Abstract: On the basis of the synergy of machine learning and experimental design, an iterative optimization framework for efficiently screening NbTaW-based refractory high-entropy alloys with high yield strength was proposed. With literature and experimental data, a database containing 39 alloys was constructed, and 23 material features related to yield strength were extracted. Feature screening was conducted through Pearson correlation coefficient and maximum information coefficient separately, and the predictive performance of five machine learning models, namely Bayesian regression, kernel ridge regression, nearest neighbor regression, random forest regression and support vector machine, was compared in three rounds of iterative processes. The results show that with the increase of the number of iteration rounds, the prediction accuracy of each model was gradually improved. The kernel ridge regression model after the third round of iteration had the highest prediction accuracy of 87.5%, and this model selected the top 17 features that were highly correlated with the yield strength arranged by the Pearson correlation coefficient. During the first round of iteration, the yield strength of the alloy with a relatively high carbon content predicted by the Bayesian regression model was relatively high, and the maximum relative error of the prediction was 16.5%, verifying the reliability of the model. In the second round of iteration, the maximum relative error of the yield strength of the alloy with added molybdenum predicted by the nearest neighbor regression model was 11.3%, and the generalization ability of the model was improved. The synergy between machine learning and experimental design could significantly shorten the alloy design cycle, providing new ideas for the development of high-performance refractory high-entropy alloys.