Page 107 - Machine Learning for Subsurface Characterization

Shallow neural networks and classification methods Chapter 3 89

depositional texture with sufficient thickness result in good prediction perfor-
mance. Median values of the log do not affect the prediction performance; how-
ever, extreme values of coefficient of variation and skewness will have adverse
effects on the prediction performance of the ANN model.

4 Conclusions
Twelve conventional logs, ten inversion-derived logs, and five log-derived cat-
egorical flags in a shale system were processed by two distinct artificial neural
network (ANN) models with fully connected layers to synthesize the NMR T 2
distribution responses, which approximate the in situ fluid-filled pore size dis-
tribution. The first predictive model generates T 2 distribution discretized into 64
bin amplitudes, whereas the second predictive model generates 6 parameters of
a bimodal Gaussian distribution that fit the T 2 distributions. The first predictive
2
model is a better performing method exhibiting median R of 0.8549 on the test-
ing dataset. However, the first model takes 30% more computational time as
compared with the second model.
2
The median R and median NRMSE of predictions of T 2 distributions on the
testing dataset are 0.8549 and 0.1218, respectively. This testing performance is
remarkable given the hostile subsurface borehole conditions when acquiring the
logs, which result in low signal-to-noise ratio, and the limited size of the dataset
available to build the model, which gives rise to overfitting and poor general-
ization in the absence of cross-validation. All testing depths have prediction
2
performance higher than R of 0.6, and 90% of testing depths have prediction
errors lower than NRMSE of 0.2.
2
A few reservoir properties, such as ϕ N , T 2,gm , and T 2,gm ϕ N , were derived
from the synthetic T 2 distribution at reasonable accuracies. Therefore, ANN-
based predictions of NMR T 2 can be reliably used to estimate permeability
based on the SDR model. Complex pore size distribution caused by complex
grain size distribution and textures can impede the robust ANN-based synthesis
of NMR T 2 distribution. Moreover, ANN did not perform well for thin beds due
to the lack of data (i.e., statistically significant information) corresponding to
the thin beds, which hinders the accuracy of data-driven models. This study pro-
vides a workflow to generate in situ fluid-filled pore size distribution, approx-
imated as NMR T 2 distribution, in hydrocarbon-bearing shale reservoirs using
neural network models. Notably, the workflow uses features constructed using
the k-nearest neighbor classifier to significantly improve the predictive perfor-
mance of the ANN model. The proposed workflow holds value in the absence of
NMR logging tool due to financial and operational challenges.

Appendix A Statistical properties of conventional logs and
inversion-derived logs for various depth intervals

See Figs. 3.A1–3.A3.