Unconventional Wells Interference: Supervised Machine Learning for Detecting Fracture Hits

Featured Application

The methodology developed in this study provides a robust way of tracking down specific events in hydraulic fracturing. The innovative workflow produced an effective diagnostics tool. Ultimately, the featured event detection and diagnostics can make it possible to find optimal solutions for oil/gas well completion design, spacing and infill, production management, and wellbore protection.

Abstract

The primary objective of the study was development of a machine learning (ML)-based workflow for fracture hit (“frac hit”) detection and monitoring using shale oil-field data such as drilling surveys, production history (oil and produced water), pressure, and fracking start time and duration records. The ML method takes advantage of long short-term memory (LSTM) and multilayer perceptron (MLP) neural networks to identify the frac hits due to hydraulic communication between the fracking child well(s) and the producing parent well(s) within the same pad (intra-pad interaction) and/or on different pads (inter-pad interaction). It utilizes time series of pressure and production data from within a pad and from adjacent pads. The workflow can capture time variable features of frac hits when the model architecture is deep and wide enough, with enough trainable parameters for deep learning and feature extraction, as demonstrated in this paper by using training and testing subsets of the field data from selected neighboring pads with over a couple of hundred wells. The study was focused on frac-hit interaction among paired wells and demonstrated that the ML model, once trained, can predict the frac-hit probability.

1. Introduction

Hydraulic fracturing, commonly known as fracking, is the mainstream technique employed in unconventional fossil-fuel field development, which enables extraction of petroleum (oil) and natural gas from shale and other “tight” rock formations. It is a well stimulation technique used for the fracturing of bedrock formations by a pressurized liquid with a proppant designed to keep the hydraulically induced fractures open and thus facilitate extraction of hydrocarbons after the pressure from the injection subsides [1,2]. Tighter well spacing to maximize the utilization of all available pockets of fossil fuel may lead to inter-well interference. A fracture hit (aka “frac hit”) is a form of hydraulic inter-well communication, usually between neighboring horizontal wells, where an existing offset well (often termed the parent well) is affected by pumping hydraulic fracturing fluids into a new well (called the child well) or another active well, either at the same or a neighboring pad. With progressing field maturation and infill drilling, at some point, the induced frac hits may start to critically affect the efficiency of fluid recovery.

Hydraulic fracturing, in combination with horizontal drilling, increased the profitability of thin petroleum-rich deposits in shale formations; its use as a source of U.S. oil and natural gas production began steadily growing in the early 2000s. The process involves drilling of a vertical well to the depth of the formation and then bending the drilling path until it extends horizontally to the designed lateral length [1,2], as shown in Figure 1. (The fracture network includes pre-existing natural fractures as well as hydraulic fractures). As field operators increasingly expand the production by the infill drilling of new wells to maximize the utilization of the asset’s acreage, frac hits can impact the exploration, field development, drilling and completion design, production optimization, and cost-effectiveness of operations, which include choke management, reservoir drainage, estimated ultimate recovery (EUR), well placement, and wellbore protection [3,4].

Over the past decade, a few publications have provided guidance on conducting analysis of frac-hit impacts on production and wellbore integrity. Sardinha et al. [5] presented a process for characterization of the hydraulic fracturing pressure interference from the viewpoint of persistent inter-well connectivity mapping. Lehmann et al. [6] improved the pressure interference analysis by adding alignment with high-resolution microseismic data. Another notable discussion on the types of inter-well interference (between producing wells, stimulated wells, and offset wells) was presented by Rimedio et al. [7] based on a fieldwide interference study during the development of Vaca Muerta shale resources. King et al. [8] and Daneshy et al. [9] reported the pressure interference effects, root causes, damage potential, and emerging mitigation strategies in a series of papers based on several case studies. Moreover, Ajisafe et al. [10] discussed their results, by analyzing production in the Avalon Shale of the Delaware Basin, that show the child-well oil production being approximately 30% lower than what was observed at the corresponding parent well. Lindsay et al. [11] reported that depending on the well spacing, there was an up to 50% probability of the infill/child well outperforming the parent well. One of the main reasons for underperformance in several unconventional basins was that such infill wells were drilled with closer well spacing than the parent wells. On the other hand, depending on the basin, time-dependent productivity gains may also be induced by frac hits due to enhancement of the parent-well stimulated volume. In addition to production interference, frac hits may inflict wellbore damage; wellbore damage is another high-profile problem related to the larger, but less obvious, issue of how fracturing affects the reservoir structure between tightly spaced wells. King et al. [12] communicated similar findings based on their analysis as well. To facilitate their diagnosis, Liu et al. [4] adopted decision-tree techniques for analysis of the pumping unit and production records and found the reduction in well spacing to be the primary contributor to frac hits. A multilevel approach to understanding the comprehensive fracture network was tested as well [4]. The proposed approach integrated multi-resolution datasets of the well log and core data, along with completion, flowback, production, testing/monitoring (tracer data and offset pressure), and microseismic data. By applying machine learning (ML) separately to each dataset, the approach provided the well-level, segment-level, and stage-level overview of the fracture network. Then, all three levels of the network representation were aggregated via information fusion to build a multilevel framework of the holistic modeling. The multilevel fracture network can be interactively visualized to support decision making on planning of the infill drilling and well completion design, as well as improving unit operations such as choke management and pressure control [4].

