Looking under the surface: Interpreting geological data with AI assistance

Projekt: ThermoOptiPlan: Optimierung der Planung und des Betriebs von Geothermiesystemen mittels innovativer Prognosetools

Forschende: Ning Qian, Monika Sester

Projektidee: This project aims to support the development of deep geothermal energy in northern German aquifers by reducing uncertainties about subsurface properties and operational risks. It develops AI-based methods to automatically analyze and integrate various geological and geophysical data sources into a digital twin with uncertainty assessment.

Understanding the geological structure beneath the Earth's surface is essential for applications such as energy exploration, geothermal development, and reservoir management. One of the most important sources of subsurface information is well log data. These measurements are collected along boreholes and record how different physical properties of the surrounding rocks change with depth. By analyzing these logging curves, geoscientists can infer rock types, identify geological layers, and evaluate reservoir properties.

Traditionally, interpreting well log data has relied heavily on expert knowledge. Geologists examine multiple logging curves together and use their experience to infer lithology and formation characteristics. While effective, this process is often time-consuming and subjective, especially when large volumes of data are involved. As modern drilling operations generate increasing amounts of logging data, there is a growing need for automated methods that can assist geoscientists in analyzing these datasets more efficiently.

Machine learning techniques have therefore become increasingly popular for well log interpretation. Early approaches used classical algorithms to classify lithology or predict reservoir properties based on logging measurements. Although these methods achieved encouraging results, they often treat each depth sample independently and rely on manually designed features. In reality, well log data form continuous sequences along the depth of a borehole, and geological formations usually exhibit layered structures that extend over long depth intervals. Ignoring these sequential relationships can limit the ability of models to capture meaningful geological patterns.

Recent advances in deep learning have opened new opportunities for analyzing sequential data. Neural network architectures designed for sequence modeling have been applied to well log analysis and have demonstrated the ability to automatically learn patterns from raw measurements. The attention mechanism has recently become a powerful tool for modeling complex dependencies in sequential data because it allows models to focus on the most relevant parts of the input while capturing long-range relationships. In the context of well log analysis, attention mechanisms provide a natural way to model interactions both along the depth sequence and across different logging curves. However, learning these two types of relationships simultaneously can significantly increase the complexity of the learning process.

In this work, we focus on improving how these relationships are represented in machine learning models. Our goal is to better capture the natural sequential structure of well log data while preserving meaningful interactions among different logging curves. To achieve this, we introduce a Sequence-based Decoupling Encoder (SDE) that separates cross-curve interactions from cross-depth dependencies so that each type of relationship can be learned independently. This decoupled design simplifies the learning process and reduces the computational complexity greatly. By learning these interactions separately, the model can build more structured representations of well log data and better reflect the inherent characteristics of geological sequences.

To demonstrate the potential of this approach, we evaluate our method using the publicly available FORCE2020 dataset. The experiments evaluate the proposed framework on two tasks: missing log reconstruction and lithology classification. Results (Fig.1) show that the proposed SDE encoder significantly improves reconstruction accuracy compared with baseline (coupled interaction) and recurrent models (RNN and LSTM), producing logging curves that better preserve local variations and geological features within missing intervals. Furthermore, SDE also improved lithology classification performance (Fig.2), demonstrating that the learned representations capture more informative subsurface characteristics.