Joint inversion of multi-height gravity and vertical gradient via physics-informed neural network Available to Purchase

Gravity inversion is a widely used method for inversing subsurface structures, which is cheap and environmentally friendly. The common gravity inversion methods predict 3D subsurface density from 2D surface gravity anomaly, which is troubled by strong multi-solution and low precision. Deep learning (DL) methods are adept at solving ill-posed problems such as gravity inversion. Recently, a selfsupervised gravity inversion method has solved the above problems by introducing a guideline to provide depth resolution in gravity inversion. However, the guideline is calculated from well-log or other inversion methods, which has raised costs significantly. This abstract proposes a lowcost and high-precision gravity inversion via a physicsinformed neural network (PINN). The proposed method obtains depth resolution by 3D gravity anomalies on multiheight and vertical gradient data. Thus, the inversion accuracy is greatly improved with limited cost augment. The proposed method is pre-trained on the synthetic data to learn the gravity inversion mechanism. Then the pretrained model is fine-tuned on the target data by a closed loop of gravity forward and inversion models to further improve the inversion results. The feasibility and effectiveness of the proposed method are verified by experiments.