Simultaneous prediction of porosity, saturation, and lithofacies from seismic data via multi-task deep learning

Prediction of reservoir parameters, including porosity, gas saturation, and lithofacies, from seismic data is of great significance for hydrocarbon reserves evaluation, reservoir quality assessment, and geological model building. Multi-task learning exhibits robust capabilities in simultaneously predicting multiple related parameters, which is desirable for estimating multi-reservoir parameters from seismic data. We propose the utilization of a seismic multi-reservoir parameter prediction network based on multi-task learning (SeisMRMTNet) informed by joint data distribution and physical constraints for the simultaneous inversion of porosity, gas saturation, and lithofacies. The SeisMRMTNet comprises two essential components: a shared feature extraction network and three task-specific networks. These networks adopt a three-dimensional sequence-to-sequence prediction paradigm to capture spatial features, hence improving seismic prediction stability. The shared feature extraction network extracts and maintains shared features between reservoir parameters and seismic information through the hard-sharing mechanism. Then, three task-specific networks establish nonlinear relationships between the shared features and the three different parameters, respectively. We incorporate physical constraints between reservoir parameters and integrate them into the network’s feature layer. Simultaneously, a two-dimensional joint data distribution constraint is applied between the predicted and actual values of gas saturation and porosity, which is incorporated into the loss function for optimization. The blind well tests on a deep heterogeneous carbonate reservoir demonstrate that the SeisMRMTNet achieves systematic improvement in prediction accuracy and better generalization performance compared to single-task learning. In particular, SeisMRMTNet can more effectively characterize formations where porosity and saturation vary significantly. Furthermore, SeisMRMTNet enhances the geological consistency and plausibility of reservoir prediction, yielding more reasonable results and data distribution for seismic reservoir characterization.