This report is dedicated to description of complex reservoir properties prediction method using neural-computing for stochastic seismic inversion results processing. Some case studies are provided.
It is very important for successful reservoir properties prediction using seismic and well log data to take all a priori geological and geophysical data into account correctly. Stochastic seismic inversion results (cubes of pressure and shear velocities, density, impedance) are defined all over the area of survey, fully correspond with seismic wavefield and with a priori geological and geophysical information too. Refined geological model possesses higher vertical resolution (due to reduction of reflections interference) and provide dimensional physical values unlike seismic amplitude. Application of neural-computer technology with integrated prediction stability check mechanism (Jack-Knife), provide reservoir properties prediction quickly and guarantee good quality.
Generalization of geophysical information is very important for now, especially for last time. It is because of wide application of attribute analysis in seismic interpretation as well as in other geophysical sciences. Usually if we have a number of different attributes not correlated with reservoir properties or with any log data it is very difficult to pick most significant for prediction or any combination of attributes that can be correlated with reservoir properties. Just for this purpose neural networks are widely used for a several years. But when using back-propagation networks with learning set, the most difficult problem is when to stop learning the network. Also a lot of questions is related to architecture selection and definition of learning process parameters . We can leave these problems to user, but better we provide some solution based on data current analysis. This work suggest solution of these problems using statistical analysis of data and experience from application of neural networks to real data.
Using seismic wavelet operator and convolutional model of seismic records forming, each time synthetic wave field for a prior model is calculated, it is compared to seismic.
Developed within interpretation software complex INPRES®, 3D seismic inversion software adjust medium's model using stochastic global optimization algorithm (simulation annealing algorithm [1,2,3]), and it allows to find a thin-layered acoustic model providing maximum similarity of synthetic and seismic fields, and also it is severely affected by prior geological and geophysical data. Optimizing is realized not for each trace, but for entire 3D model, using spatio-correlational datum structure. This approach provides geologically consistent solution and improves its vertical resolution.
Stochastic nature of the inversion algorithm permits evaluation of the problem solution uncertainty by generating a set of acoustic models, each of which is well consistent with available seismic and well data. In the proposed prediction technique, use is made of an averaged acoustic model that is computed from 5–10 particular inversions.
At first, Kohonen self-organizing neural network classification is applied to results of stochastic seismic inversion (fig.2). Main objective of this classification is generalization of inversion results for further prediction of reservoir properties, and also homogeneous zones distinction. Comparison of classification results with well log data may show correlation between some classes and reservoir properties degradation or enhancement.