A Hybrid CNN-LSTM Deep Learning Approach for Basement Depth Prediction in Structurally Complex Sedimentary Basins Using Gravity Data

The inversion of gravity data can be used to detect sediment-basement interfaces, which is crucial for understanding basin architecture and tectonic evolution. It is true that traditional inversion techniques are powerful, but they have issues related to their non-uniqueness, their computational cost, and their reliance on initial models. In this paper, a new deep learning-based method for inverting gravity anomaly profiles using Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks is presented. Synthetic gravity data were generated from 30,000 forward-modeled basin geometries with stochastic variability. With synthetic profiles, CNN-LSTM architecture is trained and validated based on normalized Bouguer anomalies. In synthetic validation data, a hybrid CNN-LSTM architecture successfully recovered both basin shapes and high-frequency structural features to high extent and reveals an average error less than 10%. A comparison of the model's accuracy with seismic-derived depths in Wadi Kharit Basin (Eastern Desert, Egypt) reveals an average error of 15%. Hybrid inversion outperformed conventional depth methods such as Source Parameter Imaging (SPI) and slightly exceeded constrained inversion for forward gravity match accuracy. The hybrid deep learning approach offers an alternative to conventional gravity inversion. These results illustrate the potential for deep learning-based inversion to enhance subsurface structural interpretation, especially in structurally complex or data-limited circumstances.