The case study reported here builds on the prior efforts of developing ML models that can be used for conducting analysis of frac-hit impacts based on the data from selected pads, including drilling surveys, bottom-hole pressure, oil and water production, and fracking start time and duration information [4]. Incidentally, the economic impact of frac hits and shut-ins (closing off producing wells) was estimated in the most recent report [12] based on the same datasets as the ones used in this work. The ultimate goal is either complete prevention of the major negative economic impacts or optimization of the well operation to minimize such impacts. Physics-based modeling can assist in understanding the mechanics of unconventional reservoirs, such as stress field alterations affecting well performance [13,14]. However, traditional computational approaches based on the known and trusted concepts may not provide the most comprehensive and realistic model of heterogeneous geological systems with a variety of unknown features [15]. The proposed ML method attempts to close this gap by taking advantage of long short-term memory (LSTM) and multilayer perceptron (MLP) neural networks to identify frac hits due to hydraulic communication between the fracking child well(s) and the producing parent well(s) within the same pad (intra-pad interaction) and/or on different pads (inter-pad interaction). It utilizes time series of pressure and production data from within a pad and from adjacent pads. The principal conclusion is that such a workflow can help to capture time-variable features of frac hits. This case study was conducted at the U.S. Department of Energy’s (DOE) National Energy Technology Laboratory (NETL) in collaboration with an industrial partner, a major shale-field operator in the Permian Basin, located in West Texas, United States, to investigate the opportunities for the application of ML techniques to facilitate prompt frac-hit event detection and operational decision-making support.

2. Materials and Methods

2.1. Data Subsets Selection

All available data were examined for completeness. Pad 137 was selected as the reference because the most comprehensive data was available for that location. Figure 2 shows the geolocations for all wells in the area of interest (selected from several hundred wells available in the field). The data for Pads 133 and 138 were also used in the analysis.

Table 1. Operational time span for Pad 137 wells as days counted from the Well 137-42006 DOFP.

Well ID	Frac Start, Day	Frac End, Day	DOFP, Day
137-42006	−53	−27	0
137-41996	−53	−27	2
137-42048	−22	−6	15
137-42338	−53	−27	16
137-42106	−5	4	21
137-42104	−22	−6	22
137-42740	431	453	477
137-43200	431	453	477
137-43201	431	454	477
137-42276	433	452	477
137-42007	433	452	477
137-43675	433	452	477

shows the key operational information, such as the fracking dates and the date of first production (DOFP) for all Pad 137 wells. The events timeline is shown in days counted from DOFP of Well 42-329-42006, the earliest producing well in Pad 137.

One can observe two subsets of wells on this pad, with distinctly offset timing of the well-stimulation batches. The first batch subset comprised the top six wells in Table 1, where fracking had started before Day 0. The stimulation was completed after the start of production at the reference well at only one of these wells. The second batch subset (the remainder of the Pad 137 wells) did not even have fracking started until at least 430 days after the start of production at the reference well. Distinct pressure spikes, an indication of a likely frac hit, were visually identified at the second batch subset of Pad 137 wells between Day 430 and Day 450, as well as between Day 530 and Day 550 (Figure 3). The latter could be attributed to interference from fracking at the other pads. For example, there was fracking at Pads 4, 25, and 133 (see

Table A1. Pads with fracking between Day 530 and Day 550.

