Machine learning predicts mineralogy in Bazhenov Formation with low error
Feb 24th 2026
A gradient boosting model wrapped in a regressor chain used well logs and thermal core profiling to predict mineral mass and volume fractions in a West Siberia unconventional reservoir, matching Litho Scanner results with low error.
- Gradient boosting regressor wrapped in a regressor chain delivered the best predictions for mineral weight and volume fractions.
- Compared with Litho Scanner results, the model achieved an average RMSE of 0.026 for weight fractions in the Bazhenov Formation.
- Combining conventional well logs with thermal core profiling improved predictions, especially for quartz, pyrite and the rock matrix volume.
- Mineral densities needed to convert mass to volume fractions were obtained by solving an optimization problem.
- Predicted mineral volumes fed a theoretical model that produced thermal conductivity values similar to experimental measurements.
- The workflow addresses challenges of unconventional rocks such as kerogen presence, heterogeneity and anisotropy by using multiscale data and multioutput modeling strategies.