Using IWAVE

Discussion and Conclusion

The rather large and only slowly disappearing error revealed by the examples from Symes and Vdovina (2009) suggests strong limits for the accuracy of regular grid finite difference methods. Finite element methods suffer from the same limitations: accurate solution of acoustodynamic or elastodynamic problems appears to demand interface-fitted meshed (Cohen, 2002), with the attendant increase in code and computational complexity.

The situation may not be so bleak, however. For one special case, namely constant density acoustics, Terentyev and Symes (2009) show that a regular grid finite difference method, derived from a regular grid Galerkin finite element method, has accuracy properties one would expect in homogeneous media (second order convergence, reduction of grid dispersion through higher order space differencing) even for discontinuous models: the interface error effect is attenuated. This type of result actually goes quite far back in computational geophysics (see for example Muir et al. (1992)), though theoretical support has been slower in coming.

Pure regular grid methods cannot take advantage of changes in average velocity across the model, and concommitant changes in wavelength. Coupling of local regular grids is possible, however, and can yield substantial computational efficiency through grid coarsening in higher velocity zones - see Moczo et al. (2006). IWAVE already accommodates multiple grids (in domain decomposition parallelism), and extension to incommensurable multiple grids would be a significant change, but in principle straightforward. The use of logically rectangular but geometrically irregular (``stretched'') grids is completely straightforward, on the other hand.

These and other extensions, both past and future, are eased by the reusability designed into the IWAVE core framework. This design has produced reasonably well-performing and easy-to-use applications, and has proven extensible to new models and schemes. Moreover, as explained by Symes et al. (2011), the object-oriented design of IWAVE dovetails with similarly designed optimization software to support the construction of waveform inversion software. The inversion applications resulting from this marriage inherit the features of IWAVE - parallel execution, high-order stencils, efficient boundary conditions, simple job control - without requiring that these aspects be reworked in the code extensions.

The IWAVE acoustic application supports many use cases beyond those of the scripted examples, such as various modes of parallel execution, array sources, movie output, 3D modeling, and many others described in the documentation. It is hoped that the brief overview above, the detailed description of the example parameters given in Appendix A, and the much more extensive description of use cases in (Terentyev et al., 2012) will enable the reader to constuct a wide variety of synthetic data sets with relative ease.

Using IWAVE