SEE COMMENTARY

Gas production in the Barnett Shale obeys a simple

scaling theory
Tad W. Patzeka,1, Frank Maleb, and Michael Marderb
a
Department of Petroleum and Geosystems Engineering and bCenter for Nonlinear Dynamics and Department of Physics, The University of Texas at Austin,
Austin, TX 78712

Edited by Michael Celia, Princeton University, Princeton, NJ, and accepted by the Editorial Board October 2, 2013 (received for review July 17, 2013)

hydrofracturing

| shale gas | scaling laws | energy resources | fracking

he fast progress of hydraulic fracturing technology (SI Text,

Figs. S1 and S2) has led to the extraction of natural gas and
oil from tens of thousands of wells drilled into mudrock (commonly called shale) formations. The wells are mainly in the United
States, although there is signicant potential on all continents (1).
The fracking technology has generated considerable concern
about environmental consequences (2, 3) and about whether hydrocarbon extraction from mudrocks will ultimately be protable
(4). The cumulative gas obtained from the hydrofractured horizontal wells and the prots to be made depend upon production
rate. Because large-scale use of hydraulic fracturing in mudrocks is
relatively new, data on the behavior of hydrofractured wells on the
scale of 10 y or more are only now becoming available.
There is more than a century of experience describing how
petroleum and gas production declines over time for vertical
wells. The vocabulary used to discuss this problem comes from
a seminal paper by Arps (5), who discussed exponential, hyperbolic, harmonic, and geometric declines. Initially, these types of
decline emerged as simple functions providing good ts to empirical data. Thirty-six years later, Fetkovich (6) showed how they
arise from physical reasoning when liquid or gas ows radially
inward from a large region to a vertical perforated tubing, where
it is collected. For specialists in this area, the simplicity and familiarity of hyperbolic decline make it easy to overlook that this
functional form reasonably arises only when specic physical conditions are met. For example, all early decline curves were proposed for unfractured vertical wells or vertical wells with vertical,
noninteracting hydrofractures. Physics-based descriptions of such
wells are readily available in textbooks, such as those by Kelkar (7)
and Dake (8). However, these books focus on radial or 1D ow
www.pnas.org/cgi/doi/10.1073/pnas.1313380110

from innite or semiinnite regions to pipes or planes. The resulting decline curves do not apply to the wells we describe here.
The geometry of horizontal wells in gas-rich mudrocks is quite
different from the conguration that has guided intuition for the
past century. The mudrock formations are thin layers, on the
order of 3090 m thick, lying at characteristic depths of 2 km
or more and extending over areas of thousands of square kilometers. Wells that access these deposits drop vertically from the
surface of the earth and then turn so as to extend horizontally
within the mudrock for 18 km. The mudrock layers have such
low natural permeability that they have trapped gas for millions
of years, and this gas becomes accessible only after an elaborate
process that involves drilling horizontal wells, fracturing the rock
with pressurized water, and propping the fractures open with
sand. Gas seeps from the region between each two consecutive
fractures into the highly permeable fracture planes and into the
wellbore, and it is rapidly produced from there.
The simplest model of horizontal wells consistent with this
setting is a cuboid region within which gas can diffuse to a set of
parallel planar boundaries. Fig. 1 illustrates the well as 1020
hydrofractures that are H 30m high and 2L 200m long,
spaced at distances of around 2d 100m. The fact that this is the
right starting point for these wells was recognized by Al-Ahmadi
et al. (9), and the diffusion problem in this setting has been
studied by both Silin and Kneafsy (10), and Nobakht et al. (11).
Examining Fig. 1 helps one to understand how gas production
evolves. When a well is drilled and completed, the ow of gas is
complicated and difcult to predict, particularly because the
water used to create it is back-produced. In practice, the resulting
initial transients last around 3 mo. After that time, gas should
enter a phase where it ows into the fracture planes as if coming
from a semiinnite region.
Signicance
Ten years ago, US natural gas cost 50% more than that from
Russia. Now, it is threefold less. US gas prices plummeted because of the shale gas revolution. However, a key question
remains: At what rate will the new hydrofractured horizontal
wells in shales continue to produce gas? We analyze the simplest model of gas production consistent with basic physics of
the extraction process. Its exact solution produces a nearly
universal scaling law for gas wells in each shale play, where
production rst declines as 1 over the square root of time and
then exponentially. The result is a surprisingly accurate description of gas extraction from thousands of wells in the
United States oldest shale play, the Barnett Shale.
Author contributions: T.W.P., F.M., and M.M. designed research, performed research,
analyzed data, and wrote the paper.
The authors declare no conict of interest.
This article is a PNAS Direct Submission. M.C. is a guest editor invited by the Editorial
Board.
Freely available online through the PNAS open access option.
See Commentary on page 19660.
1

To whom correspondence should be addressed. E-mail: patzek@mail.utexas.edu.

