Artificial Intelligence–Driven Reconstruction of Legacy Well Logs to Unlock Heat Storage Potential in Depleted Hydrocarbon Reservoirs: A Case Study from the Upper Rhine Graben

Depleted hydrocarbon reservoirs represent promising candidates for subsurface heat storage due to their well-organized geological architecture and long production history. However, effective reuse of many mature fields is hindered by incomplete or missing well log data sets, especially in older wells where data only exists in paper-based form. In this study, we present an integrated, data-driven workflow that leverages artificial intelligence (AI) to reconstruct missing petrophysical logs and reassess reservoir properties for geothermal heat storage applications. The approach has been demonstrated in a depleted field in Germany's Upper Rhine Graben – one of the most promising geothermal provinces in Europe.

Older well logs were systematically digitized, standardized, and subjected to rigorous statistical data cleaning to remove collection artefacts while preserving the underlying geological signal. Lithological information was obtained through multi-log cross-plot analysis and coded as an additional input function in the machine learning model. This step proved to be important in limiting petrophysical variability and reducing non-uniqueness in predictions. Several supervised machine learning algorithms were evaluated. Hyperparameter optimization was performed for each algorithm to identify the optimal model configuration and significantly reduce overfitting.

Due to a lack of data availability, only two modern wells with complete log suites were available. One well was used for model training and internal validation, while the other well was reserved exclusively for blind testing. The results show strong predictive performance across key petrophysical logs, with independent testing confirming the robustness and generalizability of the well model. Inclusion of lithological descriptors resulted in significant improvement in prediction accuracy and significant reduction in uncertainty compared to models based only on continuous log input.

The proposed workflow highlights the value of combining revitalization of legacy data with interpretable, well-constrained AI models. It provides a transferable methodology for unlocking the geothermal potential of depleting hydrocarbon reserves and supports data-driven decision-making for sustainable subsurface energy storage.