Carbonate Log Interpretation Models Based on Machine Learning Techniques
- Wei Shao (Halliburton) | Songhua Chen (Halliburton) | Mahmoud Eid (Halliburton) | Gabor Hursan (Saudi Aramco RDD)
- Document ID
- Society of Petrophysicists and Well-Log Analysts
- SPWLA 60th Annual Logging Symposium, 15-19 June, The Woodlands, Texas, USA
- Publication Date
- Document Type
- Conference Paper
- 2019. held jointly by the Society of Petrophysicists and Well Log Analysts (SPWLA) and the submitting authors
- 1 in the last 30 days
- 391 since 2007
- Show more detail
We address several practical and common issues regarding applying machine learning (ML) methods for formation evaluation with logging data. It is normal that the available training data are far from “big” and require using more general ML algorithms such as neural network, thus the selection of ML algorithm becomes important. Second, laboratory core measurements are usually used as training data, but the applications are for logging data. The discrepancy between the two may involve instrument limitations, environment, and/or fluid states inside the pores. These differences should be taken into account to make the model work better. Furthermore, physical constraints can be applied for those petrophysical parameters that are intrinsically correlated but could be predicted optimistically with different ML algorithms for individual parameters to obtain self-consistent, robust petrophysical parameter sets.
Carbonate lithology is known to be complex resulting in highly heterogeneous pore systems. Primary and secondary pores commonly coexist at the same depth, and the various degrees of post-deposition diagenesis processes have progressed, which results in poor performances from many conventional carbonate log interpretation models. The traditional model development approach of forward modeling and inversion is fundamentally challenging for carbonates due to the difficulties in modeling the tool response to the pore system heterogeneities.
For Middle East carbonate reservoirs, pore typing, permeability, and pore throat size distribution are the key factors that are frequently used to assess reservoir quality, to select the perforation points, and for designing production strategies. Nuclear magnetic resonance (NMR) logging, in principle, may respond to all these parameters with different degrees of difficulty; empirical correlations have been used that work better for certain wells but not as well for others, again because of the range of heterogeneities which has not been addressed in some of the empirical equations.
To illustrate the significance of heterogeneity, the histogram of a simple, commonly used macro and micro pore cutoff values model is shown in Figure 1.
|File Size||3 MB||Number of Pages||13|