<p>In high-performance computing and complex engineering system modeling, high-fidelity (HF) numerical simulations offer accurate results but are often prohibitively expensive for large-scale or iterative tasks. Multi-fidelity surrogate models (MFSMs) address this challenge by combining limited HF data with abundant low-fidelity (LF) data to balance accuracy and computational cost. However, many existing MFSMs may struggle with fine-scale variations or abrupt changes, and some employ complex structures that make it difficult to interpret variable contributions and interactions. To address these potential limitations, this paper proposes CoMARS, a novel MFSM framework based on multivariate adaptive regression splines (MARS). By leveraging MARS’s adaptive piecewise regression mechanism, CoMARS is expected to provide improved flexibility in representing local features while integrating HF and LF data. Its transparent additive form provides a clear model structure, which is beneficial for the interpretation of variable contributions and interactions. Extensive experiments on 26 benchmark functions demonstrate its high prediction accuracy and robustness, and the results (e.g., NME) indicate its effectiveness in controlling local prediction errors. Furthermore, its application to the optimization of the upper beam in a press machine validates its practical engineering value and adaptability to complex design scenarios.</p>