High-dimensional wavefield solutions using physics-informed neural networks with frequency-extension (Huang, X., Alkhalifah, T., and Wang, F., 2022)

Utilizing the single reference frequency loss function, we proposed an efficient framework to obtain higher accuracy for high-dimensional wavefields compared to standard PINN. Specifically, we start by learning a narrow frequency range wavefield using a small network, and then adapt the trained model to work with wider frequency ranges using neuron splitting. The convergence and accuracy of this model in predicting wavefields for a larger frequency range exceed those obtained through random initialization of the NN model.

References

Huang, X., Alkhalifah, T., and Wang, F., 2022, "High-dimensional wavefield solutions using physics-informed neural networks with frequency-extension", submitted to the 83rd EAGE Annual Conference and Exhibition.