PINN U-NET: A PHYSICS-INFORMED DEEP LEARNING MODEL FOR ACCURATE SALT IDENTIFICATION IN SEISMIC IMAGES

Abstract

Accurate identification of salt bodies from seismic images plays a crucial role in hydrocarbon exploration and subsurface geological analysis. Traditional deep learning models such as U-Net and its variants have achieved significant success in image segmentation tasks; however, they often fail to integrate the underlying physical constraints of seismic data, leading to suboptimal generalization and reduced boundary precision. To overcome these limitations, this paper proposes a Physics-Informed U-Net (PINN-UNet) model for salt identification using the TGS Salt Identification Challenge dataset. The proposed hybrid architecture combines the feature extraction power of ResNet34 with a Physics-Informed Neural Network (PINN) block, which introduces coordinate-aware spatial learning that enforces physical consistency within the segmentation process. The model was trained and validated using augmented seismic images, with careful optimization of learning rate, loss function, and normalization to enhance convergence and stability. According to experimental results, the suggested PINN-UNet performs better than the traditional U-Net and ResNet-based variations on the same dataset. The inclusion of the PINN component notably improved boundary localization and reduced false detections in complex salt dome regions. This work contributes a physically consistent deep learning framework that bridges the gap between data-driven and physics-guided approaches, making it a promising solution for real-world seismic interpretation and geophysical modeling.