A Conditioned Hybrid Transformer Surrogate for Propeller Surface Flow Field Prediction

This study develops a conditioned hybrid transformer surrogate to rapidly predict the surface flow fields of marine propellers. The model adopts a CNN-transformer-CNN backbone with a four-layer transformer bottleneck. The CNN encoder and decoder are uniformly conditioned via adaptive instance normalization (AdaIN), and three conditioning paradigms—token-prepend, adaptive layer normalization with zero-initialized gating (AdaLN-Zero), and cross-attention—are systematically compared within the transformer bottleneck. The discrete blade number and the continuous advance coefficient are jointly embedded as conditioning inputs. Model training and evaluation are conducted on a dataset of 2710 operating cases covering 560 propeller samples with three-, four-, and five-bladed configurations. The three conditioning paradigms achieve comparable accuracy, with the AdaLN-Zero scheme yielding the lowest validation loss; ablation results indicate that the CNN-level AdaIN conditioning, rather than the bottleneck paradigm, dominates the conditioning effect. On the test set, the proposed model achieves relative errors within 10% for flow field distributions and mean absolute percentage errors within 3.1% for the derived thrust and torque coefficients. The minimum pressure coefficient and the low-pressure region are also accurately captured. The proposed model achieves a speedup of two to three orders of magnitude over CFD simulations, and thus serves as a promising tool for fast hydrodynamic assessment in propeller optimization iterations.