Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Research › peer-review
The use of the neural network for traveltimes approximation for inhomogeneous velocity models. / Grubas, S.; Loginov, G.; Duchkov, A.
81st EAGE Conference and Exhibition 2019. EAGE Publishing BV, 2019. (81st EAGE Conference and Exhibition 2019).Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › Research › peer-review
}
TY - GEN
T1 - The use of the neural network for traveltimes approximation for inhomogeneous velocity models
AU - Grubas, S.
AU - Loginov, G.
AU - Duchkov, A.
PY - 2019/6/3
Y1 - 2019/6/3
N2 - The proposed approach considers the calculation of traveltimes on a coarse grid followed by neural network training for interpolating these traveltimes on a fine grid. Using the neural network approximation has two advantages: it reduces computational burden for complicated models (when numerical eikonal solvers should be used for traveltime computation on a fine grid), it also reduces memory requirements (as compared to storing all traveltimes computed on the fine grid). We derived the neural network architecture with a single hidden layer and performed the numerical tests, including the application of the proposed approach to the microseismic data imaging. The numerical test showed that for laterally inhomogeneous velocity model (2D) a neural network with 100 neurons on hidden layer provides a mean absolute error of about 2.7 ms and for thin-layered inhomogeneous velocity model (1D) a neural network with 4 neurons on hidden layer provides a mean absolute error of about 1 ms. The achieved accuracy is enough for the imaging objectives. Besides, the proposed approach allows to speed up the imaging performance by 4 times (2D) and by 20 times (1D) and also significantly reduce the memory for storage.
AB - The proposed approach considers the calculation of traveltimes on a coarse grid followed by neural network training for interpolating these traveltimes on a fine grid. Using the neural network approximation has two advantages: it reduces computational burden for complicated models (when numerical eikonal solvers should be used for traveltime computation on a fine grid), it also reduces memory requirements (as compared to storing all traveltimes computed on the fine grid). We derived the neural network architecture with a single hidden layer and performed the numerical tests, including the application of the proposed approach to the microseismic data imaging. The numerical test showed that for laterally inhomogeneous velocity model (2D) a neural network with 100 neurons on hidden layer provides a mean absolute error of about 2.7 ms and for thin-layered inhomogeneous velocity model (1D) a neural network with 4 neurons on hidden layer provides a mean absolute error of about 1 ms. The achieved accuracy is enough for the imaging objectives. Besides, the proposed approach allows to speed up the imaging performance by 4 times (2D) and by 20 times (1D) and also significantly reduce the memory for storage.
UR - http://www.scopus.com/inward/record.url?scp=85084780369&partnerID=8YFLogxK
U2 - 10.3997/2214-4609.201901193
DO - 10.3997/2214-4609.201901193
M3 - Conference contribution
AN - SCOPUS:85084780369
T3 - 81st EAGE Conference and Exhibition 2019
BT - 81st EAGE Conference and Exhibition 2019
PB - EAGE Publishing BV
T2 - 81st EAGE Conference and Exhibition 2019
Y2 - 3 June 2019 through 6 June 2019
ER -
ID: 24310666