Accelerating Mode Conversion Simulation for Ultrasonic Wave Using Physics-informed Neural Networks

Abstract

Ultrasonic wave simulation serves a crucial role in comprehending the wave propagation process across various fields, including Non-destructive testing, Medicine, and Geophysics. While past approaches relied on numerical analysis for simulation, a drawback was the time-consuming nature of analyses, particularly in complex shapes and conditions, making real-time applications challenging. Recent advancements have explored simulation research using artificial neural networks based on physical knowledge to reduce inference time. However, challenges persist in achieving multi-mode simulation and accurately inferring changes in defect locations. Therefore, this paper proposes a physics-informed neural network for mode conversion simulation of ultrasonic wave. The proposed model adopts a U-net architecture and is trained with a primary emphasis on predicting subsequent steps. The training process involves simulating mode conversion in an elastic solid, incorporating data with random walk noise, a network that encompasses defect information, a loss function based on the two-dimensional wave equation in an elastic solid, and self-adaptation weights. Ultimately, quantitative analysis is conducted using Test MAE, SSIM, and Physical evaluation, comparing the proposed model with comparative model. This study holds significance as it conducts ultrasonic mode conversion simulation through a physics-informed neural network, establishes a model capable of inference based on changes in defect location, and contributes to the acceleration of simulations.