This article contains supporting information online at www.pnas.org/lookup/suppl/doi:10.

1073/pnas.1313380110/-/DCSupplemental.

PNAS | December 3, 2013 | vol. 110 | no. 49 | 1973119736

ENGINEERING

Natural gas from tight shale formations will provide the United
States with a major source of energy over the next several decades.
Estimates of gas production from these formations have mainly
relied on formulas designed for wells with a different geometry.
We consider the simplest model of gas production consistent
with the basic physics and geometry of the extraction process. In
principle, solutions of the model depend upon many parameters,
but in practice and within a given gas eld, all but two can be xed
at typical values, leading to a nonlinear diffusion problem we solve
exactly with a scaling curve. The scaling curve production rate
declines as 1 over the square root of time early on, and it later
declines exponentially. This simple model provides a surprisingly
accurate description of gas extraction from 8,294 wells in the
United States oldest shale play, the Barnett Shale. There is good
agreement with the scaling theory for 2,057 horizontal wells in
which production started to decline exponentially in less than
10 y. The remaining 6,237 horizontal wells in our analysis are too
young for us to predict when exponential decline will set in, but the
model can nevertheless be used to establish lower and upper
bounds on well lifetime. Finally, we obtain upper and lower bounds
on the gas that will be produced by the wells in our sample, individually and in total. The estimated ultimate recovery from our
sample of 8,294 wells is between 10 and 20 trillion standard
cubic feet.

Fig. 1. Horizontal well with 1020 hydrofracture stages spaced uniformly

along its entire length. The common fracture height is H, and the tip-to-tip
length of each fracture is 2L. The distance between the hydrofractures is 2d.
Gas ows into each fracture plane from both sides, and the permeability of
a hydrofracture is assumed to be innite in comparison to the effective
permeability of the rock matrix and natural fractures feeding gas into it.

Gas ows according to Darcys law through a system of microfractures, cracks, reopened natural fractures, faults, and failed
rock. This multiscale and loosely connected ow system is created by the high-rate hydrofracturing of shale rock. It is fed by the
rock matrix, where gas is stored (adsorbed) in very small pores.
It turns out that the gas effectively ows along paths that are
straight lines (hence, the setting is sometimes called linear
ow) perpendicular to the fracture planes. During the ow, the
initially high gas pressure diffuses toward the hydrofractures,
which are kept at a low pressure. This gas pressure diffusion
creates a gas production rate proportional to the inverse of the
square root of time on production.
At some point in time, gas ow causes the pressure along the
midplane between the hydrofractures to drop below the original
reservoir pressure, and gas production slows down relative to the
square-root-of-time behavior. We call the time when this happens the interference time. Eventually, the gas is so depleted
that the amount coming out per time is proportional to the
amount of gas remaining. This is the classic condition for exponential decay. Thus, after a long enough time, the rate of
production declines exponentially. The pressure-dependent
coefcient describing the diffusion of gas pressure is called the
hydraulic diffusivity of gas. Physically, it is unrelated to the
molecular diffusion coefcients.
The more closely spaced the hydrofractures are, the higher will
be the initial rate of gas production but the more quickly will the
interference time be reached. These intuitive considerations are
consistent with the mathematical results of Silin and Kneafsy
(10) and Nobakht et al. (11), and with the analysis that follows.
Results
Model. Hydraulic fracture in horizontal wells creates a network of

