2:00 PM / Thursday, February 15, 1996
Invited Presentation 7
Chavent and others introduced the adjoint state method of gradient computation into computational inverse problems in the 1970's. It is the only feasible means to form the gradient of typical objective functions arising in seismic inversion, for instance, when the wavefields are computed by finite difference or finite element solutions of boundary value problems. It is extremely efficient for conservative problems and, when augmented with Griewank's checkpointing scheme, for certain dissipative problems. The speaker will review the use of the method in a variety of more and less standard contexts. He will also describe a flexible approach to mostly automatic implementation of efficient adjoint state schemes, making low-intensity use of AD tools, as part of an evolving programming environment for numerical work on inverse and control problems of PDEs.
Adjoint Computations for Wave Propagation
William W. Symes