Seismic lithofacies prediction via deep learning with prior geological constraints

ABSTRACT Reliable lithofacies prediction from seismic data was essential for advancing exploration and characterization in hydrocarbon reservoirs, CO2 storage, geothermal energy recovery, and groundwater management. Lithofacies prediction from seismic data under a data-driven supervised deep learning framework was often subject to issues of geological prior inconsistency, poor generalizability, and weak interpretability. To address these challenges, this study proposed the multi-seismic with multi-geological constraint Net (MSMGNet), a deep learning-based lithofacies prediction approach that not only incorporated multi-seismic information but also took into account prior geological knowledge, including stratigraphic differences, prior lithologic probabilities, lithofacies transition probabilities, and thickness distributions of the lithofacies. The input-based strategy and loss-based strategy were proposed to incorporate statistical geological features for constraining lithofacies prediction. In the input-based approach, geological features were embedded as structured input vectors, where stratigraphic units were encoded using one-hot encoding, and lithofacies proportions, transition probabilities, and thickness distributions were vectorized into numerical feature representations. In the loss-based approach, these geological features were formulated as regularization terms within the loss function, where Kullback–Leibler (KL) divergence was used to penalize deviations between predicted and prior distributions. Cross-well blind tests in a complex coal-bearing clastic reservoir demonstrated that incorporating these statistical geological features significantly improved lithofacies prediction performance. More importantly, geology-constrained deep learning approaches enhanced the ability to capture lithological variations between stratigraphic units, characterize thickness distributions, and identify thin sandstone and coal layers. The input-based MSMGNet, which integrated all four geological constraints, achieved the highest prediction accuracy and geological consistency in comparison with the baseline model without geological constraints and the loss-based MSMGNet.