Prediction of reservoir porosity distribution from seismic attributes using NEFPROX neuro-fuzzy model in the Gorgan Basin

Prediction of spatial distribution of porosity in a reservoir is an essential issue forestimating reserves and planning production operations. In most cases, however, lateralvariations of porosity cannot be delineated from measurements made at sparsely located wells. The integration of 3D seismic data with petrophysical measurements cansignificantly improve the spatial description of porosity. In the last two decades, severalmethods have been developed for the estimation of reservoir porosity. A number ofinversion methods are available in the industry to convert seismic amplitude into acousticimpedance. Acoustic impedance is indirectly related to porosity. Alternate integrativeapproaches for estimating porosity include geo-statistical methods, such as kriging andco-kriging using well and seismic data. One of the most accurate methods for estimatingreservoir parameters is the application of seismic attributes by a nonlinear estimator, suchas neural network or neuro-fuzzy model. Both neural networks and neuro-fuzzy modelscan be good estimators but the latter has the benefit of being interpretable. In this study, aneuro-fuzzy model called NEFPROX was used to estimate porosity in a gas reservoirlocated in the Gorgan Basin.NEFPROX is a Mamdani-type neuro-fuzzy model, so it has an advantage of beinginterpretable that makes it distinct from other type of neuro-fuzzy models. The timeconsumingcharacteristic of that method is irrelevant in this case because the prediction ofporosity is an offline prediction problem.The Gorgan Basin is located in the northern part of Iran, southeast of the Caspian Sea.This area consists mainly of three formations:the upper formation, called Clay-SandGroup , belongs to quaternary period. Below Clay-Sand Group  is a tertiary formationcalled Clay-Sand Group. A formation of Brown Beds is also a tertiary formation thatlies below Clay-Sand Group. All of these formations consist mainly of shale and sand.The discovery of gas in the Brown Beds Formation has persuaded explorationists toincrease their activities in the Gorgan Basin. The purpose of this study is to recognizeshale and sand bodies in the Brown Beds formation that consists of alternative sand andshale layers with variable thickness.First, a list of 20 seismic attributes was prepared to extract from raw seismic data inthe location of wells. Stepwise regression was used to select four appropriate attributes.The maximum number of attributes was set at four to avoid the model complexity. Theseattributes, in the order of priority, are instantaneous frequency, amplitude weightedfrequency, apparent polarity, and second derivative instantaneous amplitude. Then thesefour selected attributes were introduced into the neuro-fuzzy model as input to predictporosity as an output of the model.The neuro-fuzzy model was trained with the data of well GO3. Based on hydrocarboncore samples obtained from the Brown Beds formation and the potential of this formationas a probable reservoir, the data corresponding to this formation were selected as atraining data. A model blind test was also conducted with the data of well GO5.Porosity sections generated as the output of the model showed two low porosity sandyand shaly-sand channels in the Brown Beds formation. Lateral variations of thesechannels can clearly be recognized in these sections. The core samples available in wellGO3 (containing hydrocarbon) confirm the existence of the two inferred channels. Thisclear image of channels is simply unidentifiable from raw seismic data. Hence,NEFPROX can be very helpful in supplying valuable information about extent, shape andlithological variation of a reservoir. Finally, compared comparison was made between theperformance of the neuro-fuzzy model and regular neural networks in estimation ofporosity. The comparison indicates that the accuracy of the NEFPROX estimation isequal to that of MLP and is greater than that of RBF.