Abstract: Coaly source rocks have attracted considerable attention for their significant hydrocarbon generation potential in recent years. However, limited study is performed on utilizing geochemical data and well log data to evaluate coaly hydrocarbon source rocks. In this study, geochemical data and well log data are selected from two key wells to conduct an evaluation of coaly hydrocarbon source rocks of Jurassic Kezilenuer Formation in Kuqa Depression of Tarim Basin. Initially, analysis was focused on geochemical parameters to assess organic matter type, source rock quality, and hydrocarbon generation potential. Lithology types of source rocks include mudstone, carbonaceous mudstone and coal. The predominant organic matter type identified was Type III and Type II2, indicating a favorable hydrocarbon generation potential. Well log data are integrated to predict total organic carbon (TOC) content, and the results indicate that multiple regression method is effective in predicting TOC of carbonaceous mudstone and coal. However, the ΔlgR method exhibited limited predictive capability for mudstone source rock. Additionally, machine learning methods including multilayer perceptron neural network (MLP), random forest (RF), and Extreme Gradient Boosting (XGBoost) techniques are employed to predict TOC of mudstone source rock. The XGBoost performs best in TOC prediction with correlation coefficient (R2) of 0.9517, indicating a close agreement between measured and predicted TOC values. This study provides a reliable prediction method of coaly hydrocarbon source rocks through machine learning methods, and will provide guidance for resource assessment.