NNFWI

Integrating Deep Neural Networks with Full-waveform Inversion: Reparametrization, Regularization, and Uncertainty Quantification

Architecture

Forward Simulation

Marmousi modelInital 1D model
BP2004 modelInital 1D model

Inversion based on Automatic Differentiation

Loss function

Noise levelMarmousi modelBP2004 model
$\sigma=0$
$\sigma=0.5$

Marmousi model

Noise levelTraditional FWINNFWI
$\sigma=0$
$\sigma=0.5$
$\sigma=1$

BP2004 model

Noise levelTraditional FWINNFWI
$\sigma=0$
$\sigma=0.5$
$\sigma=1$

Uncertainty Quantification using Dropout

Inverted $V_p$std($V_p$)std($V_p$)/$V_p$ $\cdot$ 100%