Machine learning experiments on velocity extraction from migration images

Zhan Niu, Daniel O. Trad

In full waveform inversion (FWI), the update of velocity is obtained by calculating the gradient of the misfit between recorded and predicted data, which is defined by the cross-correlation of the reverse time of receiver wavefield and source wavefield. Benefits can be achieved by solving a direct non-linear mapping between the correlation and model update. In this report, we train a fully connected neural network with residual blocks which allows migrated images to be directly mapped into velocity models. The input images and the true velocity model comes from reverse time migration results on randomly generated 4-layer models. The training is performed with ADAM optimizer combined with L1/L2 norms as the loss function. Performance and convergence of the neural network with different hyper-parameters are also investigated systematically. We have tested the trained model with different synthetic inputs. Results show the that the trained network is relatively model dependent which performs well on the validation set but does a poor job on datasets that come from different distributions.