cracks in rock that was previously impermeable, allowing gas to

move. The true geometry is very complicated. Here, we explore
the possibility that it sufces to treat the rock surrounding a well
as a cuboid region in which the permeability is greatly enhanced
over the surrounding area but is uniform. We focus rst on the
single region depicted in Fig. 1 (Lower). The solution of a problem with N hydrofractures is obtained trivially by multiplication once the problem for a single region bounded by a pair of
hydrofractures has been solved.
In our initial treatments of the problem, we neglected variations in the hydraulic gas diffusivity as gas pressure changes. This
made it possible to solve the problem with closed-form expressions, but the results did not match well with experimental data.
The thermodynamic properties of natural gas must be treated
properly, as recognized by, for example, Kelkar (7). Natural gas
19732 | www.pnas.org/cgi/doi/10.1073/pnas.1313380110

is not an ideal gas; its compressibility and viscosity depend upon

its molar composition and vary strongly with temperature and
pressure. Failing to take variable gas properties into account led
to errors on the order of 50%.
We remark on some additional effects that we do not include.
Injecting water into the gas-bearing rock leads to interactions
between gas and water termed spontaneous imbibition. Although
these effects are amenable to precise analysis (12) and have experimentally measurable consequences (13), we are able to neglect them because we discarded the rst 3 mo of gas production,
when these and other transient effects are most pronounced. In
addition, as the pressure falls in a reservoir, gas adsorbed in the
rock may escape, producing additional contributions to the gas
ow. This process is described by the Langmuir desorption isotherm. In the particular eld studied in this paper, we have carried
out a detailed analysis and found this effect to be negligible, although it might not be so in other cases. Finally, desorption and
ow of gas in unfractured shale have nonlinear properties at the
microscopic level (14). We can neglect this phenomenon because
gas transport is dominated by the effective properties of a fracture
network, and the empirical evidence presented below shows that
the net effect is pressure diffusion at an enhanced rate in a
homogeneous medium.
Thus, we arrive at a specic nonlinear pressure diffusion problem to solve: It involves gas alone, permeability is uniform but
enhanced in a cuboid volume, the experimental equation of state
for gas is treated exactly, and spontaneous imbibition and desorption are neglected. The precise formulation and exact solution
of the diffusion problem are contained in Methods, and we describe
only the main results here.
Two planar hydrofractures in a well separated by distance 2d
interfere with one another after a characteristic interference
time (15), which we dene as
= d2 =i ;

[1]

where i is called the hydraulic diffusivity. It is related to the

permeability of the rock k by


k


;
[2]
i =


Sg g cg

Initial reservoir p;T

where is porosity, Sg is the fraction of pore space occupied by

gas, g is the gas viscosity, and cg is the gas compressibility. In Eq.
1, is a constant dened at the initial state of the reservoir. It
does not depend on the instantaneous gas pressure that varies in
space and time as the reservoir is depleted. This does not mean
that our solution relies on any approximation where quantities
are xed at reservoir values. It simply means that we adopt a time
unit that is dened in terms of the initial reservoir properties; the
nal results take a particularly simple form when we do so.
We measure time in units of , dening a dimensionless time as
~t t=:

[3]

Next, let m be the cumulative production of gas mass from a horizontal well with N hydrofractures (Fig. 1), and let M be the
original mass of gas contained in the reservoir volume drained
by this well. The exact solution of the model for cumulative gas
production is given by a dimensionless recovery factor (RF):

RF ~t = m=M:

[4]

We compute the RF by solving a particular boundary value

problem (Methods) and plot it in Figs. 2 and 3. Although this RF
is obtained from a numerical solution, it is a scaling function
that, for practical purposes, provides the benets of insight and
convenience commonly associated with closed-form analytical
Patzek et al.

where ~t0 is the dimensionless time necessary to extinguish the

initial transients in gas ow.
The constant depends on the gas composition and temperature, as well as on the limiting pressures, pi and pf. For the wells
we present here from the Barnett Shale, we set it equal to a
typical value of 0.645. Table S1 shows that it varies rather little as
the limiting pressures range over realistic values. Once the scaled
time ~t reaches 1, the growth in gas recovery slows, and it eventually reaches a plateau, which describes the maximum recovery
possible for the given problem parameters. The way this slowing
down occurs depends in detail upon the thermodynamics of gas
expansion, the reservoir permeability, and the initial and nal
pressures in the reservoir. Eventually, as also shown in SI Text,
production declines exponentially.
As a rst illustration of Eqs. 4 and 5, suppose one knows the
original gas in place, M. After transients of the rst few months

