Ask any technical question and we will get you answer from experienced technical person
Talk to our recruitment expert for recruitment services.
Well-log correlation is one of the most important step in subsurface reservoir characterization. Numerous algorithms have been developed to automate well to well correlation. In this article we would discuss a novel approach of doing well to well correlation through machine learning between well logs and standard stratigraphic correlation techniques.
There have been numerous attempts to automate well-log correlations since early 1970s, these early approaches mainly relied on auto and cross-correlation techniques, dynamic time warping functions and even neural networks. All these techniques achieved limited success in complex data situations as a result most of the geologist instead of using these techniques likes to rely on manual correlation approach which is often time consuming. One major issue of going for manual approach is reduced ability to input and interpret all available data for improving confidence in data. On the other hand challenges in automated correlation are problems related to poor data quality, missing data and highly heterogeneous geological section. Automated approach to log correlation may require normalized data across the field which means software might not consider variability in geology. However in complex geological scenario high amount of data with lots of data variability might be helpful. Hence automated log correlation tool should be able to accommodate high data variability and geological changes.
As discussed above log correlation in complex geology is challenging in automated workflows however if large amount of data is available for training, machine learning can do a good job of well log data correlation.
To tackle challenges arising due to geological complexity it is suggested to use supervised deep learning approach in artificial neural network for pattern recognization between petrophysical logs of different well bores. Supervised learning in artificial neural network approach employs several layers of non-linear filters for generating multi-dimensional data in order to classify patterns, trends and structures as per training data.
In automated workflow deep learning neural network model is also constrained by different correlation techniques and established geological approaches for correlation. These constraints help by reducing false correlations. Also network should comply with basic rules such as formation top does not cross, automated picks must follow structural and isopach trends defined by geologist etc.
First step of automated well log correlation begins with geologist correlating few wells out of all the wells from database to be correlated; these initial few wells are also called seed wells. Deep learning neural network does not need to be re-trained every time but this initial picks provide example of tops to be picked by network. These seed wells should have enough heterogeneity to properly characterize geological complexity within the field for reliable correlation. Log data selected for purpose should be available in all the wells of database.
Well log data from seed well is trained to recognize similar features between different well logs rather than formations. This helps in applying this training data to wide range of geology. Correlations done produces a match score between 0 to 1. Better the match closer the value to 1.Training using this method needs large dataset, which might need Cloud capabilities.
For proper application of machine learning, whole of the training data set needs to classified into set of correlating and non-correlating sample pairs as per goodness of match. As training data set is large, this classification creates robust model. Classification of training dataset should also consider applicable geological factors as constraints input by analyst one critical example of such geological constraint is constraints applied to mitigate numerous positive correlations due to cyclical depositional phenomenon. Major critical aspect of this technique is validation of dataset. It is recommended to keep 80% of input data as training data and rest of the 20% for validation of results.
This can result into robust deep learning neural network model which can be used to automatically correlate huge amount of well log data across the field with proper geological constraints. If applied properly created model works with extremely high amount of accuracy.
Once automated correlation is completed and validated, Analyst generated formation tops can be populated from seed wells to new wells correlated.
As huge dataset is trained, analyst can always look at locations of wells and curve responses to select small amount of training data closely matching with geology of area to be correlated for using them as seed wells.
Correlation done in this manner produces very good correlation without much user intervention in areas of moderate to high well density. Thus allowing the analyst to correlate large data set in very less time. This approach is great in correlating distinct markers allowing analyst more time to focus on complex geological areas.