SPE/IADC-194105-MS

Prediction of Penetration Rate Ahead of the Bit through Real-Time Updated

Machine Learning Models

Yuanjun Li, University of Southern California; Robello Samuel, Haliburton

Copyright 2019, SPE/IADC International Drilling Conference and Exhibition

This paper was prepared for presentation at the SPE/IADC Drilling International Conference and Exhibition held in The Hague, The Netherlands, 5-7 March 2019.

This paper was selected for presentation by an SPE/IADC program committee following review of information contained in an abstract submitted by the author(s).
Contents of the paper have not been reviewed by the Society of Petroleum Engineers or the International Association of Drilling Contractors and are subject to correction
by the author(s). The material does not necessarily reflect any position of the Society of Petroleum Engineers or the International Association of Drilling Contractors,
its officers, or members. Electronic reproduction, distribution, or storage of any part of this paper without the written consent of the Society of Petroleum Engineers or
the International Association of Drilling Contractors is prohibited. Permission to reproduce in print is restricted to an abstract of not more than 300 words; illustrations
may not be copied. The abstract must contain conspicuous acknowledgment of SPE/IADC copyright.

Abstract
Rate of penetration (ROP) in petroleum engineering refers to the speed of forward motion of the drilling tools
during the drilling process. This is an important parameter that has long been optimized for maximization,
keeping in mind human, safety, and environmental factors along with consideration to downhole tools. Its
importance is validated by estimating the drilling. Longer estimates of drilling time translate to increased
costs.
Drilling costs are affected mainly due to the following contributing factors: non-productive time, idle
time, and invisible time. Attempts have been made to reduce these times to reduce costs. Simultaneously the
time taken for drilling can also be reduced by effectively increasing the ROP. Drilling depths, on average
are between 5,000 to 10,000 feet, coupled with a formation that has complex properties are major factors
contributing to non-productive time covering a high proportion of drilling time. Thus, a large non-productive
time leads to longer drilling cycles, and eventually, a low ROP.
In an attempt to reduce the non-productive time, there is a need to optimize the ROP. Higher ROP
facilitates a decrease in time and thus costs.
In this paper, ROP is effectively predicted using artificial neural networks not at the surface, but at the
bit. The artificial neural network has several advantages that overcome the limitations of the conventional
models. By effectively predicting ROP, estimation of the whole drilling process time and cost, identification
of specific reasons that slow down the drilling process are possible, and proper measures to avoid these
issues can be implemented. The target of any ROP optimization strategy should be to have the highest ROP
mechanically possible, considering human health, safety, and environment, and factoring in conditions of
the well and drilling state.

Literature Review
The rate of penetration achieved with a bit has a direct effect on the cost per foot drilled. Some of the
variables that affect ROP and are considered in this project include the following: weight on bit (WOB),
bit rotational speed (N), torque, flow rate (Q), hook load, and stand pipe pressure. (Motahhari, et al., 2009)
2 SPE/IADC-194105-MS

Various models considered during this project utilize a range of parameters listed above and also are limited
by several constraints. (Hegde, et al., 2018)
The Bingham Model, derived after modification from the Maurer Model, is a relatively simple and
experimental model. Its applications are limited to low weight on bits and RPM. However, it does not
consider the depth of drilling, making this model less favorable and reliable.

The Bourgoyne and Young's Model describes the rate of penetration as a function of several variables or
functions such as sediments, compaction, rock strength, pore pressure, weight on bit, RPM, bit hydraulics,
teeth wear, etc. The complexity of the model can be understood by the following equation

The Warren Model is generally developed for soft formations bits under the assumption that the cutting
removals do not hinder the ROP. The model relates ROP to WOB, N, rock strength, and bit size using
dimensional analysis and generalized response curves. Initially, the model was developed to depict the inter-
relation between bit and rock and accounts for the effect of cutting generation, cutting removal, the "chip
hold down effect," and bit wear on penetration rate.

Furthermore, dimensional analysis was used to isolate a group of variables consisting of the modified
impact force and the mud properties to account for the cutting removal.
The above models present the following limitations:
i. The complexity of the model for detailed description of drilling phenomena is often compromised
by practical limitations as many parameters affecting the model are either unknown or insufficiently
studied for proper modeling.
ii. Inaccurate description of the surface and down-hole boundary conditions.
iii. The assumption of many factors, and the reduction in the number of key parameters taken into
account by the model, decreases the overall accuracy and renders it inadequate.
iv. Some parameters, such as bit wear, are subjected to human error due to lack of expertise.
v. Requires multiple regressions, simulations, or physical experiments.
vi. Lacks adaptability which requires constant modification for future applications.

