Abstract: High-speed rotating flow fields play a pivotal role in the mechanical systems used for stable heavy isotopes. The high dimensional design space of such complex rotating machinery is characterized by strong nonlinearity and substantial parameter coupling. To efficiently design the system and optimize its performance, it incurs considerable time and economic costs when relying solely on large volumes of high-fidelity experimental data. A common way to reduce the cost is by introducing low-fidelity numerical simulations as complementary information. To address the problem of constructing a mapping from a complex, high-dimensional design space to low-dimensional performance indicators under sparse experimental data, this study proposes a multi-fidelity information fusion framework. Firstly, we employ Latin hypercube sampling to uniformly explore the design space of a high-speed rotating flow system via low-fidelity simulations. Thousands of low-fidelity numerical simulations were performed, and the resulting high-dimensional, dimensionless design parameters were used to develop a Bayesian neural network based on low-fidelity datasets. Building upon this machine learnt model, a Gaussian process regression scheme was incorporated to establish a multi-fidelity surrogate model. Results demonstrate that the proposed surrogate achieves absolute mean percentage deviations within 6% for two primary performance metrics significantly lower than the prediction errors of numerical simulation and well within accepted engineering tolerances. These evidences confirm the effectiveness and applicability of the proposed multi-fidelity modeling framework for high speed rotating flow fields.