On acoustic fields of complex scatters based on physics-informed neural networks

Area of Science:

• Acoustics
• Computational Physics
• Machine Learning

Background:

• Modeling scattered acoustic fields in complex structures is challenging.
• Traditional methods often require meshing, limiting their application to continuous fields.
• Accurate simulation is crucial for applications like non-destructive testing and structural health monitoring.

Purpose of the Study:

• To propose a novel modeling method for scattered acoustic fields using Physics-informed Neural Networks (PINNs).
• To demonstrate the effectiveness of PINNs in handling complex structures and continuous acoustic fields.
• To provide a meshless and accurate alternative for acoustic field prediction.

Main Methods:

• Utilized acoustic simulation software to generate training data for scattered acoustic fields.
• Embedded physical governing equations into the loss function of the PINN model.
• Trained and validated the PINN on simulated ultrasonic wave propagation and acoustic scattering from simple and complex structures.

Main Results:

• Achieved a mean square error (MSE) on the order of 10^-4 between predicted and ground truth scattered acoustic fields.
• Demonstrated accurate simulation of ultrasonic wave propagation and reflection.
• Successfully simulated the scattered acoustic field of a real, complex damaged structure.

Conclusions:

• PINNs offer an effective and accurate approach for modeling scattered acoustic fields, even in complex geometries.
• The meshless nature of PINNs makes them a versatile tool for continuous acoustic field prediction.
• This method provides a reliable alternative to traditional numerical techniques for acoustic simulations.