Abstract : To share the understanding of the paper, the original text can be found in Peng Jin, Xitong Zhang, Yinpeng Chen, Sharon Xiaolei Huang, Zicheng Liu, Youzuo Lin, Unsupervised learning of full-waveform inversion: connecting CNN and partial differential equation in a loop. The paper was published
in The top meeting in computer science is ICLR.
1. Paper Contribution
- An unsupervised FWI network is proposed. In fact, it is a bit far-fetched to say "unsupervised", because its supervisory information (velocity model) can calculate the loss between the data obtained through forward modeling and the original data.
- Made a data set OpenFWI, which is specially introduced in another paper. It is very important for researchers in this direction.
2. Thesis work
- CNN for inversion
- Forward modeling with PDE
2.1 Forward modeling
∇ 2 p ( r , t ) − 1 v ( r ) 2 ∂ 2 p ( r , t ) ∂ t 2 = s ( r , t ) (1) \nabla^2 p(\mathbf{r}, t) - \frac{1}{v(\mathbf{r})^2} \frac{\partial^2 p(\mathbf{r}, t)}{\partial t^2} = s(\mathbf{r}, t) \tag{1} ∇2p(r,t)−v(r)21∂t2∂2p(r,t)=s(r,t)( 1 )
wherep ( r , t ) p(\mathbf{r}, t)p(r,t ) is atttTime t , positionr \mathbf{r}The pressure wavefield of r , v ( r ) v(\mathbf{r})v ( r ) is the velocity graph,s ( r , t ) s(\mathbf{r}, t)s(r,t ) is the source item.
The forward modeling process is
p ~ = f ( v ^ ) (2) \tilde{\mathbf{p}} = f(\hat{\mathbf{v}}) \tag{2}p~=f(v^)(2)
标准的有限差分法
∂ 2 p ( r , t ) ∂ t 2 ≈ 1 ( Δ t ) 2 ( p r t + 1 − 2 p r t + p r t − 1 ) + O ( ( Δ t ) 2 ) (5) \frac{\partial^2 p(\mathbf{r}, t)}{\partial t^2} \approx \frac{1}{(\Delta t)^2} \left(p_\mathbf{r}^{t + 1} - 2 p_\mathbf{r}^t + p_\mathbf{r}^{t - 1} \right) + O((\Delta t)^2)\tag{5} ∂t2∂2p(r,t)≈(Δt)21(prt+1−2prt+prt−1)+O ( ( t ) _2)( 5 )
whereprt p_\mathbf{r}^tprtmeans ttThe wave field at time t , prt + 1 p_\mathbf{r}^{t + 1}prt+1means t + Δtt + \Delta tt+Δt time .OOO indicates the same order, and the corresponding data is discarded.
According to the chain rule, the loss L can be calculated \mathcal{L}L 对应对速度的梯度
∂ L ∂ v ( r ) = ∑ t = 0 T [ ∂ L ∂ p ( r , t ) ] ∂ p ( r , t ) ∂ v ( r ) (7) \frac{\partial \mathcal{L}}{\partial v(\mathbf{r})} = \sum_{t = 0}^T \left[\frac{\partial L}{\partial p(\mathbf{r}, t)}\right] \frac{\partial p(\mathbf{r}, t)}{\partial v(\mathbf{r})} \tag{7} ∂v(r)∂L=t=0∑T[∂p(r,t)∂L]∂v(r)∂p(r,t)(7)
2.2 Inversion loss function
L ( p , p ~ ) = L pixel ( p , p ~ ) + perceptual L ( p , p ~ ) (8) \mathcal{L}(\mathbf{p}, \tilde{\mathbf{p}}) = \mathcal{L}_{\textrm{pixel}}(\mathbf{p}, \tilde{\mathbf{p}}) + \mathcal{L}_{\textrm{perceptual}}(\mathbf{p }, \tilde{\mathbf{p}}) \tag{8}L(p,p~)=Lpixel(p,p~)+Lperceptual(p,p~)( 8 )
wherep \mathbf{p}p和p ~ \tilde{\mathbf{p}}p~represent the input and reconstructed seismic wave data respectively.
You can use l 1 \mathcal{l}_1l1与l 2 \mathcal{l}_2l2Norm weighted sum to define the pixel loss
L pixel ( p , p ~ ) = λ 1 l 1 ( p , p ~ ) + λ 2 l 2 ( p , p ~ ) (9) \mathcal{L}_{\ textrm{pixel}}(\mathbf{p}, \tilde{\mathbf{p}}) = \lambda_1 \mathcal{l}_1(\mathbf{p}, \tilde{\mathbf{p}}) + \ lambda_2 \mathcal{l}_2(\mathbf{p}, \tilde{\mathbf{p}}) \tag{9}Lpixel(p,p~)=l1l1(p,p~)+l2l2(p,p~)( 9 )
L perceptual ( p , p ~ ) = λ 3 l 1 ( ϕ ( p ) , ϕ ( p ~ ) ) + λ 4 l 2 ( ϕ ( p ) , ϕ ( p ~ ) ) (9) \mathcal {L}_{\textrm{perceptual}}(\mathbf{p}, \tilde{\mathbf{p}}) = \lambda_3 \mathcal{l}_1(\phi(\mathbf{p}), \phi (\tilde{\mathbf{p}})) + \lambda_4 \mathcal{l}_2(\phi(\mathbf{p}), \phi(\tilde{\mathbf{p}})) \tag{9 }Lperceptual(p,p~)=l3l1( ϕ ( p ) ,ϕ (p~))+l4l2( ϕ ( p ) ,ϕ (p~))( 9 )
whereϕ ( ⋅ ) \phi(\cdot)ϕ ( ⋅ ) represents the VGG-16 feature extraction network trained with ImageNet.
3. Related work
- physical drive
- data driven
4. Summary
Forward and reverse, walk on two legs.