Robust Estimation of Reservoir Properties - How to Turn a Week into Minutes with Machine Learning

The more precise an accurate model of the subsurface is, the lower the risk in hydrocarbon exploration. Conventional petrophysical and rock physics workflows, such as integration of well log data and its interpretation with seismic data, are a lengthy process. It’s slowed by manual processing and geological complexity, but the integration of advanced machine learning algorithms can transform turnaround from a week to minutes while providing results that are just as high quality and more consistent.

The potential of advanced machine learning algorithms for estimation of porosity, saturation, and shear velocity was recently demonstrated on an extensive petrophysics and rock physics database from the Norwegian Sea. While the implementation of machine learning techniques for well log prediction is not new, the scale of this study encompassing an extensive dataset demonstrates that petrophysical properties can be quickly and accurately measured across numerous wells, and even across different geographic locations.

A New Approach Delivers Fast Predictions of Porosity, Saturation, and Shear Velocity

The study used a machine learning approach that is trained to derive predictions based on the measurements of properties from samples with similar observations. These decision-tree-based models are selected over neural networks, which learn to imitate a relationship between the input and target properties.

The figure below shows the generalized workflow for well log prediction. It starts with cleaning and randomly splitting the input well data from 141 exploratory wells so that 80% are used for training and 20% are kept aside for blind testing of the model, ensuring that the machine learning algorithm predicts on wells that it hasn’t experienced.

generalized_workflow_machine_learning_rock_physics
Generalized workflow for well log prediction utilized in the study.

The decision tree model then applies its built-in feature selection which is embedded in its training process. The process is similar to how a petrophysicist would address the task, concentrating on few key logs such as gamma ray, neutron porosity (NPHI), and Vp and Vs. In conventional workflows, based on an estimate of one week per well, it would take the petrophysicist more than three years to condition all wells included in this study.

The robustness and success of the approach show the potential for integrating existing petrophysical and rock physics libraries using machine learning, in order to estimate porosity, hydrocarbon saturation, and Vs in a very short time.

"As an example, we observed that it took under 25 minutes for a standard workstation to train our three different porosity models, and only a fraction of a second to predict porosity from these three models on a new well with very high accuracy”, says Roberto Ruiz.

The table below shows differences in turnaround between conventional, i.e. manual processes, and those performed by a machine. The timings from the conventional process presented here are indicative and assume that the petrophysicist starts from the same set of input logs as the machine learning algorithm after the training phase, as shown in the workflow.

Well log curve Turnaround: Conventional Turnaround: Machine Learning
Porosity 8 hours <1 min
Hydrocarbon saturation 8 hours <1 min
Shear Velocity 4-8 hours <1 min

 

Advanced Algorithms Solve Well Logs That Have Missing Measurements

Missing measurements when working with well logs is a well-known problem that can negatively impact the accuracy of the prediction. Sometimes issues arise with the logging tool and a particular log cannot be acquired over a certain interval.

One key advantage of the decision tree algorithm over neural networks is that it handles missing measurements relatively well, whereas neural networks can neither train nor predict from an incomplete set of inputs – they can only predict when the full suite of input logs is present. Both approaches rely on Mean Squared Error (MSE) to express how close the prediction is to the test set. However, there are also trends and changes of direction to consider, which the MSE cannot capture. To overcome this issue, two additional metrics were implemented - Predictability (PEP) and Goodness of Fit (R2). A model that precisely predicts the true values, scores 1.0, which is the best possible score. In this study, all porosity, hydrocarbon saturation, and Vs models showed good performance, with the majority of scores close to 1.0.

All three measures together provide complementary information – MSE helps assess the general accuracy of the models, while PEP and R2 help us understand the overall fit of the predictions.

Reservoir Elastic Properties are Consistent Across Geographic Locations

Using the PGS rockAVO library composed of 141 wells, a variety of formations and lithologies are included in the study, ranging from Tertiary to Triassic age in the Norwegian Sea.

Unlike prior studies using similar techniques, this new approach works on far larger datasets and does not require inputs such as mineralogy or fluid saturation in order to perform a robust prediction.

PGS_Norwegian_rock_physics_atlas
Distribution of 141 exploratory wells used in the study in relation to oil fields (green), gas fields (red), and PGS 3D MultiClient seismic data in the Norwegian Sea.

Porosity prediction – estimating the fraction of total fluid-filled spaces in the rock (PhiT) is a complex and extensive exercise. A robust petrophysical interpretation is required as part of the conventional petrophysical workflow, on top of careful calibration of the main parameters in the rock physics model equations, to produce a reliable final total porosity log.

Using decision tree models, the results show that a porosity model trained in the Norwegian Sea (Well 7122/4-1) can be adapted successfully for another area, despite having a very different geological history. This indicates that a machine learning model trained in one region could be used to estimate an initial PhiT in a new well, from a new area directly from raw logs.

Machine learning algorithms are a very promising option for estimating properties accurately, consistently, and efficiently, as well as an extremely useful tool for optimizing current petrophysical and rock physics workflows to provide rapid results.