SEE COMMENTARY

solutions. To describe essentially all wells in the Barnett Shale,

one has only to rescale this function in the time and gas production coordinates. In SI Text, we provide a spreadsheet (Dataset
S1) in which the function is tabulated for convenience.
Our nearly universal solution to the boundary value problem
depends, in principle, upon the initial state of the reservoir, pi ; T,
the well owing pressure, pf, and gas composition, y, although it
is independent of the details of the well geometry and the hydraulic diffusivity. In practice, pi, pf, T, and y can be set to typical
values within a given shale gas play. For dimensionless times ~t
much less than 1, we show in SI Text that the solution takes a
particularly simple intermediate asymptotic form:
  q
[5]
RF ~t ~t; for ~t0 < ~t  1;

Fig. 3. Comparison of 8,294 wells with scaling function. (A) Time history
of 2,057 wells in the Barnett Shale, scaled so as to t our scaling function
(initial reservoir pressure of 3,500 psi and well owing pressure of 500 psi),
for which the dimensionless time ~t starts below 0.25 and reaches 0.64 or
more. The burnt orange curves give the scaled production of each well,
and the black curve is the scaling function. Overall agreement is satisfactory. (B) Time history of 6,237 wells in the Barnett Shale for which the
scaled maximum time comes out as ~t max < 0:64 (burnt orange). These wells
are too young to trust our estimate of the interference time ; therefore,
we simply compare them with a square root function (black line). Time is
scaled by the maximum time tmax reached for each well, and production m
p
is scaled by K tmax .

Fig. 2. Cumulative production and production rate from scaling theory. (A)
Dimensionless RF RF~t vs. dimensionless time computed from the scaling
solution (black) compared with ve typical wells (burnt orange). The fracture
pressure pf is 500 psi, and the initial reservoir pressure pi is 3,500 psi. (B)
Dimensionless well production rate RF~t=~t vs. dimensionless time (black)
under the same conditions compared with the same ve typical wells (burnt
orange). Production rates of individual wells are noisy, although cumulative
production matches the scaling function well. Because the production rate
becomes linear on a semilog plot, production decline is exponential for
t= = ~t  1.

Patzek et al.

Therefore, to estimate the time after which well production

declines exponentially, measure K from the rst year of production, estimate M from the well geometry, and insert = 0:645,
and follows from Eq. 6.
However, the practical difculty we face with gas production
from the hydraulically fractured horizontal wells is greater than
this example indicates. Neither the total mass of gas in place nor
the time scale for interference to begin is known with any precision. The original mass of gas in place is uncertain, mainly
because the effective hydrofracture length, 2L, and the number
of active hydrofractures are uncertain. The time to interference
is uncertain because the hydrofracturing process greatly increases the effective permeability k of the rock in the vicinity of the
well; laboratory values of k obtained from core samples are on
the order of nanodarcies (16), whereas accounting for observed
well production requires effective values of k on the order of
100-fold greater.
Thus, we arrive at the following question: Can we extract
enough information from existing eld production data to estimate both the interference time and the original gas in place
M at the same time? In the early stages of gas production,
when
p
t0 < t  , the production rate declines purely as 1= t and and
M are impossible to determine separately. Wells delivering a
small ultimate amount of gas at a relatively high rate cannot be
distinguished from those where lower permeability rock or a small
PNAS | December 3, 2013 | vol. 110 | no. 49 | 19733

ENGINEERING

of productionphave
subsided, cumulative production takes the

form mt K t. The constant K is obtained by tting a curve of

this form to the measured cumulative production. Then,
p
p
M t= = K t = M=K2 :
[6]

number of hydrofractures deliver ultimately larger quantities of

gas at a relatively lower rate. Only the onset of interference between adjacent hydrofractures makes it possible to disentangle the
two scenarios.

Summing up the production of the 8,294 wells in our sample,

we obtain the lower and upper bounds on cumulative production
over time, as shown in Fig. 5.