Introduction
Neural networks, which is also usually referred as deep learning, is a very powerful machine learning
tool. With the help of neural networks, we can quickly set up the connection between different parameters,
which is extremely valuable for the industry application. Since last century, neural networks have been
tremendously and effectively applied in oil industry under different realms, such as well testing analysis
with time series data (Tian and Horne, 2015), geostatistical data analysis (Zhao, et al. 2018) and production
optimization (Li, et al. 2018). Those methodologies and applications have been approved beneficial for the
oil companies.
With artificial neural networks, predictive models are relatively easy to build and can overcome the
limitations of the traditional methods that involve mechanism and expert analysis. In an attempt to overcome
SPE/IADC-194105-MS 3

these challenges and predict ROP not at the surface but at the bit, i.e. downhole, the use of artificial neural
networks is employed.

Novelty
Using artificial neural networks, where the data is the model, an attempt to create a model to predict the rate
of penetration ahead of the drill bit was performed. The model was trained on surface and downhole data.
Further, predicting the ROP (Figure 1), ahead of the bit by a distance in feet, was limited by the stability
of the results. By predicting ahead, the problem of time lagged data arriving from the bottom of the well
can be alleviated and also possibly solve the discrepancies in surface and bottomhole data communicated
by downhole tools and sensors.

Figure 1—ANN Window for ROP Prediction

Predicting ahead can also validate the need for downhole sensors and telemetry technology from an
economic standpoint. If downhole data can be predicted using surface data within an acceptable tolerance,
then the use of expensive downhole telemetry such as an optical fiber wireline can be avoided. The use
of surface data extrapolated to downhole conditions using the ANN model would be sufficient for ROP
optimization and making effective decisions in real time, thereby avoiding any latency issues of obtaining
downhole data. Downhole data can initially be used solve discrepancies of prediction to improve the ANN
model.

Methodology
In this study, depth data and the corresponding parameters are provided, which are the weight on bit, torque,
and RPM at both surface and downhole (Table 1). The goal is to predict the value of ROP, ahead of depth,
by using these parameters. The research is to evaluate the importance of downhole data for ROP prediction
because downhole facilities are expensive. If the prediction of ROP from surface data is accurate enough,
downhole data would not be needed.
4 SPE/IADC-194105-MS

Table 1—Three Input Scenarios.

WOB Avg (SF) (klb)

Scenario 1 Torque Abs Avg (SF) (f-p)

RPM Surface Avg (SF) (rpm)

Depth (ft)

WOB Avg (SF) (klb)

Scenario 2 Torque Abs Avg (SF) (f-p)

RPM Surface Avg (SF) (rpm)

ROP Avg(fph)

Depth (ft)

WOB Avg (SF) (klb)

Torque Abs Avg (SF) (f-p)

RPM Surface Avg (SF) (rpm)

Scenario 3
WOB Avg (DH) (klb)

Torque Abs Avg (DH) (f-p)

RPM Surface Avg (DH) (rpm)

ROP Avg(fph)

Data Format
For the given question, three types of input combinations have been designed. The first scenario is using
weight on bit (WOB), average absolute torque, and RPM only. Scenario 2 added depth information and
previous ROP. Because this study predicts ahead of bit, the previous ROP can also be used to predict new
ROP. These two scenarios are only using surface data. The third scenario added downhole data to the inputs
for scenarios 1 and 2.

ANN Configuration
The inverted triangle structure was used as the neural network configuration. Neuron numbers for the first
layer are 20, 12, and 10 (Figure 2). Fitting for the network for the first two layers is log-sigmoid; for the
third layer is tangent-sigmoid. The training function used was Levenberg–Marquardt backpropagation.

Figure 2—Configuration of the Neural Network

Predicting Ahead of Depth

Normally, ANN is utilized to predict parameters at the same depth such as permeability prediction. However,
in this study, a staggered-layer prediction was proposed. For a short distance, the formation lithology was
assumed to remain the same. As a result, the rate of penetration ahead of depth can be predicted (Figure 3).
For example, the parameters from depth X are used to predict depth (X+Δx). Meanwhile, in the second and
third scenarios, the ROP will also be used to predict the new ROP.
SPE/IADC-194105-MS 5

Figure 3—Configuration of the Prediction Model

For the given data set, the data was divided into training data (X1) and testing data (X2). The division
method for training data is by dividing into 70%:15%:15% ratio. For testing data, data that stays at the
prediction distance (D) from testing data was used. The prediction distance (D) should also be determined
before training the neural network. Therefore, for the given data set, the length of the prediction distance
was incrementally increased and evaluation of the regression results and prediction accuracy was done.

