Machine Learning Rebuilds Missing Porosity Logs from 1950s Oil-Field Paper Records

A machine-learning workflow that turns decades-old paper well logs into modern porosity estimates could help convert a depleted German oil field into a high-temperature underground heat store. Researchers at the Karlsruhe Institute of Technology and their collaborators report that geologically informed machine learning can reconstruct neutron porosity logs across the Leopoldshafen oil field in the northern Upper Rhine Graben, where nearly all wells lack the modern logging data needed to plan seasonal thermal energy storage. The study, published in Geothermal Energy, shows that adding a simple lithology flag derived from rock-physics reasoning substantially improves porosity predictions compared with purely data-driven models.

The work targets a specific obstacle in repurposing mature hydrocarbon fields for High-Temperature Reservoir Thermal Energy Storage, or HT-RTES. Such systems are designed for injection temperatures exceeding roughly 100 degrees Celsius, and the DeepStor research infrastructure at KIT is planned to test storing heat at injection temperatures of up to 140 degrees Celsius in the depleted reservoir. Mature fields are attractive storage targets because their reservoir properties are comparatively well constrained, and reusing existing infrastructure and historical geological data reduces the exploration risks of greenfield geothermal projects. But at Leopoldshafen, most wells were drilled between 1950 and 1963, predating modern neutron porosity logging suites, so the petrophysical record is sparse exactly where it matters most.

The data inventory reflects that history. Across 24 boreholes, 23 of them within the Leopoldshafen field and one, Stutensee-1, in the adjacent tectonic block to the east, the only measurements consistently available are self-potential and short-normal resistivity logs. Gamma-ray logs exist in just nine boreholes, and neutron porosity and density logs are restricted to two wells drilled in the early 1980s, LH-5 A and Stutensee-1. Because the predictive model had to apply across the entire field, the researchers restricted the input features to the ubiquitous self-potential and short-normal logs, digitized from analogue paper records.

Before any modelling, the team built a quality-control pipeline for the legacy curves. The logs were depth-matched to ensure stratigraphic correlation across the field, and extreme outliers were flagged with an interquartile-range filter at 1.5 times the IQR threshold. Importantly, statistical outliers were not discarded automatically: each flagged point was checked against log continuity, local geological trends, and evidence of acquisition or digitization artifacts, so that only isolated spikes, abrupt discontinuities, or demonstrable digitization errors were removed. This filtering eliminated roughly 1.8 percent of all data points while preserving genuine geological variability. The analysis was further limited to depths greater than 650 meters, focusing on the reservoir and caprock succession relevant to the planned storage system.

Standardization then harmonized responses across decades of drilling campaigns, different tool calibrations, and varying mud properties. Self-potential logs were rescaled to a target range of 10 to 80 millivolts, and short-normal resistivity logs to 1 to 2.5 ohm-meters, using min-max scaling that preserves the shape of each distribution while aligning relative responses between wells. The procedure corrected notable acquisition faults, including a polarity reversal in well LH-1 A, where permeable sandy formations had incorrectly produced positive self-potential deflections. After standardization, the logs showed stable baselines and physically plausible values, in contrast to the erratic baseline shifts and drift of the raw data.

The geological ingredient came from cross-plotting the standardized self-potential against the standardized short-normal response. Guided by independent lithological interpretations from drilling reports, the team drew empirical decision polygons in this two-dimensional feature space, iteratively refined until the classification agreed maximally with the reference geological information. Shale intervals cluster at elevated self-potential values with low resistivity, while sandstones cluster at lower self-potential and higher resistivity. The result was a binary lithology flag, zero for shale and one for sandstone, that groups marls and other fine-grained intervals into a non-reservoir class and reservoir-relevant sandstones into a sandstone class. The researchers emphasize this flag is a simplified petrophysical classification rather than a complete facies model.

