Automatic Interpretation of Well Logs with Lithology-Specific Deep-Learning Methods
- Aymeric-Pierre Peyret (The University of Texas at Austin) | Joaquín Ambía (The University of Texas at Austin) | Carlos Torres-Verdín (The University of Texas at Austin) | Joachim Strobel (Wintershall GmbH)
- 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
- 2 in the last 30 days
- 258 since 2007
- Show more detail
Accurate and reliable interpretation of well logs often requires a high level of expertise from a petrophysicist along with enough relevant borehole and core measurements. As an alternative and complementary approach, deep learning has been proposed to capture the experiential knowledge gained from petrophysical interpretations, as well as the physical and heuristic models often used for that purpose. We test the latter idea using a set of wells previously interpreted and explore the critical aspects that can yield a successful automatic well log interpretation.
Some of the questions we attempt to answer here, as a guide for future applications of artificial neural networks (ANNs), are: how much data does the petrophysicist need to explicitly interpret before relying on ANNs? What is the best suited deep-learning network architecture? How widely can the ANNs be generalized? Can we automatically classify different wells for improved ANN usage? In this study, all the wells used come from the same hydrocarbon reservoir and intersect multiple formations. The focus is on estimating clay fraction (VCL), effective porosity (PHIE), water saturation (SW), and permeability (K).
We compare the performance of various architectures of deep artificial neural networks (ANNs) using different numbers of layers and neurons in each layer. Once an ideal ANN architecture was found for a specific formation, it was tested against different formations, but results were relatively poor, corroborating the specificity of ANNs to the lithology where they were trained. Furthermore, we propose a self-organizing map (SOM) as a way of partitioning wells into classes, which should be treated separately (independent ANNs). This strategy yielded the best results.
The amount of data required to train our ANNs was relatively small (5-9 wells), considering the amount of data typically required for training more general ANNs. This is only possible if different lithologies and/or rock classes are treated independently.
The use of artificial intelligence (AI) in well-log interpretation has been proposed multiple times, and for different purposes. Classification of facies within well trajectories has been done by relying on well logs, and using ANNs (Saumen, et al., 2007; Rogers et al., 1992), modular neural networks (Bhatt, and Helle, 2002(1)), deep convolutional neural networks (CNN) (Imamverdiyev, and Sukhostat, 2019), and SOMs (Fung, et al., 1995).
|File Size||2 MB||Number of Pages||20|