The demo code for "Soft residual attention-based physics-informed neural networks for acoustic phased array-based forward radiation wavefield modeling in logging-while-drilling"
Authors: Qiang Feng, Wei Zhang, Yibing Shi, Zhipeng Li and Weihong Xie
Abstract: Phased array-based look-ahead of drill bits technology serves as a common method in logging-while-drilling (LWD). It utilizes an array of acoustic sources to excite synthetic forward acoustic wavefields in the medium and detect geological anomalies. Therefore, understanding multi-source synthesized forward wavefield propagation and diffusion properties in geological structures is crucial. Previous numerical modeling strategies mainly depend on finite element methods (FEM) with refined meshes that are computationally expensive. Recently, physics-informed neural networks (PINNs) have attracted considerable attention as a novel paradigm for solving partial differential equations (PDEs). However, it encounters over-fitting and point-wise weight imbalance challenges when learning multi-source complex wavefields, which makes phased array-based acoustic wavefields reconstruction difficult. This work presented a soft residual attention-based physics-informed neural network (SRA-PINN) to address the above problems and simulate phased array-based synthetic forward wavefield propagation and diffusion phenomena in the medium. Firstly, the acoustic wave equation is embedded into the loss function as a penalty term to supervise the training process of the deep neural network (DNN) to characterize the propagation and diffusion behavior of the acoustic wavefield. Furthermore, the proposed soft residual attention mechanism assigns point-wise adaptive weights based on the point-wise PDEs residual. This prompts DNN to alleviate excessive attention on easy-to-learn No-wave areas and shifts to difficult-to-learn multi-source overlapping areas to improve reconstruction quality. Finally, the several numerical results demonstrate that the proposed SRA-PINN-based wavefield simulation method is not only more accurate than the original PINN but also significantly faster in acoustic wavefield prediction compared to the FEM. It efficiently simulates multi-source overlapping acoustic wavefields with limited data and without meshes.
- Dependent environment:
- Pytorch = 1.13.1
- numpy
- matplotlib
- pandas
- scikit-learn
- Run the code: -Example for Homogeneous velocity Model: SRAPINN_IBS4A45.py Simulation of four acoustic sources in an infinite domain, with a wavefield deflection angle of 45 degrees. Additionally, we provide model checkpoints that can be run directly to obtain results.
Corresponding author: [email protected] (Wei Zhang) or [email protected] (Qiang Feng)