Omar M. Saad, Randy Harsuko, and Tariq Alkhalifah

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

SiameseFWI is introduced as a crucial component in full waveform inversion (FWI) for effectively comparing simulated and observed seismic data. Utilizing a Siamese network, it transforms data into a shared latent space, allowing a robust comparison of representations. SiameseFWI employs two identical Convolutional Neural Networks (CNNs) with shared weights, operating in an unsupervised manner, eliminating the need for labeled data.  Evaluation of Marmousi2 and Overthrust models demonstrates its superior inversion performance over conventional FWI. Furthermore, application to Western Australia field data demonstrates its efficacy in real-world scenarios, proving resilient to noise and linear initial models. SiameseFWI offers improved performance with minimal additional cost compared to the conventional FWI and the benchmark deep learning methods.