Hydraulic Fracturing: Using Machine Learning to Get the “Right Recipe”


Hydraulic Fracturing: Using Machine Learning to Get the “Right Recipe”

The “tried and true” way to assess the performance of an oil well is to compute the total cost per barrel of oil produced once the very last barrel has been recovered.

However, other measures are used as a proxy, such as $/bbl of oil produced in the first year. This is reasonable given the dramatic first-year production declines exhibited in unconventional plays.

When trying to find the right “recipe” to complete a well, the end goal is to optimize the two most important variables; profitability (i.e. minimizing the cost function to complete the well) and production (i.e. first-year production, ultimate recovery). There are three main types of models used today:

1. Historical Analog Models

Historical analog models involve using data from previous wells in close proximity (and/or with similar parameters) to make an educated guess about the best methods to complete the current well.


  • High level and fast analysis that is much less rigorous than other methods.


  • Lack sophistication compared to other models
  • Could potentially leave a lot of potential production and profitability on the table.
  • They are very rudimentary and less capable in terms of looking multiple variables in combination.
  • Good at solving “what”, but not “why”.

2. Simulation Models

Simulation models utilize computer software to account for multiple variables at once and have the capability to predict output and create visualizations of what events physically transpire in a petroleum reservoir.


  • Better addresses “why”


  • Intensive in terms of resources and computing run-time, especially for physics-based models.
  • Trade-off between model accuracy and timeliness and feasibility of the model.

3. Machine Learning Models

Digitization, big data, artificial intelligence, analytics, machine learning (ML), and automation are some of today’s most frequent buzzwords in energy, but why are machine learning models becoming so interesting in the context of well completions and hydraulic fracturing?

In a recent paper in the Journal of Petroleum Technology, a study was conducted in the Permian basin where well completions were optimized using a machine learning model. “The data set consisted of [parameters such as]… [d]irectional surveys, daily production data, artificial lift method, formation tops, and stage-level completion data such as the number of clusters per stage, proppant amount, grain size, type of proppant, fluid volume and type, and stage length…”. Some of the features of the highest importance included proppant type, grain (mesh) size, proppant volume, and cost data (comprising the price of frac sand).

The research enabled decision-makers to select a strategy that minimized cost without sacrificing too much of the initial 12-month oil production.

The machine learning model has the potential to be the best of both worlds compared to the simulation model and the historical analog model. The machine learning model is easier to run relative to physics-based simulations, while still capturing the key drivers and the “why” behind optimizing production and minimizing cost.

This is exceptionally important in a low commodity price and cash-constrained environment for E&P and oilfield services companies, as completions (encompassing hydraulic fracturing treatments) are the largest capital expenditure in upstream oil and gas.

Learn more about how PanXchange can be of use to provide up-to-the-minute data for your optimization needs. High-quality model inputs lead to quality model outputs.