Comparison with Field Data. We display the dimensionless RF in

i) We have found the minimal ingredients that sufce to model

thousands of wells in the Barnett Shale with acceptable accuracy. The geometry of each well is a cuboid volume with a uniform array of absorbing boundaries. Between those boundaries,
rock permeability is enhanced above laboratory values but is
constant. Spontaneous imbibition can be neglected, but the gas
equation of state must be treated realistically. Gas desorption
is also negligible in the Barnett Shale but not elsewhere (e.g.,
in the Fayetteville shale). The scaling curve we nd as a result
provides surprisingly good agreement with all wells that can
reasonably be analyzed in the Barnett Shale.
ii) Inserting characteristic values into Eqs. 1 and 2, one deduces
rock permeability k of 50 nanodarcies for of 50 y and 500
nanodarcies for of 5 y. These values of permeability are 20to 200-fold larger than the values of a few nanodarcies found
for shale core samples in laboratory experiments (16). This
enhanced permeability must result from the hydrofracturing
process. Many processes could be involved, including the
reopening of preexisting fracture networks.
iii) Cumulative gas production follows a nearly universal function scaled by two parameters, interference time and mass
of gas in place M.
iv) For 2,057 of the horizontal wells in the Barnett Shale, interference is far enough advanced for us to verify that wells
behave as predicted by the scaling form. The typical interference time in these wells is around 5 y.
v) For 6,237 additional horizontal wells, no signicant deviation from cumulative production growing as the square root
of time is observed; these wells are too young to show evidence of interference. We provide upper and lower bounds
on time to interference and original gas in place for each of
these wells. The median lower bound on time to interference
is 5 y, and the median upper bound is 100 y. The bounds on
gas in place are somewhat tighter; the mean of the lower
bounds is 1 billion standard cubic feet (Bscf), and the mean
of the upper bounds is 7 Bscf. The lower bound on cumulative
production from the wells we analyzed is 10 trillion standard
cubic feet (Tscf) extracted over the next 10 y, whereas the upper
bound is more than 20 Tscf that will continue to be recovered,
at declining rates, over the next 50 y. By way of comparison,
a recent estimate of the total gas production from all wells
to be drilled in the Barnett Shale by 2050 is 40 Tscf (17, 18).
vi) The contributions of shale gas to the US economy are so
enormous (SI Text) that even small corrections to production
estimates are of great practical signicance.

Fig. 2. To illustrate its correspondence to data, we begin with

a sample of 66 wells hand-selected by an experienced reservoir
engineer as examples of good wells. In 5 of them, we nd evidence of interference, meaning
that cumulative production is not
p
acceptably t simply by K t. They do, however, t the full scaling
curve well, as we show with a graph of the cumulative production
and production rate of these ve wells in Fig. 2.
We then proceed to a more comprehensive study. We obtained data for 16,533 wells in the Barnett Shale, and from them,
we selected the 8,807 horizontal wells that had operated continuously for 18 mo or more and had not been recompleted (i.e.,
the hydrofracturing process was not repeated to increase production). We allow ourselves only two tting parameters on a
per-well basis, horizontal and vertical scale factors, which correspond physically to the interference time, , and the original
mass of gas in place, M. Details of the tting process are contained in SI Text. We nd 2,057 horizontal wells for which the
dimensionless time ~t starts with a value less than 0.25 and reaches
a value greater than 0.64. These are the wells for which interference is sufciently advanced that it can be detected with an
average uncertainty in parameters of less than 20%. We plot the
RF of all these wells vs. scaled time and compare the results with
the predicted scaling function in Fig. 5 and Fig. S3. Most of the
wells show interference because the interference time is around
5 y, but a few of them have interference times of 10 y or more
(Fig. 4). The fact that production from these more than 2,000
wells falls so well on the predicted curve provides evidence that
the simple model we adopted is sufciently realistic to estimate
gas production in the future. We note that upper bounds on the
total mass of gas in place are available from measurements of
well geometry. As a check on our results, we show in Fig. S4 that
the volumes of gas M we calculate with our theory from production data are indeed less than these upper bounds.
We acknowledge that for any given well at particular points in
time, production is noisy for many reasons (Fig. 2B). However,
the cumulative production of individual wells falls remarkably
well on our scaling curve (Fig. 2A), as does the expected production of thousands of wells (Fig. 3A).
There are an additional 6,237 wells for which interference is
not yet visible, and which we say are in the square root decline
phase. We cannot calculate and M for these wells, but we can
make use of our theory to put upper and lower bounds on them.
We present these bounds in Fig. S5.

Fig. 4. Values of interference time and gas in place M for the 2,057 wells
in Fig. 3A. Error bars indicate two standard uncertainties. Maximum interference times here are around 10 y due to the fact that wells more than
10 y old are still rare; interference times of, say, 30 y will only be reliably
detected when wells are 19 y old or more.

