2-D Niblett-Bostick magnetotelluric inversion - MTNet
2-D Niblett-Bostick magnetotelluric inversion - MTNet
2-D Niblett-Bostick magnetotelluric inversion - MTNet
You also want an ePaper? Increase the reach of your titles
YUMPU automatically turns print PDFs into web optimized ePapers that Google loves.
C = − ∑ ∑ [ −∑C<br />
= a a∑<br />
] ( σ σ − −∑ˆ<br />
σ ) [ −=<br />
∑ ∑a<br />
[ σσ<br />
+ − ∑ a σ ] σ.<br />
+ (10)<br />
J. RODRÍGUEZ et al.<br />
2-D <strong>Niblett</strong>-<strong>Bostick</strong><br />
extremely cumbersome, for it is necessary to cover many where<br />
possibilities during each epoch training session.<br />
Another approach to <strong>inversion</strong> based on neural networks<br />
bypasses the learning sessions by directly providing the<br />
algorithm with the relevant physics behind the particular<br />
(13)<br />
application. Zhang and Paulson (1997) applied a simple<br />
recursive regularization algorithm associated with a<br />
Hopfield artificial neural network (HANN) in order to E is the Ising Hamiltonian (Ising, 1925) which Tank<br />
solve the 1-D MT inverse problem with excellent results. and Hopfield (1986) showed is a never increasing quantity<br />
In the same paper the authors explored the application to as the dynamics defined by the following equation evolve<br />
the 2-D full nonlinear <strong>inversion</strong> of MT data. Our interest from given starting values:<br />
in this algorithm was triggered by the fact that there are<br />
no references in the literature attempting to continue their<br />
(14)<br />
research in this direction, even though the results presented<br />
are very promising. In this methodology there are no<br />
dedicated input-output defined units (as with the feed- sj<br />
forward network architecture). Instead, every neuron si<br />
is interconnected with every other sj neuron through a Tij<br />
weight link as detailed below. Every neuron has a twofold<br />
input-output purpose. We assume M apparent conductivity<br />
measurements and N blocks whose conductivities are<br />
unknown. According to equation (6), for given conductivity<br />
values sj, j = 1,...,N, the responses ˆsak,, k = 1,...,M, of the<br />
model can be expressed as<br />
(9)<br />
The misfit between sak (the data) and ˆsak, (the model<br />
responses) can be written as<br />
(10)<br />
Squaring as indicated and rearranging terms, this<br />
expression corresponds to<br />
(11)<br />
It can be noted that this is an unnecessarily long, and<br />
certainly very cumbersome way of representing C, the sum<br />
of squares. However, this particular form helps to recognize<br />
its correspondence to<br />
(12)<br />
Geologica Acta, 8(1), 15-30 (2010)<br />
18<br />
DOI: 10.1344/105.000001513<br />
(t+l) is the updated conductivity value of sj (t) which is either<br />
a given starting value or the result of a previous iteration.<br />
The relevance of dynamic is that as iterations proceed<br />
E decreases at each step. Since E corresponds to C, the<br />
dynamics ensure that the sum of the squared differences<br />
between data and model responses decrease at each step,<br />
