A practical method for well log data classification
- PDF / 4,706,258 Bytes
- 11 Pages / 595.224 x 790.955 pts Page_size
- 43 Downloads / 209 Views
ORIGINAL PAPER
A practical method for well log data classification Wawrzyniec Kostorz1,2 Received: 27 March 2020 / Accepted: 29 September 2020 © The Author(s) 2020
Abstract In this work, a method for well log data classification is presented. The method relies on a coordinate transformation to restructure the data in an optimal way and a quasi-probabilistic interpolation technique capable of smoothing noisy data. The approach does not require case-specific design, is computationally efficient and provides a statistical characterization of the classification problem. Consequently, transition zones between facies can be modelled in a realistic fashion and intermediate rock types can be identified with ease. Apart from being capable of classifying unseen data with high accuracy, the technique can also be used as an informative quality and consistency assessment tool for manually classified data. The properties of the method are demonstrated on a realistic test case study. Keywords Classification · Well log · Smoothing · Nearest neighbour
1 Introduction Well log data classification is an important problem in the field of reservoir modelling. In recent years computer-aided classification has been gaining traction, with a particular interest in machine learning methods. Multiple authors investigated the applicability of machine learning/neural networks to the problem with varying degrees of success [1–6]. Additionally, in 2016 an open competition targeting machine learning methods was organized with over 300 entries from 40 teams according to the official report [7]. A variety of techniques were implemented by the contestants with the best performing ones including boosted trees, random forest, majority voting, multilayer perception, support vector machine, deep/convolutional neural network and nearest neighbours. Boosted trees approach dominated the top part of the leaderboard, with top eight scores all employing the method. A recent 1D-CNN model by Imamverdiyev and Sukhostat [8] deserves particular attention due to its relatively good performance. Wawrzyniec Kostorz
[email protected] 1
Geoscience Research Centre, Total E&P UK, Westhill, Aberdeen, AB32 6JZ, UK
2
Department of Earth Science and Engineering, Imperial College London, Prince Consort Road, London, SW7 2BP, UK
However, classifiers relying on machine learning (ML) suffer from a number of typical issues associated with this class of methods, two examples of which are provided. First, ML typically requires the model structure to be provided by the user and the model layout is generally designed via trial and error. Furthermore, an appropriate layout or fine tuning is generally case-specific and does not generalize easily. Consequently, a new model has to be devised for each individual problem. Second, the more general issue of the lack of explainability results in detecting errors and estimating model uncertainty particularly challenging. As a result, model prediction is hoped, rather than trusted, to be accurate. This combined with the case-de
Data Loading...
