Page 216 - Machine Learning for Subsurface Characterization

186 Machine learning for subsurface characterization

Few disadvantages and limitations of deep learning are as follows:
1. Not suited when the data size is in the order of thousand samples or smaller.
2. Not easy to interpret and explain due to the multiple levels of abstractions
without data augmentation techniques.
3. Slow and expensive to setup and train the model, which requires substantial
computation power.
4. Data preprocessing is required to avoid garbage-in/garbage-out scenarios
because deep learning methods can easily pick up spurious correlations
and biases.
5. Not suitable for implementations in limited/constrained memory devices.
6. For structured data, deep learning performance should be compared with
ensemble and nonlinear kernel-based machine learning methods for
structured data problems.
7. Models do not generalize well due to the tendency to overfit for
simple tasks.
Few advantages of deep learning are as follows:
1. Suited for complex problems based on unstructured data.
2. Best-in-class performance that beats traditional methods by a large margin.
3. Deep learning algorithms scale with data, that is, they continue to improve
as the size of data increases, while the performance of the traditional method
flattens out.
4. Scalable across various domains and easily adapted to new problems, unlike
domain specific methods such as those traditional ones specialized for
natural language processing and image analysis.
5. No need for feature engineering. Deep learning does not require domain
expertise because it does not need feature engineering. The focus is more
on data engineering.

1.3 NMR logging
Nuclear magnetic resonance (NMR) logging tool was introduced for well
logging applications in the 1980s. These instruments are typically low-field
NMR spectrometers. NMR tool is deployed in an open borehole to primarily
acquire the depth-wise T2 distribution response of the subsurface formation
volume intersected by a borehole. T2 distribution response is produced
because of the T2 relaxations of the excited hydrogen nuclei as they move
through the pore-filling fluid while interacting with the grain surfaces. T2
distribution is further processed to obtain the physical properties of the
formations, such as pore size distribution, fluid-filled porosity, bound fluid
saturations, and permeability, critical for developing and producing from
hydrocarbon- and water-bearing reservoirs. However, the acquisition of