Most data in material science are multi-fidelity data. From the viewpoint of data producer, there is a system error for any quantum method. For machine learning algorithm, as a data consumer, various methods have been designed to maximize the number of knowledges extracted from the multi-fidelity data. In this paper, a quantitative method of noise addition was used to evaluate the influence of different noise types and intensities on some multi-fidelity data learning methods. And the effective scope of the data correction method was verified via iterative noise reduction. The results show that the ways to exploit the multi-fidelity data are crucial. It is necessary to consider comprehensively both the size and the noise level of the datasets. On a variety of datasets constructed with different noise types and intensities, the "Onion" training method that gradually deletes lower fidelity data is better than the "one by one" training method in the direction of noise reduction due to the synergistic effect of different multi-fidelity data. No matter what kind of noise intensity and training method, linear noise has less impact on the final performance of model. However, the data with sampled noise added, which the final testing results are similar to the real multi-fidelity data, were recommended to be adopted in a future research. Also, the complex noise in data is difficult to be corrected by a small amount of true data, thus being more suitable for the iterative noise reduction processing.