Direct Petrophysical Inversion from pre-stack seismic data using deep learning

C. L. Lew, C. D. MacBeth, A. ElSheikh, M. S. Jaya, M. I. Ahmad Fuad

Research output: Contribution to conferencePaperpeer-review


A method to directly invert for porosity, Vclay and hydrocarbon saturation (Shc) simultaneously from pre-stack seismic data using deep learning has been discussed. We implemented UNet architecture as the network solution where the encoder part of the UNet is incorporated with ResNet-18. The inputs for the network are the seismic angle stacks whereas the outputs are the petrophysical properties of interest. We introduced L1 norm in the loss function which can help to promote sparsity for the estimated Shc. To increase the variability of possible scenarios, the process of generating the synthetic dataset (pairing of synthetic seismograms and petrophysical properties of interest) for training involves the application of 1) three different variogram model ranges, 2) four types of suitable rock physics models, 3) oil and gas cases for the hydrocarbon fluid types, and 4) nine different sets of possible angle dependent source wavelets. The initial generated synthetic seismograms are noiseless. We then prepare second synthetic dataset by adding noise into the previously generated synthetic seismograms. Two network model (ML1 and ML2) are trained on these two synthetic datasets respectively, then applied to the modelled test data and field dataset. ML1 which is trained on the noiseless synthetic dataset initially provides better results when applied to modelled test data. However, through the validation on the field data, it is shown that ML2 trained on the synthetic data with added noise gives the best performance.
Original languageEnglish
Number of pages5
Publication statusPublished - 5 Jun 2023
Event84th EAGE Annual Conference & Exhibition 2023 - Vienna, Austria
Duration: 5 Jun 20238 Jun 2023


Conference84th EAGE Annual Conference & Exhibition 2023


Dive into the research topics of 'Direct Petrophysical Inversion from pre-stack seismic data using deep learning'. Together they form a unique fingerprint.

Cite this