Quantitative gas saturation determination for reservoir monitoring purposes became possible with the introduction of a new generation of multi-detector pulsed neutron tools and interpretation algorithms. One distinctive feature of these interpretation algorithms is that they rely heavily on modeling of tool responses for the given completions and fluid types present in the system. This modeling is usually achieved through nuclear Monte Carlo simulations and involves long computing times, significant computer resources, and human intervention. However, despite the time and cost drawbacks of this approach, an associated benefit is the ever-growing library of models being computed for wells with different attributes. The existence of such Monte Carlo computed model libraries lends themselves to the deployment of machine learning to substitute the lengthy and expensive Monte Carlo-based model building process. As a result, the associated cost and time management cease to be an issue in the data acquisition planning and interpretation for gas saturation determination. Machine learning is a sub-branch of artificial intelligence, and encompasses a category of statistical algorithms that can 'learn' from existing data without explicit programming. These algorithms can be used to build models to predict the outcome for a given set of conditions. In this specific instance, the conditions are completion, formation, and fluid parameters. For example, borehole size, number of casing strings, presence of cement, annular fluid parameters, lithology, porosity and fluid types in the pore space are all needed to predict the response of an instrument designed for reservoir monitoring. The ratios of count rates from two detectors placed at two distances from the pulsed neutron source are typical outcomes from a Monte Carlo modeling exercise. The machine-learning activity is a substitute for this process, providing fast and accurate inelastic and thermal gate ratio values for gas saturation determination. Various machine-learning algorithms such as random forest and extreme gradient boosting were applied to the data to generate prediction models for the ratios mentioned above. Results showed that over 90% accuracy can be achieved between the predictions from the machine-learning models and the ratios calculated from the Monte Carlo simulations on a validation data set. The paper will first discuss the Monte Carlo-based model building and the existing model libraries used in quantitative gas saturation analysis along with the data processing methodology used to generate input data for the machine-learning algorithms. It will be followed by a discussion of various machine-learning models applied and their prediction accuracies along with variable values. Next, the trained machine-learning models will be deployed on blind test datasets (i.e., completion, lithology and formation parameter sets that the model has never encountered before), and the performance of the models on these completely new datasets will be demonstrated by comparing the predictions with those of the Monte Carlo–based models. Finally, the success of the trained machine-learning model will be demonstrated by deploying it on an actual gas saturation log, thereby showcasing the time and cost benefits of having data-driven models that can accurately predict inelastic and thermal gate ratio values.
展开▼
