Data-driven modeling has been widely employed to alleviate the computational burden in simulation-based design optimization. Once trained, such models can be evaluated at negligible computational cost, making them particularly suitable for repeated tasks such as sensitivity analysis and optimization. Nevertheless, despite continuous advances in computing resources, acquiring a sufficiently large dataset remains challenging when individual simulation runs are computationally expensive. To address this limitation, multi-fidelity (MF) surrogate modeling has emerged as an effective paradigm that integrates high-fidelity data, which are highly accurate but costly, with low-fidelity data, which are less accurate yet inexpensive to obtain [1]. By combining information across different fidelity levels, MF modeling can exploit their complementary characteristics and often achieve improved performance compared to single-fidelity approaches [2,3]. Despite this potential, many existing studies do not fully exploit cross-fidelity information during hyperparameter optimization, leaving substantial room for improvement in terms of both accuracy and robustness [2,3]. To bridge this gap, this study proposes a hyperparameter optimization framework for MF surrogate modeling that explicitly incorporates physical information. Rather than selecting hyperparameters solely based on statistical goodness-of-fit, the proposed approach integrates physics-informed cues—such as expected trends and admissible ranges derived from low-fidelity models—into the hyperparameter search process, thereby enhancing the effectiveness of MF data fusion. The framework is evaluated using a set of benchmark functions and a real-world engineering problem. The numerical results demonstrate that, under the same computational budget, the proposed method consistently achieves higher predictive accuracy than conventional hyperparameter optimization strategies.