velocity inversion using the quadratic wasserstein metric...oct 03, 2020 · -8 -6 -4 -2 0 2 4 6...
TRANSCRIPT
Velocity Inversion Using the Quadratic WassersteinMetric
Srinath MahankaliMentor: Prof. Yunan Yang (NYU)
Stuyvesant High School
October 17, 2020MIT PRIMES Conference
Srinath Mahankali Velocity Inversion October 17, 2020 1 / 17
Motivation
Goal: find materials lying underground and their positions
Srinath Mahankali Velocity Inversion October 17, 2020 2 / 17
Motivation
Goal: find materials lying underground and their positions
Srinath Mahankali Velocity Inversion October 17, 2020 2 / 17
Approach
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin J(g(C), h) where J is the objective function
Convexity with respect to the velocity is beneficial
Srinath Mahankali Velocity Inversion October 17, 2020 3 / 17
Approach
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin J(g(C), h) where J is the objective function
Convexity with respect to the velocity is beneficial
Srinath Mahankali Velocity Inversion October 17, 2020 3 / 17
Approach
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin J(g(C), h) where J is the objective function
Convexity with respect to the velocity is beneficial
Srinath Mahankali Velocity Inversion October 17, 2020 3 / 17
Problems With the Squared L2 Norm
Beydoun-Tarantola, 1988
Nonconvexity of squared L2 norm and local minima
Brossier et al, 2010
Sensitivity of L2 norm to noise
-8 -6 -4 -2 0 2 4 6 8
-2: < x < 2:
0
0.05
0.1
0.15
0.2
0.25Graph of f(x)
-10 -5 0 5 10
-4: < s < 4:
0
0.05
0.1
0.15
0.2
0.25
L2 D
ista
nce
Squared L2 Norm of f(x) - f(x-s)
Srinath Mahankali Velocity Inversion October 17, 2020 4 / 17
An Alternative Objective Function
Engquist-Froese, 2013 introduced the Wasserstein metric for FWI
Yang, 2019
Convexity in translations and dilations of the dataInsensitivity to noise
Engquist et al, 2020
Low-frequency bias
-8 -6 -4 -2 0 2 4 6 8
-2: < x < 2:
0
0.05
0.1
0.15
0.2
0.25Graph of f(x)
-10 -5 0 5 10
-4: < s < 4:
0
50
100
150
W2
Dis
tanc
e
Squared W2 Distance of f(x) and f(x-s)
Srinath Mahankali Velocity Inversion October 17, 2020 5 / 17
What is the Wasserstein Metric?
Optimal transport introduced by Monge
Definition
The pth Wasserstein metric is defined as
Wp(f , g) =
(inf
T∈M(µ,ν)
∫Ω|x − T (x)|pdµ
) 1p
.
Srinath Mahankali Velocity Inversion October 17, 2020 6 / 17
What is the Wasserstein Metric?
Explicit formula when data is in one dimension
Theorem
Let g and h be two probability distributions defined on R, and letG (t) =
∫ t−∞ g(s) ds and H(t) =
∫ t−∞ h(s) ds. Then
W 22 (g , h) =
∫ 1
0(G−1(s)− H−1(s))2 ds.
