach number variations on transition patterns was systematically analyzed. A CNN (convolutional neural network) model was employed to map flow field patterns to transition locations, incorporating a transfer learning strategy with progressive unfreezing and layer-wise learning rates.

Results demonstrate that transfer learning significantly outperforms direct training: for Ma ≤ 0.4, only 1/10 of the target domain samples are required to achieve a mean absolute error below 2.04% of the average ground-truth value; for Ma ≥ 0.8, a progressive domain adaptation strategy controls the error within 6.19%.

The approach enhances transition prediction under small-sample conditions and provides a reliable technical pathway for cross-condition flow modeling.