Welding has been considered as one of the most efficient and reliable joining technologies in fabricating metallic components. However, imperfections and/or defects in the welded components are inevitable during and after the welding process which have significant effect on weld quality as well as integrity of the finished products. For this reason, the standards for weld quality assessments are issued by various regulatory bodies in which the limits for occurrence or dimensions of imperfections are described. However, in some studies, welded joints in low quality category are found to have higher fatigue strength than those with high weld quality specification. These studies reveal the need to consider the fatigue strength in relation with the standards of imperfections while specifying the quality of a welded component. Fatigue strength assessments are often performed using experimental approach which requires extensive man-hours and cost. On the other hand, numerical methods are also widely used to determine the fatigue life of welded components. However, parameter uncertainties can lead to inaccurate results and considering weld defects takes considerable computational hours. In this regard, multi-fidelity surrogate models are considered as one of the promising approaches for fatigue strength predictions. They are developed and trained using the data constructed from high-fidelity models which are calibrated using experiments, and low fidelity ones which are simplified versions of high-fidelity models. In this study, a multi-fidelity surrogate model which can predict the fatigue life in relation to the weld quality was developed for butt-welded joints containing imperfections. Different from other machine learning models and conventional assessments, the proposed model can provide fatigue strength predictions while keeping the balance between accuracy and computational efficiency by taking advantage of high- and low-fidelity models. The methodology and procedures in developing surrogate model as well as the predictions and findings given by the model are discussed. Furthermore, extending the geometry of the welded joints and types of imperfections will make this surrogate model versatile and robust for practical applications.