19734 | www.pnas.org/cgi/doi/10.1073/pnas.1313380110

Discussion

Gas released by hydraulic fracturing can only be extracted

from the nite volume where permeability is enhanced. Exponential decline of production once the interference time has
been reached is inevitable, and extrapolations based upon the
power law that prevails earlier are inaccurate. The majority of
wells are too young to be displaying interference yet. The precise
amount of gas they produce, and therefore their ultimate profitability, will depend upon when interference sets in.
For the moment, it is necessary to live with some uncertainty.
Upper and lower bounds on gas in place are still far apart, even
in the Barnett Shale with the longest history of production.
Pessimists (4) see only the lower bounds, whereas optimists (19)
look beyond the upper bounds. A detailed economic analysis
based on the model presented here is possible, however, and is
being published elsewhere (17, 18, 20, 21). The theoretical tools
we are providing should make it possible to detect the onset of
interference at the earliest possible date, provide increasingly accurate production forecasts as data become available, and assist
Patzek et al.

p*

p dp
:
g Z g

[14]

SEE COMMENTARY

Zp
mp = 2

Here, p* is a reference pressure that will be set to pf. After differentiation of

Eq. 14 and cancelation of terms, one obtains the following nonlinear diffusion equation for gas pseudopressure:


Sg g cg mp
2 mp
1
mp
=
,
=
2
x
t
pm t
k

[15]

with
p = 

with rational decisions about how hydraulic fracturing should

proceed in light of its impact on the US environment and economy.

The initial condition for Eq. 15 is

mpx,t = 0 = mpi = mi :

mpx = 0,t = mpf = mf ,

g ug

h
i
Sg g + 1 a kg gas
=
3 ,
m s
t

[7]

where ug is the Darcy (supercial) velocity of gas, Sg = 1 Swc is gas saturation (with Swc being the connate water saturation), g is the free gas density,
a is the adsorbed gas density (kilograms of gas per cubic meter of solid), and
is the rock porosity.
By applying Darcys law to the linear, horizontal ow of gas, we can
substitute
k p
ug =
g x

g p
kg p
g p
Sg
+ 1 a
:
x g x
p t
g p t

[9]

m


= 0:
x
x=d

Mg p
,
Zg RT

~t = t=;
~=
m

cg =

1 g
g p


T =const

1 1 Zg
:
=
p Zg p

[11]

We dene Ka p,T as the differential equilibrium partitioning coefcient of

gas at a constant temperature (e.g., ref. 22):

Ka =

a
g


:

[12]

T=const

By inserting Eqs. 11 and 12 into Eq. 9, the general nonlinear equation of

transient, linear, and horizontal ow of gas is obtained:
!


kg p
p
= Sg + 1 Ka cg g :
x g x
t

[13]

This nonlinear differential Eq. 13 can be simplied by introducing the

Kirchhoff integral transform of gas pressure after Al-Hussainy et al. (23),
which, in the present context, is also called the real gas pseudopressure:

Patzek et al.

d2
i

x~ = x=d



1

cg p g Zg =p2 mx,t:
i
2

[20]

Here, the subscript i refers to the quantities at the initial reservoir pressure
pi and temperature T.
Consider the linear ow of gas into a transverse planar hydrofracture
of height H and length 2L, and separated by distance 2d from the next
hydrofracture planes, as depicted in Fig. 1. The scaled transport equation is
~
~ 2 m
m
,
=
~t i ~
x2
 


~ x~,~t = 0 = m
~ i x~ ,
m

[10]

where Zg p,T,y is the compressibility factor of gas, Mg is the pseudomolecular mass of gas, R = 8,314:462 J/kmol-K is the universal gas constant, and
T is a constant temperature of the reservoir.
The isothermal compressibility of gas is dened as

[19]

Eq. 15 is most useful in a scaled form. We dene dimensionless time,

distance, and pseudopressure by

The gas density is related to its pressure and temperature through an equation
of state for real gases:
g =

[18]

where the hydrofracture pseudopressure, mf, might be a constant or a

slow function of time. At the midpoint between two fractures, one has
by symmetry

[8]

and obtain the following nonlinear partial differential equation:

[17]

