Creating Synthetic Well Log Using Machine Learning Techniques

• Main Author: Herdiansyah, Rafli
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/40067
• Institution: Institut Teknologi Bandung

Petrophysics contains a large set of data and recognizing patterns such as well logs. Well logs are recorded in various types and measurements. Hence, these consume man power and time to build a reservoir characterization and modelling. When using wireline log to characterize formation properties of an area, sometimes it experiences into incomplete datasets because of missing data points. Another case, not all types of well logs properties is covered during running which results in unmeasured data sets that can be used in the future. One way to overcome these problems and minimize the task intensity is to create synthetic well logs. Machine learning is a computer science that contains algorithm to build model by recognizing the pattern. Therefore, it can create algorithm and model for predictive analysis. As a result, it can be reliable and cheaper alternatives than running new dataset. This paper presents two main ideas to obtain workflows of well log incomplete dataset prediction and new well log dataset prediction by using another existing well log dataset. Synthetic well logs were generated using single well for the first workflow. Besides, well log dataset from two well is used to predict new dataset that unmeasured in another well for the second workflow. During the analysis the datasets are divided into two parts, where 80 percent is to build the model, and 20 percent is used as validation set by using maximum accuracy sensitivity on the number of percentage training datasets in the range of 50 percent to 80 percent. The predictions are supported with algorithms by SVR and KNN methods. Then, the best machine learning is selected based on the least mean square error (MSE) and highest coefficient of determination (R2) to predict incomplete dataset and new dataset. Prediction results of each model are compared. The performance of SVR is outstanding because it can run effectively in high dimensional spaces. Although, there are some data that cannot be predicted properly with the aim of avoiding over-fitting. The performance of the KNN is also good. Because KNN is a simple machine learning that only used single hyperparameter, the accuracy depends only on the single hyperparameter. Therefore, KNN is less effective than SVR to be used on high dimensional space such as well log prediction that require many predictors. In predicting incomplete dataset, the accuracies of SVR and KNN are respectively 83.7% and 71.8%. On the other hand, the accuracies of SVR and KNN for creating a new dataset are respectively 96.5% and 90%. The result of two case studies reveal that SVR technique provides high accuracy in creating synthetic well log. The novelty of the synthetic well log is a simple, concise, and efficient workflow approach that involve the effect of percentage of training data and parameter sensitivity within machine learning to obtain optimum model. This workflow can also be used to predict various problems.