Brittleness index prediction in shale gas reservoirs based on efficient network models

- Formation brittleness is an important parameter to describe the likelihood of complex fracture network during hydraulic fracturing.
- We presented two network techniques to predict brittleness by well logs and mineralogical data.
- Two neural network models were compared and back-propagation artificial neural network has the highest performance.
- Extreme learning machine model has a great advantage of processing time in brittleness evaluation.

Brittleness index is one of the critical geomechanical properties of unconventional reservoir rocks to screen effective hydraulic fracturing candidates. In petroleum engineering, brittleness index can be generally calculated from the mineralogical composition by X-ray diffraction (XRD) test or rock mechanical parameters by tri-axial experiments and well logs. However, mineral composition analysis or tri-axial experiments cannot produce continuous brittleness profile. Well log-based brittleness index prediction conventionally relies on Young's modulus and Poisson's ratio, but sometimes shear compressional velocity is not available to derive elastic inputs for the brittleness index calculation. This study proposes some data-driven practical brittleness prediction approaches based on back-propagation artificial neural network (BP-ANN), extreme learning machine (ELM) and linear regression using commonly available conventional logging data and lab mineralogical-derived brittleness. A dataset of 71 mineralogical-derived brittleness measurements from Silurian Longmaxi marine shale, Jiaoshiba Shale Gas Field, Sichuan Basin, China were established. The model comparisons and error analysis reveal that the application of artificial intelligence models can be more effectively applied to brittleness prediction compared with simple regression correlations. Both BP-ANN and ELM models are competent for brittleness prediction while BP-ANN model can produce slightly better brittleness prediction results with same inputs and ELM model require less running time. Thus, more choices can be made according to accuracy and computational speed demand. Moreover, an overall ranking of sensitivity degree is then provided to show the impacts of different well logs as inputs on the BP-ANN and ELM model, which is helpful to find optimal inputs in given case. Comparing to traditional well-log based brittleness approaches, data-based approaches show its wider applications because the integration of mineralogical composition and well log information can provide continuous brittleness profile in terms of high accuracy while acoustic full waveform velocities are no longer necessary inputs in brittleness evaluation.