Accurate evaluation of gas content in coalbed methane (CBM) reservoirs is essential for resource assessment and development planning. Current approaches often rely on predictive modeling but seldom incorporate geological models. With the increasing use of machine learning, conventional commercial modeling software is found to be inadequate for diverse predictive modeling requirements. A three-dimensional (3D) geological attribute modeling method for CBM gas content was developed by integrating potential-field theory with machine learning. The No. 2 coal seam of the Shanxi Formation in the Zhongyu block of the Qinshui Basin was selected as the case area. An implicit potential-field method combined with universal kriging interpolation was applied to construct the reservoir’s structural model. Support vector machines, random forests, stochastic gradient descent, and gradient boosting were then employed to predict gas content. K-fold cross-validation was used to optimize hyperparameters and assess model performance. Among the algorithms tested, stochastic gradient descent demonstrated the best predictive accuracy, with a coefficient of determination (R²) of 0.81 and an average relative error of 8.86%. A 3D gas content attribute model was subsequently generated by combining kriging interpolation with machine learning predictions. The resulting model shows significantly improved accuracy compared to conventional approaches and provides a clear, quantitative visualization of the spatial distribution of gas content. This integration of complex predictive models with geological modeling not only fully leverages the predictive capabilities of machine learning but also significantly enhances the accuracy of coal seam attribute models. The methodology presented here establishes a foundation for the future development of domestic software capable of 3D coalfield geological modeling.