Physics-based, deep learning model improves screen-out predictions

By Stephen Whitfield, Associate Editor

Screen-out usually occurs when proppants carried in fracturing fluid create a bridge across the perforations in a fracture or a similar restricted flow area, creating a sudden and significant restriction to fluid flow that causes a rapid rise in pump pressure. This condition can cause significant disruption to a hydraulic fracturing operation, delaying the placement of subsequent stages and wellbore cleanout operations, and ultimately leading to lost production days. 

Because of that potential impact, advanced warning of screen-out is critical to improving operational safety and efficiency, according to Dr John Sun, Senior Data Scientist at Noble Energy. Speaking at the 2020 Unconventional Resources Technology Conference (URTeC) in July, Dr Sun discussed how his company has evaluated the use of machine learning in combination with existing methods  to improve screen-out predictions during frac treatments. 

The methods currently used to predict screen-out often rely on either post-minifrac diagnostics or pure physics-based modeling approaches, which Dr Sun said are not always appropriate for real-time application. A number of physics-based approaches have been developed to predict screen-out to some degree, such as the inverse slope method, which involves the visual inspection of the slopes of surface pressure plots during frac treatments. While these approaches can be effective, they leave a lot of room for improvement in terms of accuracy. For instance, the inverse slope method can produce false positives and false negatives, making it difficult for operators to make decisions in real time, Dr Sun said. 

“We see these false positives, in part, because the inverse slope model is very sensitive to changes in surface pressure, and it’s only visually inspecting the pressure alone,” Dr Sun said. “The model ignores other changes in the formation, such as proppant concentration and the flow rate, that can affect surface pressure.” 

Deep neural network-based approaches have been useful for predicting screen-outs, especially in terms of anomaly detection. The model Dr Sun discussed was a combination of physics-based inverse slope model and CNN-LSTM (convolutional neural network-long short-term memory) network architecture. It is a type of deep learning system used to generate textual descriptions of images to help reduce false positives and false negatives. Dr Sun said combining the two modeling methods can help provide more accurate advance warnings of screen-out from real-time pumping data.

Noble utilized three predictive models as part of its test of the ensemble model – a physics-based inverse slope model, a baseline LSTM model and a CNN-LSTM model – using input datasets from different formations and geological areas in the Niobrara-DJ Basin. Results from each model were validated against one another, and the input datasets were then run in an ensemble model that incorporated elements of the other models to establish whether that ensemble model could make more accurate prediction of screen-outs than any of the models could individually. 

Noble researchers split the field into smaller regions, then selected wells from each region that best represented the geology of the area. Ultimately, they selected a sample of 65 screen-out stages to examine for the model, as well as normal stages from the same wells. In running the machine learning models, Noble divided the stages into three datasets: one for training the models, one for validating the models and one for testing the models. 

Researchers normalized surface pressure against the flow rate of the fracturing fluid to remove pressure changes due to flow rate variation, divided this normalized surface pressure into 1,280-second sequences and mapped those sequences out in an inverse slope model. 

From there, the slopes represented in each timestep were analyzed to determine if and when screen-out has been triggered. Negative surface pressure slopes occur when the hydrostatic pressure in the wellbore increases gradually with the addition of pumping proppant slurry. Once the slurry reaches the perforation, surface pressure stops declining – Dr Sun said the slope may remain constant or even increase gradually. The rate of pressure increase should be roughly equivalent to the rate of pressure decrease. If that pressure increase starts to deviate, it is a sign that the treatment is screening out at the near-wellbore area. 

The test showed that the inverse slope model’s predictions aligned well with observations from the field, indicating its capability in identifying pressure spikes and providing warning messages in real time. However, two of the three pressure spikes measured in the model were false positives.

Dr Sun said the neural network incorporated a wider range of features than the modified inverse slope model. Instead of only using normalized surface treatment pressure, the neural network incorporated engineering features like cumulative flow rate, cumulative proppant volume, surface pressure divided by flow rate, geological areas and formation types. These features helped the model extract information about how much fracturing fluid and proppant are injected. The CNN-LSTM model consisted of an additional two groups of feature extraction layers.

“The difference between a deep learning model and the inverse slope model is that the deep learning model uses a lot more parameters to make its predictions,” Dr Sun said. “It’s much easier for a deep learning model to predict a number of situations, like when the proppant concentration is very low. We’re not going to see the false positives that pop up in the inverse slope model.” 

The machine learning models did not show false positives, but the screen-out predictions were still a mixed bag. The CNN-LSTM model showed noticeable improvement over the baseline LSTM model in the validation dataset; true positive rate of screen-out prediction improved from 57% for the baseline model to 74% for the LSTM, and true negative rate increased from 89% to 94%. However, it did not perform as well with the test dataset, with the true positive rate dropping from 86% to 76% and the true negative rate dropping from 100% to 90%. The model also predicted screen-out earlier than the field observations, while the inverse slope model’s prediction was fairly accurate when subtracting the false positives. 

Noble utilized weighted averages of the inverse slope model and the CNN-LSTM model to build its ensemble model, which outperformed the other models in terms of true positive rate, true negative rate and false positives. The true positive rate of screen-out prediction from the test dataset increased to 95%, and the true negative rate of increased to 96%.