Analysis of the well data input and its suitability for the prediction process is achieved through built-in validation steps, enabled by the Data Validation Microservice, that expose specific deficiencies, if they exist, for each well.
Trained lithological models built by data scientists are provided by default within the Assisted Lithology Interpretation application to deliver rapid and consistent lithology interpretations. The models have been trained using supervised machine learning (ML) techniques, learning from a number of wireline or logging-while-drilling (LWD) data. Algorithms are encoded with intelligence to recognize combined features in well log curves and quantitatively assess the likelihood that these represent a particular lithology, based on previous examples seen by the system.
Prior probability calculations measure the similarity between the training data and the test data. When plotted together with the lithology prediction down a well, the prior probability score highlights those predictions that are likely to be of lower confidence, allowing interpreters to focus rapidly on areas of the well where further analysis may be required. The application provides the ability to set a prior probability cut off limit so interpreters only see the most confident predictions.
Many ML algorithms are probabilistic classifiers able to predict the likelihood that input data belongs to a given class. These posterior probability distributions can be output to provide additional information for geologists when validating or refining the initial prediction results. This measure appears as cumulative likelihood of a particular lithology.
Using trained lithology models, unseen wells with pertinent log curves can be processed and a predicted lithology provided, alongside confidence in that prediction, through prior and posterior probability calculations. Mapped features are classified into expected lithology categories, using the trained predictive model, and the resulting classification undergoes post-processing. This workflow is able to deliver rapid, detailed, and consistent lithology predictions for a well in less than five seconds.
The Automatic Model Selection mode is selected by default and permits the application to select the most appropriate model from the model based on the prior probability and the log curves present at any given well interval. Archive. Sixty-four models reside in the archive for the purposes of Automatic Model Selection. This intelligent mode enables a full-depth prediction to be presented to users, meaning lithology predictions will be performed irrespective of the log combinations present throughout the well(s) selected, if a suitable model is found. The minimum requirement for lithology to be predicted is the presence of a gamma-ray log.
The in-app Data Viewer offers the Well View for visualizing the predicted lithologies, prior probability, and posterior probability against the input wireline/LWD data. The Data Viewer also houses the Correlation View, which allows you to QC the lithology predictions for the wells side-by-side. The Correlation View also gives you the option to display the input well data if required.
The Curve Alias Mapping tool offers functionality for you to map curves from either OpenWorks® software or zip upload to the model(s) selected in the Assisted Lithology Interpretation application for lithology prediction. You are presented with the current state of your curve mappings where you have the option to edit the in-app curve dictionary as required and/or select a preferred curve where more than one option exists.
The ability to track and access previous runs allows users to monitor historical work and retrieve stored output.