Skip to main content
U.S. Department of Energy
Office of Scientific and Technical Information

Seismic interpretation using Support Vector Machines implemented on Graphics Processing Units

Conference ·
OSTI ID:896565

Support Vector Machines (SVMs) estimate lithologic properties of rock formations from seismic data by interpolating between known models using synthetically generated model/data pairs. SVMs are related to kriging and radial basis function neural networks. In our study, we train an SVM to approximate an inverse to the Zoeppritz equations. Training models are sampled from distributions constructed from well-log statistics. Training data is computed via a physically realistic forward modeling algorithm. In our experiments, each training data vector is a set of seismic traces similar to a 2-d image. The SVM returns a model given by a weighted comparison of the new data to each training data vector. The method of comparison is given by a kernel function which implicitly transforms data into a high-dimensional feature space and performs a dot-product. The feature space of a Gaussian kernel is made up of sines and cosines and so is appropriate for band-limited seismic problems. Training an SVM involves estimating a set of weights from the training model/data pairs. It is designed to be an easy problem; at worst it is a quadratic programming problem on the order of the size of the training set. By implementing the slowest part of our SVM algorithm on a graphics processing unit (GPU), we improve the speed of the algorithm by two orders of magnitude. Our SVM/GPU combination achieves results that are similar to those of conventional iterative inversion in fractions of the time.

Research Organization:
Lawrence Livermore National Laboratory (LLNL), Livermore, CA
Sponsoring Organization:
USDOE
DOE Contract Number:
W-7405-ENG-48
OSTI ID:
896565
Report Number(s):
UCRL-CONF-222359
Country of Publication:
United States
Language:
English

Similar Records

Randomized Sampling for Large Data Applications of SVM
Conference · Sat Dec 31 23:00:00 EST 2011 · OSTI ID:1059336

Approximate l-Fold Cross-Validation with Least Squares SVM and Kernel Ridge Regression
Conference · Thu Apr 10 00:00:00 EDT 2014 · 2013 12TH INTERNATIONAL CONFERENCE ON MACHINE LEARNING AND APPLICATIONS (ICMLA 2013), VOL 1 · OSTI ID:1567338

A GPU-COMPUTING APPROACH TO SOLAR STOKES PROFILE INVERSION
Journal Article · Thu Sep 20 00:00:00 EDT 2012 · Astrophysical Journal · OSTI ID:22092321