Understanding Frac Hits Using Machine Learning

Can machine learning help asset teams plan for the new and potentially industry-threatening problem of frac hits? This problem affects any shale producer looking to extract more oil from their reservoir by closely spacing wells. The impact to production and unplanned costs and delays are prioritizing the need for a solution to better predict and manage frac hits. While frac hit research has ramped up over the past few years, most solutions are expensive and limited in their applicability between pads. Given the physics complexities associated with frac hits and the large amount of available reservoir and well data, applying machine learning to this problem seems the ideal solution. 


What are frac hits?
Frac hits, alternatively known as frac bashes or fracture interference, refer to a specific kind of well-to-well interference event where fractures initiated during completion of one well (the child) contact another previously drilled well (the parent) (See Figure 1).

Figure 1: Schematic of parent well with two children: (a) At time 0, single parent well, (b) At some time t, the same parent well with a depletion area around it, (c) Also at time t a child well completed with adequate spacing from the parent well, (d) Also at time t, a second child well completed with inadequate spacing from the parent well, which results in a frac hit.

Frac Hits are Difficult to Predict
There is a lot about hydraulic fracturing that is poorly understood and hard to predict. Fracture propagation is notoriously hard to predict at small scales, let alone the large-scale field cases which are required. While other problems can tolerate a large degree of uncertainty for good economic outcomes, hydraulic fracturing increasingly cannot.

Several potential solutions have been explored with limited success: RTA analyses; (expensive) pad tests using some combination of microseismic, downhole pressure gauges, and/or fiber; and high-fidelity simulations to capture fluid flow and fracture dynamics (see Figures 2 and 3 for examples). The ideal solution would transfer learning from one frac hit event to another pair of wells in an inexpensive, scalable way. Each of these solutions fall short as being too expensive, not repeatable, or not scalable. 

Figure 2: Example of an RTA analysis for a frac hit (from Lawal et al. 2013)

Figure 3: Microseismic activity showing a scenario where parent well depletion results in asymmetric fracture growth (taken from Walser et al. 2016)

Applying machine learning to understand and predict frac hits is more cost-effective than the aforementioned solutions and provides the scalability required for full-field development. A few papers provide examples of machine learning algorithms being applied to the frac hit problem, however they are focused on a few pairs of wells where stage-by-stage data is available. (see SPE 191789 and Machine learning-based fracture-hit detection algorithm using LFDAS)


This creates the same challenges as the other mentioned solutions - requiring data from expensive, specialized measuring equipment for each child completion stage (like microseismic, downhole pressure gauges) and cannot easily transfer learning from one pad to a new pad. 


OAG’s Approach

At OAG, we combine our understanding of accepted theories of pressure depletion and inter-well interference with machine learning algorithms. Success in understanding and predicting the probability of frac hits has been achieved with this approach:


1. Gather data about as many frac hit events as possible. The goal is to find frac hit events in as many wells as possible using all available historical data. For each child well completed, the OAG algorithm examines offset wellhead pressures and offset volumes in that time period which could indicate a frac hit. This is often how frac hits are detected in the field. 


2. Generate features for each well pair. This could include produced offset volumes (a proxy for depletion); neighborhood information for the pair, like rock types and oil/gas content in the reservoir; presence of faults; and macro-scale stress orientation. 


3. Run a machine learning model to predict the probability of a frac hit occurring for a given well pair. A good frac hit model will include categorical variables (e.g. what is the rock type, what is the type curve) and others are continuous variables (e.g. depletion area around parent well, proppant volume, treatment rate). The output of this step will identify the probability of a frac hit event occurring for that well pair. 


4. Run a separate model to predict impact of the frac hit on production volumes of the parent and child. Frac hits may have a positive or negative impact on production. Identifying the frac hit is not enough to determine mitigation plans. This model considers other factors such as when the hit occurred relative to the offset well’s life, difference between parent/child designs, etc.


The combination of understanding the probability of a frac hit on a specific well and predicting the impact on production makes it easier for an asset team to plan future wells.  Results strongly support OAG’s theory that machine learning is a more affordable and scalable solution to manage frac hits. This approach has been validated by predicting historic frac hits in multiple data sets. The models have been accepted by a large independent operator and will be tested in the field. Watch this blog for a future article on the results.

 Article by Rahul Verma

Published Nov 5 2019 

  • LinkedIn - White Circle
  • Twitter - White Circle

(844) 624-9355

1330 Post Oak Blvd, Suite 1825

Houston TX 77056

© 2019 OAG Analytics All rights reserved