Project ID	Well ID	County ID	Fracking Start	Fracking End	Fracking Start Day	Fracking End Day
4	4-42283	1	17 February 2020	22 February 2020	524	529
4	4-42284	1	24 February 2020	1 March 2020	531	537
4	4-42285	1	24 February 2020	1 March 2020	531	537
133	133-41267	2	25 February 2020	3 March 2020	532	539
133	133-41268	2	25 February 2020	3 March 2020	532	539
25	25-39736	3	4 March 2020	11 March 2020	540	547
25	25-40356	3	4 March 2020	11 March 2020	540	547
133	133-41261	2	7 March 2020	13 March 2020	543	549
133	133-41262	2	7 March 2020	13 March 2020	543	549
25	25-39735	3	12 March 2020	19 March 2020	548	555
25	25-39737	3	12 March 2020	19 March 2020	548	555
25	25-3951	3	22 March 2020	29 March 2020	558	565

) during the same time span.

The likelihood of inter-well communication may depend on the local geology and the distance between the wells. Accordingly, Pad 133 was selected for a preliminary investigation into potential inter-pad interferences, because it was closer to Pad 137 than to Pads 4 and 25 (Figure 2). Pad 138, located between Pads 133 and 137, was then picked as the verification pad for the same case study. Eight of the twelve Pad 133 wells were stimulated between Day 91 and Day 128 (see

Table A2. Timeline for Pad 133 wells as counted from the Well 137-42006 DOFP.

Well ID	Frac Start, Day	Frac End, Day
133-40798	91	97
133-40854	93	103
133-40853	93	106
133-40787	100	109
133-40845	109	121
133-40846	109	121
133-40832	113	119
133-40847	122	128
133-41267	532	539
133-41268	532	539
133-41261	543	549
133-41262	543	549

), chronologically taking place between the first and second well-stimulation batches at Pad 137 (Table 1). The second batch at Pad 133 followed the second batch at Pad 137 and was of primary interest in the case study. DOFP data were not available for Pad 138 wells, which somewhat limited the data analysis. However, it did not critically affect the model validation and verification workflow.

Well stimulation at Pad 138 was done between Day 294 and Day 327 (

Table A3. Timeline for Pad 138 wells as counted from the Well 137-42006 DOFP.

Well ID	Frac Start, Day	Frac End, Day	DOFP, Day
138-42651	294	308	335
138-42682	294	308	332
138-42683	294	308	339
138-42751	296	306	333
138-42758	296	306	329
138-42759	310	324	340
138-42761	310	324	337
138-42762	310	324	344
138-42797	316	327	348
138-42805	316	327	347

). Some of the relevant data were missing, but the pressure and fracking timeline data were available for all 10 wells on Pad 138 (Figure 4a), for the second batch on Pad 137 (Figure 4b), and for the second batch on Pad 133 (Figure 4c). The average treatment pressure (calculated for each stimulated well) was the lowest for Pad 138 and the highest for Pad 137. The treatment (well-stimulation) pressure data were deemed to be particularly important because higher pressure at the child well could be more likely to lead to either pressure communication with the offset wells or the loss of well control (or well blowout) whenever this pressure is transferred to wells that have less pressure integrity. Higher treatment pressure is also more likely to produce fracture propagation far enough to impact the nearby producing wells.

2.2. ML Model Development

The objective of this case study was to develop and train ML model(s) that could accurately estimate the probability of a frac-hit type of communication between two wells, a hydraulically stimulated child-well and a producing parent-well, in near real time.

2.2.1. ML Model Structure

The data from the selected pads used for ML model development included time-series data for fracking time and duration, production rates for oil and water, and non-temporal geometric data from a drilling survey (only the distance between the well pairs was used in the study). These data were preprocessed and ingested as feature variables. The output of the trained ML model was expected to be the probability of a frac-hit type of communication between the two selected wells at any time.

All ML models in this study were constructed using long short-term memory (LSTM) and multilayer perceptron (MLP) neural networks [16]. These models take advantage of LSTM’s ability to handle temporal data, i.e., the time-series data in this study, and utilize MLP for feature learning and feature extraction. The effect of the number of LSTM and MLP layers, as well as the layer sizes (the number of nodes in each layer), on the ML model performance was tested using the configurations shown in

Table 2. Comparison of ML model configurations.

