Machine learning can play an important role in making subsurface quantitative interpretation workflows more efficient, consistent and potentially more accurate. Two workflows are shown in 1D and 3D applications. It is argued that the 1D cases are more about improving efficiency whilst the 3D cases have the potential to improve the accuracy.
Examples are shown from conventional and unconventional basins. Beyond that it is demonstrated how one can combine deep learning and physics-based models to provide fast and accurate subsurface predictions.
In this paper applications of the above two categories of machine learning workflows to some key quantitative interpretation workflows in both conventional and unconventional reservoirs will be discussed. In the conventional reservoirs case, a machine learning based petrophysical interpretation, workflow 1 mentioned above, will be discussed in which uncertainty in the prediction is captured using a Bagging approach. In the unconventional case, a machine learning approach based on workflow 2 above is shown for 3D sweet spot analysis by collectively predicting pore pressure, volume of Kerogen and geomechanical attributes. Both these examples are generally time consuming tasks if performed manually and hence it is demonstrated that machine learning can lead to both efficiency and accuracy. Beyond that it is demonstrated in workflow 2 how deep learning can be combined with physics-based models to provide fast and accurate subsurface predictions.
Introduction
Machine learning can play an important role in making subsurface quantitative interpretation workflows more efficient and consistent which should ultimately lead to a more confident decision making process. There are two categories of machine learning workflows in subsurface quantitative interpretation and prediction:
1.Train in 1D and apply in 1D:
a) Training step: a model is calibrated to a relatively small number of wells (logs, cores) in the relevant basin or sub-basin.
b) Application step: the calibrated model is applied to all the other wells in the same region of interest.
This workflow is by and large about efficiency. For example, train a supervised model to predict, say, porosity, on 10 wells with manual interpretation, and apply to the other 90 wells.
2.Train in 1D and apply in 3D:
a) Training step: a model is calibrated to the data (logs, cores) of all wells, or a representative subset of the wells, in and around the 3D volume.
b) Application step: the calibrated model is applied in 3D to seismic attributes and seismic inversion results (e.g. elastic properties).
This workflow is mostly about improving accuracy and confidence. To date, upscaling the well-based models into 3D has been performed using Rock Physics, e.g. a Rock Physics Model (calibrated to well data) transforms elastic properties into rock properties such as porosity or pore pressure. Machine learning improves on this by incorporating more information than merely the elastic properties, such as: well coordinates (so that lateral trends are captured), depth below datum (to incorporate compaction trends), temperature information (e.g. from a basin model), etc.
However, it should be stressed that using machine learning to predict absolute quantities (e.g. porosity, pore pressure etc.) directly from seismic, a relative measure, is unphysical. Low frequencies must somehow be inserted, a process known as (model-based) seismic inversion. A facies-based seismic inversion (Zabihi Naeini and Exley, 2017) is optimal for this purpose, which means the outputs from this process are not only elastic properties but also a discrete facies distribution.