Srinath Mahankali Velocity Inversion October 17, 2020 7 / 17
The Constrained Optimization Problem
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin W 22 (g(C), h)
Previous results involve convexity in changes of the data
Srinath Mahankali Velocity Inversion October 17, 2020 8 / 17
The Constrained Optimization Problem
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin W 22 (g(C), h)
Previous results involve convexity in changes of the data
Srinath Mahankali Velocity Inversion October 17, 2020 8 / 17
The Constrained Optimization Problem
1C(x)2
∂2ψ∂t2 −
(∂2ψ∂x1
2 + ∂2ψ∂x2
2 + · · ·+ ∂2ψ∂xn2
)= f (x, t),
ψ(x, 0) = 0, ψt(x, 0) = 0 on Ω,
∇ψ · n = 0 on ∂Ω
C∗(x) = argmin W 22 (g(C), h)
Previous results involve convexity in changes of the data
Srinath Mahankali Velocity Inversion October 17, 2020 8 / 17
Challenges and Solutions
Wave data is generally not a probability distribution
Normalize wave data k(t) by k(t) = k(t)+γ∫ T0
k(t)+γ dt
Computing the W2 distance is difficult for higher dimensions
Data is only a function of time → apply one dimensional formula
No explicit solution for wave equation in general
Srinath Mahankali Velocity Inversion October 17, 2020 9 / 17
Challenges and Solutions
Wave data is generally not a probability distribution
Normalize wave data k(t) by k(t) = k(t)+γ∫ T0
k(t)+γ dt
Computing the W2 distance is difficult for higher dimensions
Data is only a function of time → apply one dimensional formula
No explicit solution for wave equation in general
Srinath Mahankali Velocity Inversion October 17, 2020 9 / 17
Challenges and Solutions
Wave data is generally not a probability distribution
Normalize wave data k(t) by k(t) = k(t)+γ∫ T0
k(t)+γ dt
Computing the W2 distance is difficult for higher dimensions
Data is only a function of time → apply one dimensional formula
No explicit solution for wave equation in general
Srinath Mahankali Velocity Inversion October 17, 2020 9 / 17
Challenges and Solutions
Wave data is generally not a probability distribution
Normalize wave data k(t) by k(t) = k(t)+γ∫ T0
k(t)+γ dt
Computing the W2 distance is difficult for higher dimensions
Data is only a function of time → apply one dimensional formula
No explicit solution for wave equation in general
Srinath Mahankali Velocity Inversion October 17, 2020 9 / 17
Challenges and Solutions
Wave data is generally not a probability distribution
Normalize wave data k(t) by k(t) = k(t)+γ∫ T0
k(t)+γ dt
Computing the W2 distance is difficult for higher dimensions
Data is only a function of time → apply one dimensional formula
No explicit solution for wave equation in general
Srinath Mahankali Velocity Inversion October 17, 2020 9 / 17
Velocity Models in One Dimension
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Position
1.5
2
2.5
3
3.5
Vel
ocity
Velocity Model 1
0 0.5 1 1.5 2 2.5 3 3.5 4
Position
1.5
2
2.5
3
3.5
Vel
ocity
Velocity Model 2
0 0.5 1 1.5 2 2.5 3 3.5 4
Position
1.5
2
2.5
3
3.5
Vel
ocity
Velocity Model 3
Theorem
Suppose f (t) is nonnegative and compactly supported, and let k be thevelocity parameter in these three cases. Then W 2
2 (g(k), h) is convex in kover the interval (0, k∗].
Srinath Mahankali Velocity Inversion October 17, 2020 10 / 17
Ray Tracing
Velocity function of the form C(X , z) = a + bz
Distance and traveltime formulas:
X = 2p
∫ zp
0
dz√u2(z)− p2
T = 2
∫ zp
0
u2(z)√u2(z)− p2
dz
For source wave data f (t), the predicted wave data is approximatedby Apredf (t − Tpred)
Srinath Mahankali Velocity Inversion October 17, 2020 11 / 17
Ray Tracing
Distance and traveltime formulas:
X = 2p
∫ zp
0
dz√u2(z)− p2
T = 2
∫ zp
0
u2(z)√u2(z)− p2
dz
For source wave data f (t), the predicted wave data is approximatedby Apredf (t − Tpred)
Srinath Mahankali Velocity Inversion October 17, 2020 11 / 17
Ray Tracing
Distance and traveltime formulas:
X = 2p
∫ zp
0
dz√u2(z)− p2
T = 2
∫ zp
0
u2(z)√u2(z)− p2
dz
For source wave data f (t), the predicted wave data is approximatedby Apredf (t − Tpred)
Srinath Mahankali Velocity Inversion October 17, 2020 11 / 17
Ray Tracing
Distance and traveltime formulas:
X = 2p
∫ zp
0
dz√u2(z)− p2
T = 2
∫ zp
0
u2(z)√u2(z)− p2
dz
For source wave data f (t), the predicted wave data is approximatedby Apredf (t − Tpred)
Srinath Mahankali Velocity Inversion October 17, 2020 11 / 17
Velocity Model in Two Dimensions
Theorem
Assume f (t) is nonnegative and compactly supported. Then, W 22 (g , h) is
jointly convex in a, b over the following region U:
U := (a, b) ∈ R2 : a, b > 0,bXr
2a≥ S0, T (Xr , a, b) ≥ T (Xr , a
∗, b∗)
for some positive constant S0.
