Machine learning can acquire knowledge efficiently and has become a research hotspot in the field of artificial intelligence. Among them, the machine learning algorithm based on the XGBoost model has many advantages such as high computational efficiency, strong generalization, effective handling of missing values, and strong prediction performance, which shows excellent performance and wide prospect in solving the problems of regression, classification, ranking and industrial process prediction. To address the current problems of large sample size and limited model extrapolation capability in the establishment of cigarette tar amount prediction models, this paper proposed a XGBoost model based algorithm that can realize the inverse prediction of the combination of auxiliary material designing parameters under given tar amount, and the XGBoost models were trained based on the data set accumulated during the process of cigarette production. The results demonstrated that: (1) Four XGBoost based models were cross-trained with the XGBoost algorithm for four features: tar amount, tipping paper permeability, filter rod pressure drop, and cigarette paper permeability, each model was generated with one of the features as the objective function and the other three features as input parameters; (2) Based on the trained XGBoost models, a machine learning inverse prediction algorithm for given cigarette tar value is proposed, which can predict the cigarette design parameters such as tipping paper permeability, filter rod pressure drop, and cigarette paper permeability under a given tar amount. (3) With the predicted cigarette design parameters, the accuracy between the predicted tar amount value and the given tar amount value is higher than 99%, which effectively improves the accuracy of cigarette product design and product development efficiency.