Page 222 - Machine Learning for Subsurface Characterization

192 Machine learning for subsurface characterization

3.2 Data preparation
Data preparation, or data preprocessing, affects the convergence time of neural
networks during the training phase. Each input has a different range; for
3
3
example, illite volume fraction ranges from 0.1 to 0.5 m /m , whereas water
3
3
and oil volume fraction is below 0.1 m /m . Each input data is scaled to
have zero mean and unit variance, which is also referred as the standard
scaler. Such scaling is suitable when there are only a few outliers in the data.
As the number of outliers increase, it is advisable to use robust scaling.
Output/target data generally do not require either scaling or normalization.
Outliers in targets were screened and removed by fitting NMR T2
distribution, discretized into 64 T2 bins, with a cubic spline and then
removing those with three peaks and unusual characteristics when fitting the
NMR T2. Depth shift was performed to ensure that the 10 inversion-derived
logs and the corresponding NMR T2 distribution correspond to the same
depth. This ensures that the deep neural networks are well trained to relate
10 features (inversion-derived logs) with 64 targets (NMR T2 distribution
discretized into 64 T2 bins).
4 Neural network architectures for the NMR T2 synthesis

4.1 Introduction to NMR T2 synthesis
We implement four deep neural network architectures that can learn to process
the seven mineral volume content logs, namely, illite, chlorite, quartz, calcite,
dolomite, anhydrite, and kerogen, and three fluid saturation logs, namely, bound
water, free water, and oil, to synthesize NMR T2-distribution responses along a
well length. These 10 logs will be referred as the inversion-derived logs. The
first half of the VAE-NN, GAN-NN, and VAEc-NN architectures learn to
abstract and understand the concept of NMR T2 distribution. These three
architectures first train deep neural networks to accurately reproduce the
NMR T2 distribution in the training dataset by extracting certain
representative features of the NMR T2 distribution; following that, a simple
multilayer neural network is connected with the pretrained deep neural
network to reliably synthesize the NMR T2 distribution from the 10
inversion-derived logs. Instead of individually predicting T2 amplitude for
each of the 64 T2 bins of the NMR T2 distribution, we implement deep
networks to simultaneously predict T2 amplitudes for the 64 T2 bins to
generate the complete NMR T2 spectra, thereby providing greater constraint

FIG. 7.1, CONT’D formation tops or lithology indicators provided by the geology expert. Tracks
1,2,3,and4aretheconventionallogsonwhichinversionisperformedtoobtaintheinversion-derived
logs in Track 5. CMR_ADT in Track 6 is NMR T2 distribution response that is the target for the data-
driven modeling. NMR T2 distribution response at a specific depth comprises 64 T2 amplitudes as a
function of T2 times. Deep neural networks were trained to relate the 10 inversion-derived logs to the
entire NMR T2 spectra, comprising 64 T2 amplitudes as a function of T2 times.