Inversion of fracture density from field seismic velocities using artificial neural networks

Abstract

The inversion of fracture density from field measured P-and S-wave seismic velocities is performed using a neural network trained with an output from the modified displacement discontinuity fracture model. The basic idea is to use input-output pairs generated by the fracture model to train the neural network. Once the neural network is trained, inversion of fracture density from field-measured seismic velocities is performed very quickly. The overall performance of the neural network in the inversion process is assessed by means of a loss function. The results indicate that both sources of field information (P-and S-wave velocities) predict the field fracture density with reasonable accuracy. The performance of the neural network was compared to the prediction from least-squares fitting. It is shown that the neural network out performs the least-squares fitting in predicting the field-fracture density values.

Source rock characterization of the Albian Kazhdumi formation by integrating well logs and geochemical data in the Azadegan oilfield, Abadan plain, SW Iran

“…6. The aforementioned three-layered NN was trained by using the Levenberg Marquardt training algorithm (TrainLM), which details of their computation process and training can be found in Bishop (1995) and Boadu (1997Boadu ( , 1998. The performance of the model was measured by using mean squared error (MSE).…”

Neural Network Based Automated Interpretation Algorithm for Combined Geophysical Soundings in Coastal Zones

“…ANN had been widely used in various areas of geophysics, such as crossequalization of seismic time-lapse surveys (Ibrahim et al, 2002) identification of seismic crew noise (Buffenmyer et al, 2000), identification of seismic arrival types , seismic attribute calibration (Johnston, 1993), seismic inversion (Poulton et al, 1992., Roth and Tarantola, 1994., Boadu, 1998, fractured reservoir characterization (Ouenes, 2000), first arrival or arrival picking (Murat and Rudman, 1992;Mc Cormack et al, 1993;Mac Bath, 1995, 1997), seismic event classification (Dystart and Pulli, 1990), earthquake prediction (Feng et al, 1997), discrimination of genuine from spurious seismic events in mines (Finnie, 1999), as intelligence amplification tool (Poulton, 2002), locating buried objects (Brown and Poulton, 1996), borehole resistivity modeling (Zhang et al, 2002), lithology log estimation (McCormack, 1991;Rogers et al, 1992;Chen and Fang, 1993;, log interpretation (Pezeshk et al, 1996), permeability prediction from well logs (Rogers et al, 1992;Huang et al, 1996), reservoir parameter estimation (Aminzadeh et al, 2000;Lingireddy, 1998), reservoir characterization (An and Moon, 1993), prediction of transient water levels in subsurface (Coppola et al, 2003), run off prediction (Elshorbagy and Simonovic, 2000), characterization of aquifer properties (Rizzo and Dougherty, 1994), in modeling soil water retention curves (Schaap and Boulten, 1996), in simulating the rainfall-runoff (Abrahart and Kneale, 1997;Abrahart and See, 1988;Hsu et al, 1995;Smith and Eli, 1995)...…”

Artificial neural network and statistical models for predicting the basic geotechnical properties of soils from electrical measurements

“…Therefore, a model that is capable of resolving complex relations between electrical and soil parameters is required. Such comparative performances of neural networks and traditional regression models in predicting subsurface soil/rock properties from geophysical measurements have been investigated by several researchers (e.g., Ronen et al ; Boadu ; Boadu ; Leiphart and Hart ; Hampson et al ).…”