Omar M. Saad, and Tariq Alkhalifah

 GitHub: https://github.com/DeepWave-KAUST/FreqSiameseFWI-pub

FreqSiameseFWI is a deep learning-based misfit function designed to enhance and expedite the Full Waveform Inversion (FWI) process while addressing its inherent computational challenges and high nonlinearity. By employing a self-supervised learning approach within the FWI framework, FreqSiameseFWI effectively mitigates cross-talk noise and supports Multi-source FWI (MSFWI). Using the Siamese network architecture, this framework allows for comparisons of seismic data in a latent space, extracting spectral features through Fast Fourier Transform (FFT). The method demonstrates significant improvements in inversion speed, proportional to the number of stacked sources used to generate supershots while providing robust inversion performance with accuracy that outperforms traditional misfit functions, as validated on the Marmousi2 and Overthrust models.