GPU applications for modelling, migration, inversion and machine learning
Daniel O. Trad
There is a gap in best-practices between research in industry and academia. One of the weaknesses of research in academia is that tests are often carried out on very few data sets, missing the rich feedback provided by variety in testing. In industry, on the other hand, we are forced to work with different datasets. This data diversity brings a significant inflow of information that benefits research and development. In academia, it is hard to get new datasets and the necessary preprocessing makes it difficult and time-consuming to achieve testing diversity. Another difference is that industrial environments may have at least one or two orders of magnitude larger computer resources than academia. This report is about reducing these differences by creating an environment where students can, with limited resources, quickly model, migrate and invert seismic data, starting from known models or even simple pictures of models, that can easily be found in publications.
Modeling seismic data is a key part of research for acquisition design, imaging, full waveform inversion and machine learning. From the convolutional model to more sophisticated wave propagation methods with anisotropic visco-elastic 3D wave equations, there is a wide variety of approaches to simulate seismic data in different geophysical models. The more sophisticated the approximation, the more realistic the events we see on the simulated data. The same applies to inversion methods like migration and full waveform inversion (FWI). The efficiency and accuracy of modeling directly translate to better reverse time migration (RTM) and FWI. Although there are different options for wave propagation modeling, the finite difference method (FD) is most commonly applied because of its good trade-off between accuracy and efficiency. In this report, we discuss and illustrate how Graphics Processing Units (GPU) implementations for FD using a convolutional pattern can help in closing this research gap mentioned above.