Explaining Shale: Can Machine Learning Capture Complexity?

This data-file decomposes the drivers of shale productivity in Alberta’s Duvernay play, across a correlation-matrix of 23 different variables.

Machine learning can be used to predict 78% of the variance in wells’ performance from this data-set, surpassing the 19-67% predictive power of regression models (chart above). Accordingly, $1M/well savings are suggested, while well productivity can improve by 19-97%.

Shale is a data industry. “Big data” approaches are the only way to capture the complex inter-correlations within shale’s productivity drivers. As shown below, well EURs are meaningfully correlated with 12 variables. The “largest” driver is “proppant placed”, which is itself meaningfully correlated with 16 other variables.

Machine learning is still in its infancy in the shale patch, representing c2% of total industry-research. It presents material upside.


Machine Learning to Optimise Rod Pumps

This data-file summarises progress using machine learning to maximise production from mature wells by detecting errors and optimising production. The algorithms are getting more accurate.

Methodology. First, we tabulate the accuracy of prior ML studies, touching on initiatives from Equinor, Conoco and Concho. Next, we focus in on an excellent, recent technical paper, achieving 5-10% production uplifts using machine learning to optimise 300 wells at the Bahrain oilfield.

Hence we constructed a simple model for digitising rod pumps: we estimate $100k of NPV can be created through instrumenting a typical rod pump well early in its life.

Global decline rates can be lowered by c100kbpd per annum for over a decade, using these improved algorithms.

Copyright: Thunder Said Energy, 2022.