Five supervised algorithms were then compared: Random Forest, Gradient Boosting, Extreme Gradient Boosting, Support Vector Regression, and Multi-Layer Perceptron neural networks, each tuned with a deliberately small grid search to avoid overfitting to a single training well. Two input configurations were tested. Model 1 used only the scaled self-potential and resistivity logs, while Model 2 added the lithology flag. Training and validation used the Stutensee-1 well, split 80 percent training and 20 percent blind test with k-fold cross-validation, and the independent well LH-5 A was reserved exclusively as a blind test. Model 2 outperformed Model 1 across all tested algorithms, and Random Forest was selected for subsequent analysis because of its interpretability, resistance to overfitting, and lower computational complexity compared with neural networks.

Blind validation in LH-5 A produced a coefficient of determination of approximately 0.60 over the depth interval from 1,195 to 1,250 meters, reproducing both the magnitude and vertical variability of the measured log. The main failure occurred in a gas-saturated reservoir interval between 1,204 and 1,208 meters, where the model captured a perturbation in neutron porosity but predicted it with an inverse trend. The cause traces to severe class imbalance in the training data: about 81 percent of samples, 5,813 in total, were shale, while only 19 percent, 1,112 samples, represented sandstone units. The model was therefore biased toward shale-dominated responses and had little exposure to gas-bearing signatures. A lithology-conditional correction, a linear regression fitted between predicted and measured neutron porosity in the sandstone intervals of Stutensee-1, raised the coefficient of determination to roughly 0.70 and improved agreement across the entire interval, including the gas-bearing section.

Applied across the remaining Leopoldshafen wells, the corrected workflow predicted neutron porosity values of approximately 21 percent to more than 35 percent in the calcareous sandstone layers and about 23 to 30 percent in the marls, with the Niederrödern Formation sandstones showing the highest values. The greatest variability appeared in the Niederrödern marls of well LH-2, which intersects the Leopoldshafen Fault. In LH-13, the well closest to the planned DeepStor site, predicted values locally exceeded 35 percent in the Karlsruhe Calcareous Sandstones. Directly measured effective porosity from plug samples analyzed by helium pycnometry was systematically lower, ranging from 15.8 to 26.2 percent in plugs from LH-13 and Stutensee-1. The authors attribute this difference to neutron logs capturing a more total pore-volume signal, while plugs come from the more competent parts of the well, and conclude the predicted values fall within a realistic range for these sandstones.

The researchers are careful about the limits of their results. Uncertainty in the predicted porosity remains considerable, reflecting the indirect relationship between electrical logs and porosity, the analogue origin and variable quality of the legacy data, the restriction of training to a single well, and the absence of density or sonic logs that would add independent constraints. They therefore recommend treating the reconstructed porosity as input for probabilistic reservoir modeling rather than as a deterministic parameter. The patchy distribution of predicted porosity, with no systematic vertical or lateral trend, may partly reflect mineral precipitation from highly saline formation waters of roughly 120 grams per liter, which can clog pore throats and reduce effective porosity, but without direct calibration at the DeepStor location the estimates should be interpreted cautiously until site-specific core, log, or fluid-chemistry data become available. Even so, the study demonstrates that legacy self-potential and resistivity logs, often dismissed as insufficient for quantitative petrophysical analysis, can be transformed into usable subsurface information when combined with physics-guided feature engineering and machine learning, offering a transferable approach for other mature basins with comparable archival data.

Subject of Research: Machine-learning reconstruction of neutron porosity logs from legacy well data to support high-temperature reservoir thermal energy storage in the Leopoldshafen oil field

Article Title: Geologically informed machine learning for neutron porosity reconstruction from legacy well logs: implications for high-temperature reservoir thermal energy storage

Article References: Tolba, A. T., Garipi, X., Bauer, F., Kohl, T., & Schill, E. (2026). Geologically informed machine learning for neutron porosity reconstruction from legacy well logs: implications for high-temperature reservoir thermal energy storage. Geothermal Energy, 14(1), Article 29. https://doi.org/10.1186/s40517-026-00414-4

Image Credits: AI Generated

DOI: 10.1186/s40517-026-00414-4

Keywords: machine learning, neutron porosity, legacy well logs, reservoir thermal energy storage, Leopoldshafen oil field, Upper Rhine Graben, DeepStor, petrophysics, Random Forest, self-potential log, short-normal resistivity, depleted hydrocarbon reservoir