Acquiring reliable lithology information is a critical step for geological analysis since many basic jobs in early exploration have to be completed under the application of lithological materials. Lithology prediction then is always regarded as a research hotspot in geosciences. XGBoost is proved to be more powerful on pattern recognition than classic models, as it takes advantages of gradient boosting, classification tree, regularization, and other advanced machine learning techniques, thus being more potential to provide an ideal solution for lithology prediction. Nonetheless, this model is difficult to obtain optimal results due to the employment of many hyper-parameters, and will be low-efficient when dealing with many variables. Therefore, two computing techniques, continuous restricted Boltzmann machine (CRBM) and particle swarm optimization (PSO), are introduced to improve prediction performance of XGBoost. CRBM can extract fewer while more significant features from original data, and PSO will automatically optimize hyper-parameters during training process. Data used for validation is derived from tight sandstone reservoirs of member of Chang 4 + 5, western Jiyuan Oilfield, Ordos Basin, Northern China. Three experiments are designed to verify prediction capability of the proposed model. In order to highlight validation effect, two classic predictors named support vector machine (SVM) and gradient boosting decision tree (GBDT) are applied to create a contrast. The total prediction accuracy and the respective accuracy of each lithology produced by CRBM-PSO-XGBoost are all the highest in three experiments, well demonstrating the proposed model is effective to predict the lithology of tight sandstone reservoirs and has better robustness.