thus leading to the model whose responses best fit the data<br />
in a least square sense. It can be noted that with the HANN<br />
the matrix elements Tij are known and provided by the<br />
physics of the problem, since they are the elements (AT ki kj i akj<br />
ak<br />
ki ak i 1 M<br />
kikj<br />
akj<br />
i<br />
2 2<br />
i=<br />
1 j¹<br />
i=<br />
1 k=<br />
1 2 k=<br />
1<br />
i=<br />
1 2 2k<br />
= 1k<br />
= 1 k=<br />
j1=<br />
1 ( σ ) 2 .<br />
2∑<br />
∑ ( σ ak<br />
ak)<br />
.<br />
2 k<br />
k<br />
=<br />
1<br />
N N M<br />
N M<br />
M<br />
1<br />
M 1<br />
1 2<br />
= − ∑ ∑ [ −∑<br />
a N N<br />
N<br />
( ki<br />
2<br />
σ<br />
akj)<br />
1]<br />
E −1<br />
. Niσ<br />
j N−<br />
N<br />
∑ ∑ [ − ∑<br />
ak<br />
(11)<br />
akiσ<br />
i + ∑ aki<br />
2 i=<br />
1 j¹<br />
i=<br />
1 2 k=<br />
1<br />
i=<br />
1 ∑ ∑ Tijσ<br />
2<br />
iσ<br />
k<br />
j −=<br />
1 I iσ<br />
, k=<br />
1<br />
k=<br />
1E<br />
−<br />
2∑<br />
∑ Tijσ<br />
iσ<br />
j −∑<br />
∑ I iσ<br />
i<br />
i,<br />
2 i<br />
i<br />
=<br />
1<br />
j<br />
j<br />
¹<br />
¹<br />
i<br />
i<br />
=<br />
1<br />
i<br />
i<br />
=<br />
1<br />
N N<br />
N<br />
M<br />
1<br />
M<br />
1 2 M<br />
M<br />
E = −<br />
M<br />
1 M<br />
M<br />
∑ ∑ Tijσ<br />
iσ<br />
j −<br />
Tij = −∑<br />
a∑<br />
I iσ<br />
i ,<br />
kia<br />
∑ ( σ ak ) 2<br />
kj<br />
i −1<br />
(12) .<br />
a2<br />
2<br />
∑ kiσ<br />
i + a<br />
i=<br />
1 j¹<br />
i=<br />
1T<br />
ij = −∑<br />
aiki<br />
= 1a<br />
2<br />
kj<br />
k=<br />
1 i −<br />
k=<br />
1<br />
2∑<br />
akiσ<br />
i + ∑<br />
aki<br />
ki<br />
k=<br />
1<br />
2 k<br />
k<br />
=<br />
1<br />
k<br />
k<br />
=<br />
1<br />
M<br />
MN<br />
N M N<br />
1 2<br />
T 1 1 ⎛ N<br />
ij = −∑<br />
akia<br />
kj and IE i = − ∑ a∑<br />
⎞<br />
( t+<br />
1)<br />
( )<br />
= 1 + 1 sgn<br />
⎛ N<br />
t ⎜<br />
⎟<br />
i<br />
.<br />
2 2 ⎜ ∑ Tijσ j + I<br />
⎞<br />
( t+<br />
kiσ<br />
T<br />
1)<br />
i + ijσ∑<br />
iσ<br />
ja−<br />
ki∑<br />
σ . I ak iσ<br />
i(13)<br />
,<br />
( t)<br />
k=<br />
1<br />
22<br />
ki<br />
=<br />
1<br />
j¹<br />
i=<br />
1 sgn⎜<br />
i⎟<br />
i = k+<br />
= 1 i=<br />
1<br />
⎟.<br />
2 2 ⎜ ∑ Tijσ j + I i<br />
⎝ j≠i<br />
= 1 ⎟<br />
⎝ j≠i<br />
= 1 ⎠<br />
M<br />
1 1 ⎛ N<br />
M<br />
M<br />
⎞ 1<br />
( t+<br />
1)<br />
( t+<br />
1)<br />
( t)<br />
2<br />
σ T σ ( t+<br />
1)<br />
σ sgn⎜<br />
⎟<br />
i = ij = −+<br />
∑ a<br />
i<br />
kia<br />
.<br />
2i<br />
2 ⎜ ∑ kj Tand ijσ<br />
+ I j Ii<br />
− i ⎟ ∑ akiσ<br />
i + (14) ∑ akiσ<br />
ak .<br />
k=<br />
1<br />
(t ) ⎝ 1 ⎠<br />
2<br />
j≠i<br />
=<br />
k=<br />
1<br />
k=<br />
1<br />
σ (t )<br />
σ j<br />
j<br />
( t+<br />
1)<br />
σ<br />
T<br />
ij<br />
1 1 ⎛ N<br />
⎞<br />
i<br />
ij<br />
( t+<br />
1)<br />
( t)<br />
sgn⎜<br />
⎟<br />
T i = +<br />
.<br />
σ a( x, T ) = ∫ Fh( x, x', z ', σa, T ) σ(<br />
x', (t ) z ') dx' dz '. (8) ( AA T<br />
σ<br />
( AA )<br />
ij 2 2 ⎜ ∑ Tijσ j + I i ⎟<br />
j<br />
ij<br />
⎝ j≠i<br />
= 1 ⎠<br />
T<br />
σ a( x, T ) = ∫ Fh( x, x', z ', ij σa, T ) σ(<br />