Results
Prediction Distance Evaluation
When the input data from the first scenario (Figure 5) is used, the regression results are below 0.8 regardless
of the distance (Figure 4). When using the second input scenario (Figure 5), with depth and ROP, the
regression results improved significantly. The prediction regression results are above 0.8 for short distances.
For the third scenario (Figure 6), the fluctuation of the regression curve is smoother than the previous one.
Downhole data can provide a more stable prediction for ROP. However, when the distance reached 150ft,
the regression becomes unstable. Therefore, 40 ft was used as the prediction length.

Figure 4—Regression Results for Scenario 1 (Surface Data)

6 SPE/IADC-194105-MS

Figure 5—Regression Results for Scenario 2 (Surface + Depth + ROP)

Figure 6—Regression Results for Scenario 3 (Surface + DH + Depth + ROP)

Prediction Results
Table 2 summarizes the results for the three different scenarios. A significant increase in the regression is
noticed from scenario 1, 0.56, to scenario 2, 0.87, owing to that fact that the number of parameters was
increased by 2 indicates that greater the number of parameters, the better the results. Both scenario 1 and
2 considered surface parameters only. To test the hypothesis and study the effect of downhole parameters,
for scenario 3, surface and downhole parameters were considered. The results displayed an increase in
regression by 4.6%, to give 0.91. Furthermore, the ROP can be predicted up to 40 stably. Both scenario 2
and 3 can provide high regression results.

Table 2—Regression Results.

Scenario Regression Parameters Input Data Combination

1 0.56 3 Surface only

2 0.87 5 Surface + Depth+ ROP

3 0.91 8 Surface + Downhole+ Depth+ ROP

Figure 7 is the comparison of the results given from three different scenarios. The overall regression is
only around 0.6 for using surface data only. After using depth and ROP, the ANN model in orange can
SPE/IADC-194105-MS 7

mimic the real ROP curve, in blue, more closely. And overall regression can reach as high as 0.87. For the
third type of input data, the regression for overall data is above 0.9. But for testing data, it does not improve
much. The 0.8 testing data was when using only surface data. When using downhole data, the regression
for testing is 0.82.

Figure 7—ROP Prediction Curves

Discussion and Conclusion

The accuracy achieved depends on the available data in terms of quality which can be overcome with data
processing techniques. To address the complexity of the formation gauged by surface data, there may be
some discrepancies which can be overcome by prior downhole data. The neural network employed here is
most simplistic but can be improved with complex models, like convolutional neural network (CNN) or
recurrent neural network (RNN) that work well with data in a series. This model is based on providing as
much information from as little data as possible. It is given that if the dataset was larger, the results would
be better. The results indicated here do provide a trend to follow and is feasible enough to validate the aim
of the study.
In conclusion, if the primary aim is to reduce drilling costs and non-productive time, use of telemetry
systems and downhole sensors can be avoided.

Acknowledgements
The authors would like to thank the University of Southern California and Halliburton for their resources
and contributions.

References
Motahhari, H. R., Hareland, G., Nygaard, R., & Bond, B. (2009, June 1). Method of Optimizing Motor and Bit Performance
for Maximum ROP. Petroleum Society of Canada. doi:10.2118/09-06-44-TB
8 SPE/IADC-194105-MS

Hegde, C., Daigle, H., & Gray, K. E. (2018, October 1). Performance Comparison of Algorithms for Real-Time Rate-of-
Penetration Optimization in Drilling Using Data-Driven Models. Society of Petroleum Engineers. doi:10.2118/191141-
PA
Tian, C., & Horne, R. N. (2015, September 28). Machine Learning Applied to Multiwell Test Analysis and Flow Rate
Reconstruction. Society of Petroleum Engineers. doi:10.2118/175059-MS
Zhao, X., Popa, A. S., Ershaghi, I., Aminzadeh, F., Li, Y., & Cassidy, S. D. (2018, April 22). Reservoir
Geostatistical Estimation of Imprecise Information Using Fuzzy Kriging Approach. Society of Petroleum Engineers.
doi:10.2118/190051-MS
Li, Y., Popa, A., Johnson, A., Ershaghi, I., & Cassidy, S. (2018, April 22). Dynamic Layered Pressure Map Generation
in a Mature Waterflooding Reservoir Using Artificial Intelligence Approach. Society of Petroleum Engineers.
doi:10.2118/190042-MS
Liu, Z., & Samuel, R. (2018, March 6). Probabilistic Real-Time Trajectory Control Considering Uncertainties of Drilling
Parameters and Rock Properties. Society of Petroleum Engineers. doi:10.2118/189702-MS