For large Xr , the convex region contains (a∗, b∗)
Nonuniqueness of solution fixed by adding multiple receiver locations
Srinath Mahankali Velocity Inversion October 17, 2020 12 / 17
Velocity Model in Two Dimensions
Theorem
Assume f (t) is nonnegative and compactly supported. Then, W 22 (g , h) is
jointly convex in a, b over the following region U:
U := (a, b) ∈ R2 : a, b > 0,bXr
2a≥ S0, T (Xr , a, b) ≥ T (Xr , a
∗, b∗)
for some positive constant S0.
For large Xr , the convex region contains (a∗, b∗)
Nonuniqueness of solution fixed by adding multiple receiver locations
Srinath Mahankali Velocity Inversion October 17, 2020 12 / 17
Velocity Model in Two Dimensions
Theorem
Assume f (t) is nonnegative and compactly supported. Then, W 22 (g , h) is
jointly convex in a, b over the following region U:
U := (a, b) ∈ R2 : a, b > 0,bXr
2a≥ S0, T (Xr , a, b) ≥ T (Xr , a
∗, b∗)
for some positive constant S0.
For large Xr , the convex region contains (a∗, b∗)
Nonuniqueness of solution fixed by adding multiple receiver locations
Srinath Mahankali Velocity Inversion October 17, 2020 12 / 17
Velocity Model in Two Dimensions
Srinath Mahankali Velocity Inversion October 17, 2020 13 / 17
Velocity Model in Two Dimensions
Srinath Mahankali Velocity Inversion October 17, 2020 13 / 17
Velocity Model in Two Dimensions
Srinath Mahankali Velocity Inversion October 17, 2020 13 / 17
Future Work
Study models with a larger number of parameters
Possible to study convexity using frequency instead of time domain
Srinath Mahankali Velocity Inversion October 17, 2020 14 / 17
Future Work
Study models with a larger number of parameters
Possible to study convexity using frequency instead of time domain
Srinath Mahankali Velocity Inversion October 17, 2020 14 / 17
Acknowledgements
I am extremely thankful to:
My mentor, Prof. Yunan Yang
The PRIMES program
My parents and my brother
Srinath Mahankali Velocity Inversion October 17, 2020 15 / 17
Acknowledgements
I am extremely thankful to:
My mentor, Prof. Yunan Yang
The PRIMES program
My parents and my brother
Link to my paper (posted on PRIMES website):https://arxiv.org/abs/2009.00708
Srinath Mahankali Velocity Inversion October 17, 2020 15 / 17
References
Wafik B. Beydoun and Albert Tarantola. First Born and Rytovapproximations: Modeling and inversion conditions in a canonicalexample. The Journal of the Acoustical Society of America,83(3):1045–1055, 1988.
Romain Brossier, Stephane Operto, and Jean Virieux. Which residualnorm for robust elastic frequency-domain full waveform inversion.Geophysics, 75, 05 2010.
Bjorn Engquist and Brittany D. Froese. Application of theWasserstein metric to seismic signals. arXiv preprint arXiv:1311.4581,2013.
Bjorn Engquist, Kui Ren, and Yunan Yang. The quadraticWasserstein metric for inverse data matching. Inverse Problems,36(5):055001, May 2020.
Srinath Mahankali Velocity Inversion October 17, 2020 16 / 17
References
Gaspard Monge. Memoire sur la theorie des deblais et des remblais.Histoire de l’Academie Royale des Sciences de Paris, 1781.
Yunan Yang. Analysis and Application of Optimal Transport ForChallenging Seismic Inverse Problems, 2019.
Srinath Mahankali Velocity Inversion October 17, 2020 17 / 17