SiameseFWI: A Deep Learning Model for Full Waveform Inversion

Comparing simulated data with observed ones is an essential part of full waveform inversion (FWI). The kind of comparison employed is crucial to the success of FWI. We propose using a Siamese network to transform the observed and simulated data into a latent space so we compare the data representations. So, using two identical convolutional neural networks (CNN) with shared weights in FWI, we refer to as SiameseFWI, allows the networks to learn to extract key features from both simulated and observed data, and use the Euclidean distance to subsequently quantify the loss based on the transformed data. SiameseFWI operates as an unsupervised technique, eliminating the need for labeled data. In each FWI iteration, the Siamese network and the velocity model are sequentially updated to minimize the Euclidean distance loss. Empirical evaluation on the Marmousi2 model demonstrates that SiameseFWI achieves improved inversion performance compared to traditional FWI. Moreover, SiameseFWI exhibits superiority over the benchmark deep learning method while adding a small cost to the traditional FWI. We will share applications on real data in the presentation of this work.