Model Settings	Small Model	Midsize Model	Large Model
Number of LSTM layers	2	1	1
Number of MLP layers	4	7	8
Batch normalization layer	Yes	Yes	Yes
Number of parameters	43,953	235,617	916,241

. The number of model parameters ranged from approximately 44,000 to 916,000. The details of a midsize model structure are shown in Figure A2. The corresponding number of parameters in each layer is shown in Figure A3.

2.2.2. Data Preparation for ML Model Training

As shown in Figure A2, the non-temporal data were first transformed into a constant time series, of the same size as the other time-series datasets, and then concatenated and input into the LSTM layer. Additionally, all input and output variables were normalized, so that the model was agnostic of the parameters’ physical units. Proper normalization factors, based on the minimum and maximum parameter values that the model ingested (full ranges of pressure, production rates, distance, etc.), were used to normalize each of the parameters within the range of [0, 1].

To prepare the data for the training and testing of the model, the source data from all 35 wells on Pads 133, 137, and 138 were used. Then, those 35 wells were mixed and matched to generate the binary permutation of wells with repetition, for a total number of 1225 ordered well pairs. Transforming the input data structure into vector data reflects the inherent causality of the communication between the wells within each data vector. A standard time range of 1000 days was used for all training datasets. Each temporal dataset had the time values computed with reference to DOFP at Well ID 137-42006 in Pad 137. All data prior to the reference point were truncated.

The LSTM layer’s input data had dimensions of (m, t, n), where the first element, m, is the total number of the datasets in a batch supplied to the model (number of training/validation or testing datasets), the second element, t, is the size of time series, and the third element, n, is the number of input features. For the sample model illustrated in Figure A2 and Figure A3, t = 1000 and n = 14. The feature variables were the initial well distance based on a drilling survey, the binary fracking indicator (0 for a non-fracking day and 1 for a fracking day), and the remaining twelve parameters, which were the measured values, interpolated values, time derivatives of the measured values, and time derivatives of the interpolated values for each pressure, as well as the oil and water production measurements. The model output had dimensions of (m, t, s), in which the first two dimensions are the same as in the model input, and the last dimension, s, is a single output variable for the probability of a frac hit at each time step.

To generate training targets for the ML model, the algorithm for computing the frac-hit probability, p_fh, as a function of the input features was uniformly applied as shown below:

$$p\_fh={\sum }_{k=1}^{n}f\left({\nu }_k\right)$$

(1)

where ν_k is the kth input feature, and f is the probability contribution function of ν_k. The frac-hit probability can thus be computed as a sum of the outputs of individual feature-dependent functions. The following linear function was used to initially approximate the target frac-hit probability, P_fh:

$$P\_fh={\sum }_{k=1}^n{c_k\nu }_k$$

(2)

where the coefficients, c_k, for the key features related to the frac hits were used in the calculations. After inspecting the available data and manually testing for various coefficients, their non-zero values were estimated as shown in

Table 3. Coefficients used in calculating frac-hit probabilities.

Feature	p	$\dot{\mathit{p}}$	O	$\dot{\mathit{o}}$	w	$\dot{\mathit{w}}$
$c_k$	0.1	0.8	0.01	0.04	0.01	0.04

Note: See the symbol definitions at the end of the article.

.

These values were arrived at by trial and error after extensive review and analysis of the available data and domain-knowledge guidance, with some degree of expert judgement in identifying the frac hit features hidden in the data. Initially, fourteen relevant features were pre-screened and then down-selected to six. The major challenge was the lack of a well-defined frac-hit likelihood measure, especially in low likelihood cases.

Figure 5, Figure 6 and Figure 7 show examples of the computed frac-hit probability with the corresponding pressure, oil production, and water production data from the parent well (137-42106 in Pad 137) around the time of fracking at the child well (133-40787 in Pad 133). As shown in Figure 5, a large spike in the time derivative of pressure due to fracking is the top contributor to the instant jump in frac-hit probability to 0.88.