Note that mi is a constant only in a virgin reservoir. During refracturing, it

will vary with the distance to the old hydrofractures.
We apply this equation to a nite region between two fractures, as shown
in Fig. 1 (Lower):

Methods
We begin with an expression for mass balance of gas owing in a porous rock:

[16]

 
~ x~,~t = 0
m

for x~ = 0

~ x = 0
m=~

for ~
x = 1:

[21]
ENGINEERING

Fig. 5. Upper and lower bounds on cumulative production from 8,294 wells
in our sample. Vertical wells are excluded from the analysis, whereas twofold
more wells will ultimately be drilled; thus, the upper bound is not an upper
bound on the whole eld.

k

:
Sg + 1 Ka g cg

and

Our approach is somewhat more general than that of Silin and Kneafsey
(10) because we do not require any particular equation of state for natural
gas and do not use the more limited p2 formulation (8). The mp and p2
solutions are equivalent only if p=g Zg is a linear function of pressure;
however, generally, it is not (ref. 8, pp. 254255). The price we pay is that our
model must be solved numerically, but the cost is just a couple of seconds of
delay before the full solution is computed on an average laptop. For the
Barnett Shale, we use the values of well owing pressure pf = 500 psi and
initial reservoir pressure pi = 3,500 psi.
The supercial velocity of gas owing into the right face of the hydrofracture at the origin is


k p


uf =
:
[22]
f x
x=0
The mass ow rate into this fracture is
_ = 2HLf uf :
m

[23]

Using Eq. 22,

PNAS | December 3, 2013 | vol. 110 | no. 49 | 19735

_ = 2HLf
m

k p


:
f x
x=0

[24]

Next, we replace the pressure with the real gas pseudopressure in Eq. 14:
_ = 2HL
m

k Mg m


:
2 RT x
0

[25]

The partial derivative can now, in turn, be rewritten with use of the scaled
pseudopressure from Eq. 20, and the permeability, k, can be eliminated
in favor of the gas diffusivity, i , and the characteristic interference
time, .
Let M be the total mass of gas contained originally in the reservoir within
the volume 4LHd between two consecutive hydrofractures, M 4i LHdSg ;
then the gas ow rate into the fracture plane at the origin takes the nal form:
_ =
m

~

M m

,
2 ~
x
0

[26]

~ x~j0 ~t depends only on gas composition, the initial

where the function m=
~ has been
and fracture pressures, and reservoir temperature. The scaling of m
devised so that this relation is exact.
The total ow into each pair of hydrofractures is twice that in Eq. 26. More
generally, when there are N fracture stages, as depicted in Fig. 1, and we
include the contribution to mass ow from the exterior faces of the left- and
right-most hydrofractures, the original mass in place is
M N + 14i LHdSg ,

[27]

and the total mass transport out of the reservoir is given by

_ =
m

~

M m

:
x~
0

[28]

1. Birol F (2012) Golden Rules for a Golden Age of GasWorld Energy Outlook Special
Report on Unconventional Gas (IEA, Paris).
2. Osborn SG, Vengosh A, Warner NR, Jackson RB (2011) Methane contamination of
drinking water accompanying gas-well drilling and hydraulic fracturing. Proc Natl
Acad Sci USA 108(20):81728176.
3. Vidic RD, Brantley SL, Vandenbossche JM, Yoxtheimer D, Abad JD (2013) Impact of
shale gas development on regional water quality. Science 340(6134):1235009.
4. Hughes JD (2013) Energy: A reality check on the shale revolution. Nature 494(7437):
307308.
5. Arps JJ (1945) Analysis of decline curves. American Institute of Mining Engineers
Petroleum Transactions 160:228247.
6. Fetkovich M (1980) Decline curve analysis using type curves. Journal of Petroleum
Technology (June):10651077.
7. Kelkar M (2008) Natural Gas Production Engineering (PennWell, Tulsa, OK).
8. Dake LP (1978) Fundamantals of Reservoir Engineering, Developments in Petroleum
Science (Elsevier, Amsterdam), Vol 8.
9. Al-Ahmadi HA, Almarzooq AM, Wattenbarger RA (2010) Application of linear ow
analysis to shale gas wellsField cases. (Society of Petroleum Engineers) 130370:110.
10. Silin DB, Kneafsey TJ (2012) Gas shale: Nanometer-scale observations and well modeling. Journal of Canadian Petroleum Technology 51:464475.
11. Nobakht M, Mattar L, Moghadam S, Anderson DM (2012) Simplied forecasting of
tight/shale-gas production in linear ow. Journal of Canadian Petroleum Technology
51:476486.
12. Schmid KS, Geiger S (2013) Universal scaling of spontaneous imbibition for arbitrary
petrophysical properties: Water-wet and mixed-wet states and Handys conjecture.
J Petrol Sci Eng 101:4461.
13. Dehghanpour H, Zubair HA, Chhabra A, Ullah A (2012) Liquid intake of organic
shales. Energy Fuels 26:57505758.

