Now suppose I know a set of n Green's function traveltime fields for n source positions . Estimating the traveltime field , due to a single intermediate surface source position , from the n given fields , is defined as traveltime interpolation.
For the set of , (7) can be generalized to a linear matrix equation of the form:
(12) |
or into a general weighted least-squares system :
(13) |
System (13) can be solved by standard damped least-squares,
(14) |
or by the optimal (but slow) Singular Value Decomposition (SVD) method, for the components of the unknown vector .I refer the reader to Strang (1980) for an excellent review of damped least squares and SVD linear algebra techniques. The diagonal weighting matrix is of dimensions (nxn), and the weights wi could be inverse-distance: , for example. The matrix is (nx3) in 3-D and (nx2) in 2-D (ignoring the terms), and contains the vector components of the known traveltime gradient fields. The data vector is (nx1) and contains the values. For a 2-D geometry, at least two traveltime gradient fields must be known, and for a 3-D interpolation at least three traveltime gradient fields must be available.
In the case of interpolation, a refinement to the estimate of the values is possible. In this paper, I use the Law of Cosines given by (8) as an initial estimate of, say, and , where the traveltime field at is to be interpolated from two bounding source locations and . Please refer to Figure to consider the appropriate interpolation geometry. However, the angle is known precisely from the given traveltime (gradient) fields at and :
(15) |
Therefore, given the angle estimates , and , and the true angle , I can refine my initial (primed) estimates such that the total angle is conserved as a sum of , :
(16) |
Then, the correction factor f is given as
and
(17) |
As an example, for a 2-D interpolation of given and , I first set up a 2x2 matrix system per (13). I evaluate the required data vector values and using (8), and refine the cosine estimates using (17). I then invert system (13) by Cramer's Rule if the determinant is not too small, or by SVD if otherwise. The entire procedure is repeated in a point-by-point independent manner, making it an obvious candidate for massively parallel implementation, for all subsurface grid locations .
Finally, the traveltime gradients are spatially integrated to traveltime such that, for a source at (x2,z2) on a flat surface z = z2,
(18) |
(19) |
where H is the Heaviside function. This process completes the traveltime interpolation of from and .