Symmetry-informed surrogates with data-free constraint for real-time acoustic wave propagation

A deep neural network is developed as a symmetry-informed surrogate model to estimate the stress of acoustic waves at any arbitrary location, time, and wave excitation width, as a kind of implicit representation of acoustic wave propagation. To improve the model performance, we suggest a physic-informed training residual by employing the symmetry constraint of acoustics. Compared to the FEM simulation using COMSOL Multiphysics, the R2 score of the surrogate model was greater than 98%, and its inference performance was about 215 times faster. The proposed residual can be characterized by three benefits: no architectural changes to the surrogate model are required, no labeled data or reference solutions are necessary, and the generalization ability of the surrogate model is enhanced in all input directions. Numerical investigation validates the superior generalization of the model trained with the symmetry constraint, along with the examination of the regularization rate and sampling approach, the additional hyperparameters induced by the proposed residual. Furthermore, we demonstrate the potential viability of real-time simulations from surrogate models through a practical application of the trained model for flaw detection. The results indicate that the capability of the data-driven surrogate model for generalized implicit representation can be enhanced by adopting the physic-informed training residual.

Automated summary

A deep neural network with symmetry-informed training is developed to estimate the stress of acoustic waves at any arbitrary location, time, and wave excitation width. The model outperforms FEM simulation in terms of R2 score and inference speed. The results demonstrate the superior generalization ability and faster inference of the model trained with symmetry constraint.