Well Logs and Machine Learning

The Well Log Visualization capability in Paradise® enables corroboration among machine learning classification results, seismic attributes, and traditional well logs. Seismic data is presented at borehole and lithofacies resolution in feet/meters. Use the Well Log Visualization capability to:

• Compare machine learning classification results and other seismic attributes to traditional well logs
• View well logs, formation tops, and extracted log curves together within a cross-section
• Customize the Well Log view by adding tracks and changing the properties of data

Well Log Visualization is a powerful tool to check and optimize machine learning results vs. reservoir properties computed from conventional well logs and traces extracted at well locations. In the example to the right, a Self-Organizing Map (SOM) – unsupervised ML – was used to distinguish shales from chalks.

The Well Log Cross Plot tool graphically displays the relationship between log curves and machine learning results, revealing pay intervals on the Well Cross Section depth track. Cross Plots highlight specific neurons in conjunction with log values most associated with production. Apply the Well Cross Plot tool to obtain these results quickly and easily:

• Bring seismic data resolution to borehole log and lithofacies scale in feet/meters via attributes, Machine Learning (ML) Stratigraphic Analysis using Self-Organizing Maps (SOMs), and ML Geobodies
• Extract seismic data along the borehole as attributes, SOM results, or probability results, and easily modify the associated color bars
• Display digital well logs, TD charts, formation tops, and cross-sections in straightforward displays

Apply the companion Well Log Viewer to analyze logs versus machine learning results, and identify which neurons correspond to reservoir petrophysical properties. Examine different ML topologies versus well logs to determine which topology is optimum. Then, convert the classes corresponding to production into ML geobodies.

A powerful technique for reservoir characterization is overlaying ML Stratigraphic Analysis with Fault Detection, as shown here in a region in the Caspian Sea, per Laudon et al. (2021).

On the right is a SOM that is sculpted on the top of the best section of the reservoir. Note the well is drilled in a small fault compartment. In the vertical view on the right, the well appears to have been drilled down a fault plane (shown in gray). The wells logs are VClay left and PHIE on the right. In the upper part of the well, the log character changes sharply, where the SOM indicates the well penetrated a large fault. In deeper sections of the well, the PHIE curves generally show lower porosity through lengthy fault sections, which are also slightly shalier. There is a significant amount of sand and porosity along the well path; however, the ‘ratty’ appearance in the logs suggests it has been drilled very close to a fault for much of the reservoir section. Source: Laudon, et al., (2021)

The bi-variate statistical analysis complements the Stratigraphic Analysis – Multi-Attribute Classification (SOM) – results with quantitative analysis based on well logs. Once the best SOM result is determined; then, this step uses Paradise to extract the neurons at well locations. Additionally, a category log with flags “Net_Reservoir” and “Non_Reservoir” is generated by applying petrophysical cutoffs to wireline logs within the zone of interest. Once the data are collected and organized, a bi-variate statistical analysis is conducted to assess the relationship between neurons and reservoir rocks. The Chi-squared test is applied to determine if a statistical correlation exists between neurons and petrophysical cutoff values. The histogram shown here is an example output of the process.