Memory efficient implementation of gravity inversion

The distribution of density plays a crucial role in interpreting subsurface geological structures for resource exploration and deep-structure research. However, conventional gravity inversion often involves storing a dense, large sensitivity matrix, demanding significant computational resources. To enhance the memory efficiency of gravity inversion, we introduce an intelligent data-sensing-based inversion to minimize the data involved in the inversion process. Initially, we validate the effectiveness of our approach using two synthetic data. Our intelligent data sensing approach employs a loss-controlled compression technique to sense the optimal subset from the original gravity anomaly. A user-specified threshold determines the degree of the information loss. Subsequently, we invert the optimal subset, but not all is used in each iteration. Instead, we implement an inverting-testing strategy during the iteration, where part of the optimal subset serves as the inverting dataset, and another part serves as the testing dataset. The amounts of inverting and testing datasets are equal in each iteration but may vary across iterations based on the inversion process. We conduct statistical analyses to assess the impact of the intelligent data sensing approach on memory requirement during gravity inversion. The proposed method is applied to two field datasets, demonstrating that it requires only a small percentage of the total gravity anomaly to reconstruct density distribution nearly identical to those obtained using the complete optimal subset.