Application of data-driven and geostatistics-guided machine learning for petroelastic modelling

Calibration of dry rock frame models is a critical step in petroelastic modeling in 4D quantitative interpretation to accurately estimate changes in elastic properties resulting from saturation changes. Assigning mineral elastic moduli and rock frame hyperparameters has been extensively debated, with petrophysicists and rock physicists often adopting distinct approaches. This study evaluates traditional calibration strategies and explores the potential of machine learning (ML) for directly estimating elastic properties from petrophysical parameters and saturation changes. Two cases are examined: in the first, a data-driven approach employs a DNN trained on petrophysical properties from multiple wireline logs; in the second, where log data are insufficient, the DNN is trained on synthetic petrophysical data generated using geostatistical techniques. Performance evaluation demonstrates that the DNN-based approach is computationally efficient, requiring low memory while achieving accuracy comparable to, or exceeding, conventional grid-search techniques. Elastic properties predicted by the DNN on synthetic datasets further confirm its robustness. However, both ML and traditional methods struggle in specific depth intervals of the blind well, indicating potential errors in input parameters, output elastic properties, or rock frame models. To quantify uncertainty in DNN predictions, the Monte Carlo dropout technique is employed, providing a rigorous assessment of model reliability.