Transfer learning from seismic data for predicting reservoir productivity and lithologies for upstream optimization

Unofficial AI-generated summary based on the public title and abstract. Not an official translation.

🔬研究者:The transfer learning approach and feature engineering techniques are novel and could be extended to other subsurface prediction tasks.

🏢実務担当者:This method can be used as a de-risking tool for CCUS and geothermal project development by leveraging existing seismic data and limited well control.

Abstract Although seismic inversion is an excellent tool for reservoir characterization, its reliability hinges on extensive well control that is area-specific and nontransferable between surveys. This limitation becomes particularly acute in deepwater environments, where developing a dependable inversion cube can span more than a decade. To address these constraints, a transfer learning approach was presented for reservoir prediction that leverages well-dynamic data and poststack seismic attributes available during initial exploration phases. Data from a well-known oilfield (pre-salt Mero oilfield, offshore Brazil), covered by narrow-azimuth seismic data, were used to train and validate machine learning models, primarily self-organizing maps and random forest regression. From a depth-migrated amplitude volume, nine poststack attributes (complex trace, geometric, and voxel-based texture) were derived. After data similarity checks across different seismic surveys using principal component analysis, exploratory data analysis, and feature engineering, transfer learning was applied to other areas with a few wells that were used as blind tests for model validation. Next, self-organizing maps and random forest regression of flow capacity (trained with Mero data) were run in two other surveys hundreds of kilometers distant from the Mero field: Bacalhau and Lapa oilfields; there is no similar approach in the literature. At the Bacalhau field, a nonproductive volcanic mound was differentiated from hyperproductive carbonate mounds (blind test performance = 85%). At the Lapa field, the lower reservoir flow capacity at the perforated zones was correctly predicted (blind test performance = 75%). A disruptive method was created that can transfer machine learning from well-characterized areas to others with poor well control and make accurate reservoir productivity predictions. Therefore, it can be used as a powerful reservoir de-risking tool in upstream phase projects. Likewise, it can be applied to optimize any other subsurface projects like carbon capture, usage, and storage (CCUS), geothermal, and hydrogen storage in the energy transition context.

• openaire https://doi.org/10.1190/int-2025-0010first seen 2026-05-14 22:01:01