Creating Trends for Reservoir Modelling Using ANN Markus Lund Vevle*, Jon Magne Aagaard

Creating Trends for Reservoir Modelling Using ANN Trends help describe large scale reservoir characteristics There can be both a lateral and depth trend Trend analysis can be time consuming Linear regression is one way of extracting trends ANN can do regression modelling, how does it compare? Emerson Confidential

Outline Data used Algorithm Results Summary Emerson Confidential

Data used Gullfaks, offshore Norway: 151 wells in total 5 zones, with 49-118 per zone with data Average cell thickness ~2 m Amplitude seismic Maui, Taranaki basin, New Zealand: 5 wells with data (synthetic) Acoustic impedance Average cell thickness ~ 6.5 m Emerson Confidential

Algorithm ANN, structure and backpropagation algorithm in C++ SoftSign for activation function Root Mean Squared Error RMSProp for Mini Batch training Custom adaptive algorithm for epoch training. Network architecture restricted to equal sized hidden layers Training and verification data grouped according to K-fold cross validation methodology 3D output of one epoch, optimum epochs and from linear regression Emerson Confidential

Gullfaks - settings Input settings: Runtime graph showing Softsign activation function for hidden layers Linear activation function for output layer 3 hidden layers 3 neurons in each layer 10 cross validation groups Mini-Batch size 128 Blocked well data, porosity, and depth information Runtime graph showing Linear regression prediction error (green) NN prediction error (red) NN optimum prediction error (purple/violet) NN training error (black) Emerson Confidential

Gullfaks Tarbert – Well data vs trend output Results Both wells used for training Subtle difference between NN (blue) and Linear reg (black)

Gullfaks Tarbert – Well data vs trend output Results Both log tracks are the same well Left figure when well was used for training Right figure when well was left out

Gullfaks Tarbert – Well data vs trend output Results More variability in well data → potential for overfitting NN (blue) captures better the decreasing trend in the bottom (left figure)

Gullfaks Tarbert – Well data Results Apparent trend first increasing porosity, then decreasing Emerson Confidential

Gullfaks Tarbert – Well data and Linear regression Results Lower part of the trend is captured Emerson Confidential

Gullfaks Tarbert – Well data and ANN (one epoch) Results Both upper and lower part of the trend is captured Emerson Confidential

Gullfaks Tarbert – Well data and ANN (one epoch + optimum) Results Both upper and lower part of the trend is captured Emerson Confidential

Gullfaks Tarbert – Output in 3D Results Blue dots – Wells Input data centered around the middle Least knowledge about the east flank

Gullfaks Tarbert – Output in 3D Results

Gullfaks Tarbert – Cross section west-east Results Subtle differences where there is well control

Gullfaks – other zones Results Is it always necessary? Often seen prediction error from linear regression vs NN is very small ~ 1%

Gullfaks – other zones Results Sometimes you get strange effects

Gullfaks – other zones Results

Gullfaks – other zones Results

Maui Results 5 wells with synthetic data AI parameter

Maui Results AI ANN output trend

Maui Results

Maui Results

Maui Results Overall trend captured

Maui Results More details are captured. Does not try to fit the biggest outliers.

Summary and discussion ANN can be used for extracting trends With only well data Smooth output A combination of well data and spatially distributed data, for better lateral understanding More details emerge Overfitting was not observed (?) Runtime is from 10’s of seconds to a couple of minutes More data means slower Defining the network takes time Is the settings we used valid across many fields? Data is needed