Research on Multiphysics Fields Reconstruction Combining Physics-Informed Neural Networks and Acoustic Tomography
To improve the accuracy of noncontact temperature and velocity measurements based on acoustic tomography (AT), this article proposes a physics-informed neural network integrated with AT (PINN-AT) for the simultaneous estimation of temperature and velocity fields. Sparse acoustic travel-time data are acquired by transmitters and receivers distributed around the target region. These measurements, together with the acoustic travel-time integral relationship and the nonlinear differential equations governing curved acoustic ray propagation, are incorporated into a physics-informed inversion framework that establishes a complete measurement model from acoustic sensing to field estimation. Unlike conventional approaches that rely on large-scale training datasets or complex inversion algorithms, our method solves the problem using only the acoustic data acquired from the current measurement task, together with the underlying physical laws. Consequently, it offers greater practicality and enhanced generalization capability. By embedding acoustic physical information directly into the inversion process, PINN-AT enables high-resolution reconstruction of temperature and velocity fields with enhanced physical interpretability, improved accuracy, and significant robustness to measurement noise. The effects of key parameters, including the number of hidden-layer neurons and the learning rate, are systematically investigated. Numerical results show that the proposed method accurately traces acoustic rays in accordance with Snell’s law. More importantly, the proposed method exhibits marked robustness to measurement noise, achieving accurate coupled reconstruction of temperature and velocity fields across a range of noise levels. Both computational fluid dynamics (CFD) simulations and experiments validate its effectiveness. In numerical simulations, the mean relative errors of the reconstructed velocity and temperature fields are below 17 % and 10 %, respectively. In experiments, the maximum relative error of the reconstructed temperature field is 1.89 %, while the mean relative error of the reconstructed velocity field is 8.84 %.