A hybrid unsupervised-supervised deep learning framework for sandstone thickness prediction from seismic data

Accurate sandstone thickness prediction from seismic data is vital for reservoir characterization and well placement optimization. However, conventional deep learning methods are often hindered by inefficient sequential processing or excessive computational costs when handling long seismic traces. To overcome these limitations, we propose a two-stage deep learning framework. First, an unsupervised feature extraction network derives high-dimensional latent representations directly from seismic data. Second, a novel reservoir sequence prediction network—utilizing efficient ProbSparse self-attention and self-attention distilling—maps these features to sandstone thickness, even with limited well-log training data. When applied to a field dataset with limited borehole control, our method resolved sandstone bodies thickness about 15∼20 m and achieved a Mean Absolute Percentage Error of just 3.7% at blind validation wells. This hybrid approach offers a robust, computationally efficient solution for high-precision reservoir prediction in data-constrained environments. • A hybrid unsupervised-supervised framework for thickness prediction. • Computationally efficient long sequence modeling for seismic data. • Enables high-accuracy thickness mapping despite significant well data constraints.