Computational fluid dynamics (CFD) simulations of aortic hemodynamics provide valuable clinical insights but remain computationally expensive. We present an ensemble deep learning approach for fast prediction of 3D velocity, pressure, and wall shear stress (WSS) fields in patient-specific aortic geometries. Five multi-layer perceptrons predict PCA-reduced fields from B-spline geometric features and inlet velocity, trained on 5,477 converged steady-state RANS simulations with SST k-omega turbulence. On held-out synthetic data, the ensemble achieves MNAE = 1.04%, 1.62%, 1.39% for velocity, pressure and WSS respectively, placing the method among the most competitive reported aortic CFD surrogates. Validation on real patient geometries distinguished 7 cases fully external to the statistical shape model from 26 geometries involved in shape-model fitting. The corresponding MNAE values were 4.6%, 4.6%, and 5.1% for the fully external cases, and 1.5%, 1.7%, and 2.1% for the shape-model-derived cases. Leave-one-velocity-out cross-validation supports generalization to unseen inlet conditions (R^2 >= 0.891, MNAE <= 0.96%). Inference requires 0.1 s per case (more than 3000x speedup over CFD).