Effective gravity inversion of basement relief with unfixed density contrast using deep learning

Abstract

Recovering basement relief from gravity data plays a crucial role in understanding regional tectonics and advancing resource exploration. Traditional inversion methods typically assume a known density contrast between sedimentary layers and basement rocks to simplify the inverse problem, despite the reality that this contrast varies significantly. To overcome this limitation, we propose a deep learning approach to estimate basement relief from gravity data without requiring a fixed density contrast. We develop two distinct model generation methods to prepare the dataset and validate our neural network through comprehensive synthetic studies. Utilizing a CNN-LSTM architecture, which performs robustly across all tests, we apply this method to both synthetic and field case studies. The results demonstrate that our approach accurately estimates basement relief under variable density contrasts. Furthermore, our testing framework identifies the most effective network architectures and model generation strategies for tackling complex, multi-source geophysical problems.