After generating the data for all 1225 cases (ordered well pairs), a random data shuffling was performed with the same permutation for the input and output datasets. Then, ~90% of the data cases, or 1100 cases, were used for model training/validation, and the remaining ~10%, or 125 cases, were used for model testing. Cross validation was performed during the training process with an approximately 80%/20% split into training and validation data, respectively.

3. Results and Discussion

Each model was trained and tested with the same pre-processed data. The relative performance of the previously defined three types of ML models (Table 2) during training and testing is compared in

Table 4. Performance of the three ML model types.

Metrics	Small Model	Midsize Model	Large Model
Train Loss	1.91 × 10−5	2.85 × 10−8	1.26 × 10−9
Test Loss	2.07 × 10−5	5.86 × 10−8	4.98 × 10−9

. It was observed that with an approximately five-fold increase in the number of model parameters, going from a small to a midsize model, the minimized mean squared error (MSE loss) decreased by almost three orders of magnitude. Furthermore, with a four-fold increase in the number of parameters, going from a midsize to a large model, the MSE loss decreased by another order of magnitude. The reduction in the MSE loss with the progress in the model’s training is shown in Figure 8, using the large model as an example.

For predictions, the test data were then fed to the trained model. The time to make a prediction was within one second for all three models; the time variations among different models were more due to the computational overhead variations than to the model size. The model prediction and the ground truth obtained using the algorithm for computing the target frac-hit probability, as in (2), can be compared to each other to evaluate each model’s performance. Comparisons of the predicted and ground-truth values of frac-hit probabilities for intra-pad and inter-pad interactions are shown in Figure 9 and Figure 10 (using the large model in all examples). The title on top of each subplot shows the interacting well pair, with the stimulated child well’s ID followed by the parent well’s ID. The prediction and ground truth were mostly the same; however, the discrepancy was relatively larger for the cases with low probability values, which reflects the inherent difficulty in identifying potential frac hits and estimating their likelihood whenever the tell-tale features are not prominent.

With the trained ML models available, it is possible to estimate the probability of a frac hit due to the interaction between two wells located either within the same pad or on different pads. For the frac hits due to interaction between two wells within a single pad, the plots in Figure 9 show two examples of relatively higher frac-hit probabilities for the well pairs within Pads 137 and 138, respectively. For inter-pad well interactions, the frac-hit probability at the wells on Pad 137 due to the wells on Pad 133 (Figure 10, plots (a) and (b)), as well as at the wells on Pads 137 and 133 due to the wells on Pad 138 (Figure 10, plots (e) and (f), respectively) were below 10% and lower than those at the wells on Pad 138 due to the wells on Pad 137 (Figure 10, plots (c) and (d)).

The objective function used in this case study was designed to enforce a generally linear relationship with the pressure and production responses as well as with their first-order time derivatives, for concept demonstration purposes. However, this function definition is flexible and can be tailored to emphasize specific observations and understandings of the reservoir and the wells’ properties, based on the data availability and their quality. For example, the gas production variable was not included in the loss function only because of the associated data fidelity being much lower compared to the fidelity of the liquid production data, such as oil and water production, from the liquid-rich reservoir available for the case study.

4. Conclusions

The case study reported here utilized data records for three out of the ninety-six pads available in the industrial partner’s oil field. The goal was to develop an effective ML-based workflow and illustrate that a frac-hit event can be probabilistically recognized by pre-trained ML models. The predictions of the trained ML models agreed with the observations at the neighboring wells. The workflow was designed to recognize two types of frac hits, intra-pad and inter-pad frac hits. The first type of frac hits, corresponding to intra-pad well interference, is mainly due to relative offset between batches of fracking and production intervals, or the “batch” effect. There were two batches observed at Pad 137. The high frac-hit probability instances, with impacts on the early-batch production wells, were attributed to the late-batch fracking wells within the same pad. Inter-pad frac hits were also identified here, with interferences between Pads 133, 137, and 138. The methodology developed in this study provided a robust approach to tracking down frac hits based on pressure and production data. The innovative workflow produced a valuable tool for frac-hit diagnostics. Ultimately, the frac-hit event detection and diagnostics can make it possible to find optimal solutions for well completion design, spacing and infill, production management, and wellbore protection. The input data can be continuously streamed from the field into a ML-based workflow to enable the process to provide rapid insights into the well operations and support real-time decision-making such as drilling operations planning, shut-in control, and stimulation management.