Abstract: The gas sealing behavior of coal reservoirs is affected by lithological assemblages in coal measures, and the accurate identification and prediction of such coal assemblages is critical for evaluating source reservoir cap configurations and predicting gas bearing potential. Traditional lithological assemblages rely on labor interpretation, with disadvantages of high cost, low inefficiency and strong subjectivity. Three machine learning algorithms were compared in this research, including light gradient boosting machine (LightGBM), random forest (RF) and extreme gradient boosting (XGBoost), and the roof/floor strata of No.8 coal seam of Benxi Formation in the northern Mizhi area of Ordos Basin was chosen as the study area. And the favorable lithological assemblages in the coal measure were assessed using logging parameters inversion by optimal algorithm in conjunction with coalbed methane content, based on well logging and core observation data. Results indicated that LightGBM model yielded the highest prediction accuracy at 0.93, compared with 0.92 for Random Forest and 0.91 for XGBoost. Regional inversion by LightGBM showed that mudstone-coal-mudstone assemblage was the most widely distributed, followed by sandstone-coal-mudstone assemblage and the mudstone-coal-sandstone assemblage. Integrated analysis of coal thickness and measured gas content revealed that the limestone-coal-mudstone assemblage exhibited the highest gas abundance, with an average of 168.6 m³/m². And it was the most favorable lithological assemblage in the study area under present samples and gas bearing content indices, and it was predominantly distributed in the southeastern part of the study area. This study effectively overcomes the limitations of traditional lithological assemblage identification, providing new insights and technical references for the comprehensive exploration and development of coal-measure unconventional natural gas.