x', z ') dx' dz '. ( t+<br />
1)<br />
σ<br />
(8)<br />
a( x, T ) = ∫ Fh( x, xT', z ', σa, T ) σ(<br />
x', σ iz<br />
') dx' dz '. (8)<br />
F<br />
( AA ) ij<br />
h<br />
(t )<br />
) =<br />
σ ∫ F h F σ<br />
h( x, x', z ', σa, T ) σ(<br />
x', z ') dx' dz '. (8) j<br />
a<br />
F σ h( x, ax<br />
', z ', σa, T ) σ ( x', z ') dx' dz '. (8) T ij<br />
T<br />
σ j<br />
( AA ) ij<br />
TN , σ j<br />
ij (8)<br />
σ<br />
, M σ T a( x, T ) =<br />
j , j = 1,...,<br />
N,<br />
∫ Fh( x, x', z ', σa, T ) σ(<br />
x', z ') dx' dz '. (8)<br />
2<br />
ij<br />
ˆ σ k M N<br />
A)ij,<br />
ak , σ j , = j1=<br />
,..., 1,...,<br />
N,<br />
ˆ σ a , k 1,...,<br />
M.<br />
which in turn come from the integral equation.<br />
ak = ∑ kjσ<br />
j N =<br />
(9)<br />
2<br />
ˆ σ ak , k = 1,...,<br />
Mj=<br />
1 ˆ σ ak = ∑ akjσ j , k = 1,...,<br />
M.<br />
(9)<br />
j=<br />
1 N<br />
Strictly speaking, equation (14) applies only to binary<br />
σ ,<br />
ˆ σ ak = ∑ akjσ j , k = 1,...,<br />
M.<br />
variables (9) that can take values of zero and one, as in the HANN<br />
ak N<br />
j=<br />
1<br />
architecture (Hopfield, 1982). The corresponding equations for<br />
σˆ ˆ σ<br />
2<br />
ak = akσ<br />
∑ a<br />
M kjσ<br />
j , k = 1,...,<br />
M.<br />
(9)<br />
N<br />
M<br />
N<br />
1ak ,<br />
1<br />
an assemblage of such variables, to make up for arbitrary values<br />
j=<br />
1 ,..., M 2<br />
M<br />
N 2<br />
∑ = N a, ∑ ( σ ˆ ak 1− σ ak ) = ∑ 2[<br />
σ 1ak<br />
−∑<br />
akjσ j ] . 2 (10)<br />
kjσ<br />
j<br />
of conductivity, are very similar and are given in Appendix D.<br />
σˆ 2 C<br />
, k<br />
=<br />
= 1,...,<br />
(<br />
M<br />
σ<br />
.<br />
k=<br />
1 ∑ 2ˆ<br />
ak −σ<br />
) [<br />
] .<br />
2<br />
kak<br />
= σ<br />
(9)<br />
= 1 ∑ 2 j=<br />
1 ak −∑<br />
akjσ<br />
j (10)<br />
j=<br />
1 ak<br />
1 ,..., M<br />
k=<br />
1<br />
k=<br />
1<br />
j=<br />
1<br />
M<br />
M<br />
N<br />
1 2 1<br />
2 The application of equation (14) is straightforward<br />
ˆ<br />
N N CM= N ( σ N ) M [ M ] .<br />
1<br />
(9) N N∑<br />
M ak −σ<br />
ak = σ<br />
1 N∑<br />
ak − a<br />
M ∑ kjσ<br />
M j (10)<br />
when M>N, for the process reduces to the standard least<br />
2<br />
− ∑ C∑= −[<br />
1 ˆ<br />
−<br />
σ M<br />
ak = 2 k=<br />
1 ∑ akia kj −]<br />
σ iσ<br />
j −∑<br />
[ −<br />
2 k=<br />
1<br />
− ∑ −a<br />
1 2j<br />
= 1<br />
kiσ<br />
i + ∑ + akiσ<br />
ak ] σ i + squares + problem. This corresponds to the over constrained<br />
2 ∑ ∑∑<br />
akjσN j , k = 1,...,<br />
M.<br />
(9)<br />
2 1<br />
[ ∑ akia<br />
kj ] σ iσ<br />
2<br />
σ ˆ<br />
j<br />
i=<br />
1 j¹<br />
i=<br />
1 k=<br />
1<br />
i=<br />
1 2∑<br />
[ ∑ akiσ<br />
i ∑ akiσ<br />
ak ] σ<br />
ak − σ ak ) =<br />
i<br />
M2<br />
∑ [ j=<br />
σ1<br />
ak −∑<br />
akjσ j ] . (10)<br />
i=<br />
1 j¹<br />
i=<br />
1 N k=<br />
1<br />
i=<br />
k1=<br />
1 2 k=<br />
1 k=<br />
1 k=<br />
1<br />
2 1 2 k=<br />
1<br />
j=<br />
1 2<br />
case when we have more data than unknown parameters.<br />
ˆ σ ) [ N N M<br />
N M<br />
M<br />
ak = ∑ 1 σ ak −∑<br />
akjσ j ] . (10) 1 2<br />
2<br />
In general, however, M