|Newell, A J, Woods, M A, Graham, R L, and Christodoulou, V. 2021. Derivation of lithofacies from geophysical logs: a review of methods from manual picking to machine learning. British Geological Survey Open Report, OR/21/006.
Contributor/editor: Kingdon, A
The aim of this report has been to evaluate a range of practical and accessible methods to derive lithology (or lithofacies) information from geophysical logs. All of the methods have their advantages and disadvantages (Table 6) and none will produce perfect results because of the inherent fuzziness of rock classifications and the convolved sampling signal of geophysical logs. Future work on fuzzy methods (such as Fuzzy-C-Means or Fuzzy-SVM) may be useful for this reason. Thus instead of predicting just one class as output they predict a fuzzy output i.e. a rock could belong to 40% of class A and 60% of class B etc. The method chosen is likely to depend on the quality and range of the available geophysical logs, user preferences and the scale of the project that is being considered (e.g. one borehole versus many hundreds). The best solution might be to combine a number of methods e.g. k- means to gain some initial insight into the data, classification by machine learning or cut-offs and finally some manual editing if required. An encouraging finding of the review is the large volume of high quality, open source software (mostly written in Python) that is now available for log handling, display and analysis. The widespread use of Jupyter notebooks makes this software accessible and easy to run.
Table 6. Table summarising some of the advantages and disadvantages of the different methods trialled.
Method Advantage Disadvantage Manual classification - Uses all the background knowledge and expertise of the geologist - Tools for manual digitising are built into most log handling and 3D geomodelling software - Doesn’t require high-quality log datasets or normalisation of data for multi-well projects - Can be undertaken on raster log images if no digital version is available - Slow manual process - Exact criteria for each classification decision can vary and may not be clear or reproducible by other users Cut-offs - Rapid, semi-automated method - Reproducible by any user applying the same cut-off criteria - Tools are built into most log handling and 3D geomodelling software - User has a strong degree of control on the outcome by adjusting cut-offs in different borehole regions - Can work effectively on a single log such as gamma-ray in simple siliciclastic formations - Often requires a degree of trial and error to find ideal cut-offs which slows the process - Becomes complex as the number of log types increase above two - Requires normalisation of log data if criteria are applied across multiple boreholes Mineral composition analysis by linear inversion - Rapid automated technique - Provides insight into the mineralogical composition of the formation - Useful in limestone-dolomite formations - Realistic in the sense that geophysical log measurements (with wide vertical sampling) are often ‘averages’ across different lithologies and mineralogies - Included in many log analysis packages or can be undertaken in spreadsheets and open-source software - Results must be post-processed into more conventional discrete lithofacies classifications - Requires high-quality geophysical log data, in particular a diverse suite of (relatively uncommon) porosity logs - Automated technique with little opportunity to include the users general geological knowledge of the formation in the outcome Unsupervised cluster analysis (K-means clustering) - Rapid automated technique - Reproducible by any user using the same input criteria - Deployable using open-source Python tools - Elbow plots provide unbiased insight into the likely number of facies classes that can be resolved from a particular set of log measurements - Requires a priori knowledge on the number of expected facies - No opportunity to include general user geological expertise or machine-learned support from cored intervals Supervised cluster analysis (Support Vector Machines) - Rapid automated technique, particularly once the SVM training process is complete - Easy to deploy using open-source Python tools - Incorporates geological expertise by using training examples of what constitutes the “correct” clustering of a dataset - Computes a range of metrics for the performance of the classifier based on the modelled versus the observed for the test data - Some unfamiliar terms and concepts for new starters in machine learning - Sophisticated but not infallible and (as shown by competitions against standard test datasets) will still generate lithology logs with many errors because of the inherent fuzziness of rock classifications and the convolved sampling signal of geophysical logs.