Abstract: Lithology prediction is crucial for the exploration and development of unconventional coal-rock gas. However, due to limitations in seismic resolution and strong reflections from coal seams, predicting the lithology of thin interbeds and coal seam roofs and floors remains challenging. To address this challenge, a joint model- and data-driven high-resolution seismic lithology prediction approach is proposed. First, high-resolution lithology-sensitive parameters are obtained through seismic meme inversion to construct a high-quality labeled dataset. Subsequently, a gated recurrent neural network incorporating an attention mechanism is employed to extract correlated features in high-dimensional space, which establishes a classification model for nonlinear mapping from multiple lithology-sensitive parameters to multiple lithology types. Essentially, the proposed method cascades the advantages of model-driven and data-driven approaches. The method is applied to field data from a ~ 2000 m deep coal-rock gas reservoir. Results show that it successfully predicts independent thin coal seams with thicknesses of about 3 m, as well as geologic anomalies within coal seams. Comparisons of prediction results for two wells withheld from the inversions using different methods further demonstrate the prediction performance of the proposed approach. For one well, lithology prediction accuracy improves from 62.9% using a pure data-driven method and 85.0% using an optimal thresholding method to 90.0%; for the other well, accuracy improves from 62.0% and 77.0% to 87.0%. Field data tests validate the significantly enhanced lithology identification accuracy of the proposed method in both thin interbeds and layers masked by strong coal-seam reflections.