Field Deployment of a LSTM Neural Network Tool for the Rock Formation Consolidation Inference of Brazilian Sandstone Reservoirs

Abstract

The objective of this work is to present a methodology based on the analysis of drilling parameters to infer if a reservoir formation is well consolidated or not, as a support to the selection of sand control strategies. This work proposes a statistical classification model and the usage of a memory based neural network, known as LSTM (long short-term memory) network. This model explores time series characteristics of the problem and it is validated using a cross strategy. Training performance is evaluated using F1-score, which is a metric that balances precision (percentage of true positives compared to false positives) and recall (percentage of true positives compared to false negatives), chosen because the dataset is unbalanced, there are more samples of one class than the other. The dataset consists of pre-tagged wells, each of them with at least nine hours of drilling data. Considering 48 cases from different drilled wells, the model was trained to learn how to tag between both patterns. The model analyzes 23 different drilling variables to reach a conclusion. After training the model, tests were performed and the results showed a high identification efficiency: around 90% of accuracy. That way, mechanical data analysis from the drilling process plays a very important role, supplementing that information and allowing a better understanding of formation behavior by employing what can be considered full-size and a real-time scratch test. Match the collected data with those from wells in which there is logging information, provides geomechanics calibration, and allows consistent rock profiling. It helps to define not only if there is a need for sand control but also the kind of technique to be applied to the analyzed formation accordingly to its consolidation state. The impact of that information is expressive to the completion process. This feature will be very useful in Brazilian post-salt wells that present sandstone as its reservoir rock formation. Also, as this tool was designed to run in a drilling digital twin, it can be automatically run as soon as the total depth is reached in the drilling phase, providing a fast insight to anticipate completion design. It is the first time in literature that this approach is used for this specific objective: define if a gravel pack or even any kind of sand control is indeed necessary to be installed based on information gathered while drilling the well. Its great results led this tool to the deployment phase. This work also aims to illustrate the first outcomes of that application in real-time decision-making.