Deep QI: A Machine Learning Approach to Quantitative Interpretation of Subsurface Data

For the exploitation of both conventional and unconventional plays, lithology classification, petrophysical evaluation, pore pressure prediction and geomechanical analysis play critical roles in accurate reservoir characterisation, safe well planning and execution, sweetspot identification etc. In unconventional plays specifically, the ability to predict areas of higher productivity depends on understanding the pressure and stress magnitudes. In general, any pressure-stress-property model must be supported by petrophysically conditioned logs, calibrated to core data. It’s important for the industry to develop safe and innovative methods which keep pace with the drilling activity and harness all data effectively.

Machine learning can play an important role in making sub-surface interpretation workflows faster, more consistent and in certain cases superior. This leads to quicker, more confident results and therefore improved decision making. Its adaptability means that machine learning can be deployed in different subsurface workflows as explained below. 

Speed and efficiency in 1D

In 1D, a model is calibrated (‘trained’) to targeted stratigraphic data from a relatively small number of wells in the relevant basin or sub-basin. In the application phase, the calibrated model is applied to all other wells in the same region of interest. This workflow is about speed and efficiency, for example, train a supervised model to predict, e.g., porosity, on 10 manual interpretated wells, and subsequently apply to any number of wells for which porosity has not yet been evaluated, all virtually instantaneously. This type of machine learning workflow allows personnel to focus on adding value to the interpretation process by fine-tuning the training data and QC’ing the machine learning results rather than spending a significant amount of time repeating standard workflows on a large number of wells. Note that this approach can also credibly predict missing data, for example where measurement failed or logs were not run.

Figure 1 - Blind well located 50km from training data - overall good match between manual (black) and ml (red)

Improve accuracy and confidence with 3D.

Machine learning can also be allied to new methods in 3D quantitative seismic interpretation,  offering spatial prediction in between wells. In this case, training of machine learning models takes place on well data (as per previous paragraph – see also Figure 1 for QC at a blind well), and then in the application phase, the calibrated (‘trained’) models are applied sequentially (ie. the output of one is the input to the next machine learning model) to traces of a seismic inversion results (e.g., elastic properties – see Figure 2).

Training a machine learning model to predict, say, porosity directly from seismic traces is futile, as such a model cannot ‘invent’ the required low frequency data; hence the need for a quality seismic inversion step. This workflow is not so much about increasing efficiency/speed of execution, but is more about improving accuracy and confidence in the results, so that e.g. geobodies can be extracted with more confidence or a reservoir model can be populated more accurately.

Figure 2 - Example of a pore pressure model derived using a deep neural network applied to a facies-based seismic inversion result

What about 4D?

There is the possibility to extend to 4D dynamic data. Time-lapse seismic and simulator models can be incorporated in the workflow to predict production induced saturation and pressure changes.  By integrating well surveillance data, the pressure and production rates may be correlated with seismically derived properties to establish promising relationships between 3D/4D seismic and well performance. 

Find out more

We hope we have demonstrated a promising outlook for the application of machine learning to save valuable turnaround time and potentially improve the accuracy in subsurface workflows. Differing data classes and disciplines can be integrated to achieve new insights and a common objective. 

For more details join one of our upcoming events or/and download the following publications.