Page 80 - Machine Learning for Subsurface Characterization

66 Machine learning for subsurface characterization

Appendix B Categorization of Appendix D Estimations of specific
depths using flags (categorical reservoir parameters from NMR T 2
features) 96 distributions 98
Appendix C Importance of the 12 References 100
conventional logs and 10
inversion-derived logs 97
1 Introduction

NMR log responses acquired in the subsurface are inverted/processed to generate
T 2 distribution specific to each depth. In geological formations, T 2 distribution is
the transverse relaxation time of hydrogen nuclei of the fluids in the pores. T 2 dis-
tribution is represented as a spectra of T 2 amplitudes measured across 64 T 2 time
bins. T 2 distribution is a function of the fluid-filled pore volume, fluid phase dis-
tribution, and fluid mobility in the pores. NMR T 2 distribution approximates the
fluid-filled pore size distribution. T 2 distribution concentrated around small T 2
times are primarily due to small-sized pores. Unlike the conventional logging
tools,suchasgamma ray,density,neutronporosity, and resistivity,the operational
and financial challenges in deploying the NMR logging tool and computing the
NMR T 2 distributions impede its use in most of the wells. Deployment of
NMR tool is a more severe challenge in shale reservoirs. Well conditions, such
as the lateral well section, small-diameter boreholes in deep HPHT reservoirs,
and boreholes with large washouts common in carbonates, limit the use of the
NMR logging tool. The objective of this chapter is to synthetically generate
NMR T 2 distribution from conventional easy-to-acquire logs, so that NMR T 2 dis-
tribution can be generated in the wells where NMR log is not available due to well
conditions, financial limitations, and operational constraints.
We developed two artificial neural network (ANN)-based predictive models
that process the conventional logs to generate the NMR T 2 distribution. The first
predictive model implements a generic ANN with fully connected layers that
generates T 2 distribution discretized into 64 T 2 bins, identifying relaxation times
in the range of 0.3–3000 ms logarithmically split into 64 parts. The second
ANN-based predictive model implements two steps: First, the T 2 distribution
is fitted with a bimodal Gaussian distribution characterized by six parameters,
namely, two amplitudes, two variances, and two T 2 locations of peak amplitude;
subsequently, a generic ANN model with fully connected layers is implemented
to generate the six parameters for the bimodal T 2 distribution, which are
later invoked to generate the T 2 distribution for any depth in the formation.
The second approach is based on the observation that the T 2 distribution
responses have mostly either unimodal or bimodal distributions.
ANNs have been applied to log-based subsurface characterization. Bhatt
and Helle [1] predicted porosity and permeability for wells in the North Sea
by processing well logs using committee neural networks. For the porosity pre-
diction, ANN processed sonic, density, and resistivity logs. For permeability
prediction, ANN processed density, gamma ray, neutron porosity, and sonic
logs. Al-Bulushi et al. [2] predicted water saturation in the Haradh sandstone