Supervised pre-training and prior-guided fine-tuning for deep learning-based scatter interpolation and geophysical inversion

SUMMARY Interpolating scatter data obtained from discrete observations is essential for the continuous representation of subsurface media. Traditional interpolation algorithms typically rely on weighting the relationship between interpolation points and nearby known points, which makes it more difficult to incorporate multisource data and prior constraints as the amount of information increases. This study explores the use of deep neural networks to replace traditional interpolants, constructing a deep learning-based scatter interpolation workflow that integrates prior information through isotropic or anisotropic smoothness loss functions, addressing traditional methods’ limitations. To enhance the capability of the deep neural network for sparse scatter interpolation, we synthesized a large number of scatter-velocity model pairs to pre-train it using supervised learning. The pre-trained network is further adapted to specific interpolation tasks by physics-guided unsupervised fine-tuning to achieve stable interpolation results. Due to the flexibility of incorporating multisource information through input or supervised loss and imposing constraints of geophysical laws through unsupervised loss, our DL-based interpolation can be easily extended to solve geophysical inversion problems that jointly fits both data and geophysical laws. Our experiments validate the effectiveness of this workflow and demonstrate its potential in multi-information-constrained geophysical scatter interpolation, which forms the basis for multi-information inversion. This work not only advances machine learning algorithms for geophysical scatter interpolation but also provides valuable insights for deep learning geophysical inversion involving multiple data sources, and physical laws.