Answers are grounded in this paper and cite their sources.
Accurate prediction of porosity, permeability, and water saturation is important for oil and gas reservoir evaluation. Uneven label distributions can limit regression performance for sparsely represented values. This study develops a multi-task learning framework combining convolutional neural networks, gated recurrent units, attention, and bounded density-based loss weighting to predict these three properties from well logs. Kernel density estimation of the training labels is used to construct weights through a logarithmic inverse density transformation and bounded rescaling. Three configurations are compared: an unweighted single task network, an unweighted multi-task network, and a density-weighted multi-task network. In the reported comparison, the weighted model has the lowest MSE, RMSE, and MAE for all three targets. However, its reported R2 values for water saturation and permeability remain negative, indicating performance below a test set mean reference under the standard definition of R2. The results therefore indicate relative error reductions within the evaluated configurations rather than established practical reliability. Density-stratified evaluation, repeated runs, and independent well-based validation are needed to assess benefits for rare target values and generalization across wells.
No takes yet. Share an insight, caveat, or question.
Chen et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: