Abstract: Surrogate models play a crucial role in aerospace technology. To offer a trade-off between training costs and model accuracy, we introduce multi-fidelity surrogate models based on Gaussian processes. These models leverage data with varying levels of fidelity, to enhance predictive performance while simultaneously minimizing computational expenses. By integrating information from both high-fidelity and low-fidelity datasets, multi-fidelity models can achieve higher accuracy without the need for extensive high-fidelity data, which is often costly and time-consuming to obtain. However, existing multi-fidelity GP models predominantly rely on empirically predetermined kernel functions. These kernel functions are typically selected based on prior experience and general assumptions about the data, without considering the specific characteristics and underlying structures of the datasets at hand. As a result, these kernels are not tailored to the data. To overcome these challenges, this paper introduces a novel approach to constructing kernel functions that are specifically designed to uncover and exploit the latent correlations between datasets of different fidelities. To address these issues, we first propose the Mixture of Experts Neural Kernel (MoENK) which construct a more powerful kernel based on some basic kernels. In MoENK, we use two primary components: MoE-Linear and MoE-Product to selectively mask intermediate results, thereby effectively filtering out noise which may introduce negative transfer. This is then applied to multi-task Gaussian processes, which allows for the sharing of information across different fidelity levels, thereby improving the robustness and accuracy of the surrogate models. To rigorously evaluate the performance of the proposed method, we conducted a series of experiments using three benchmark function examples and two airfoil cases. It exhibited particularly notable advantages in predicting the high-dimensional drag coefficient scenario of the NACA0012 airfoil, reducing RMSE and MAE by 40.42% and 44.70%, respectively, compared to the suboptimal method, LR-MFS. These results confirm that the proposed MoENK kernel function can adaptively predict without relying on predefined kernels, offering strong generalization capabilities and robustness. This provides a new tool for constructing surrogate models in engineering systems.