19736 | www.pnas.org/cgi/doi/10.1073/pnas.1313380110

Here, we are treating the left- and right-most exterior hydrofracture faces
approximately, as extensions of the wellbore length by d at each end. The
reason it is not appropriate to treat the two ends as semiinnite is that
without the great enhancement of permeability brought about by the
hydrofracturing process, gas transport is negligible. Our assumption is that
volumetric rock damage extends beyond the ends of the two last fractures
for characteristic distance d.
Integrating Eq. 28 with respect to the dimensionless time ~t gives the
nal result
 
m
= RF ~t , where
M

 
RF ~t

Zt
0

~
 
m
d~t

~t :
x~ 0

[29]

The initial boundary value problem (Eq. 21) is solved numerically with
an efcient fully implicit solver and a sequential implicit solver. The rst
solver has been implemented in Python, and the second has been implemented in MATLAB (MathWorks). Accurate numerical solutions can be
obtained in both cases within a few seconds on an average laptop. Essential
properties of the result are revealed by exact solution of simplied
equations, depicted in Fig. S6.
ACKNOWLEDGMENTS. We thank John Browning for help in estimating
reserves, for his deep insights into well performance in the Barnett Shale,
and for help in selection of well-behaved groups of wells. We thank D. Silin,
S. Bhattacharaya, and R. Dombrowski for detailed comments on the manuscript. The gas production data were extracted from the IHS Cambridge Energy
Research Associates database, licensed to the Bureau of Economic Geology.
This paper was supported by the Shell Oil Company/University of Texas at
Austin project Physics of Hydrocarbon Recovery, with T.W.P. and M.M.
as coprincipal investigators, and the Bureau of Economic Geologys Sloan
Foundation-funded project The Role of Shale Gas in the U.S. Energy Transition: Recoverable Resources, Production Rates, and Implications. M.M.
acknowledges partial support from the National Science Foundation Condensed Matter and Materials Theory program.
14. Monteiro PJM, Rycroft CH, Barenblatt GI (2012) A mathematical model of uid and
gas ow in nanoporous media. Proc Natl Acad Sci USA 109(50):2030920313.
15. Patzek TW, Silin DB, Benson SM, Barenblatt GI (2003) Non-vertical diffusion of gases
in a horizontal reservoir. Transport in Porous Media 51:141156.
16. Vermylen JP (2011) Geomechanical studies of the Barnett Shale, Texas, USAPhD Thesis
(Stanford University, Palo Alto, CA). Available at https://pangea.stanford.edu/departments/
geophysics/dropbox/SRB/public/docs/theses/SRB_125_MAY11_Vermylen.pdf. Accessed
October 23, 2013.
17. Browning J, et al. (2013) Barnett Shale reserves and production forecast: A bottom-up
approach. Part I. Oil and Gas Journal. Available at http://www.ogj.com/articles/print/
volume-111/issue-8/drilling-production/study-develops-decline-analysis-geologic.html.
Accessed October 23, 2013.
18. Browning J, et al. (2013) Barnett Shale reserves and production forecast: A bottom-up
approach. Part II. Oil and Gas Journal. Available at http://www.ogj.com/articles/print/
volume-111/issue-9/drilling-production/barnett-study-determines-full-eld-reserves.htm.
Accessed October 23, 2013.
19. Potential Gas Committee (2013) Potential Supply of Natural Gas in the United States
(December 31, 2012), press release. Available at potentialgas.org/download/pgc-pressrelease-april-2013-slides.pdf. Accessed October 23, 2013.
20. Glen G, Browning J, Ikonnikova S, W TS (2013) Barnett cell economics. Energy
60:302315.
21. Ikonnikova S, Browning J, Horvath S, Tinker SW (2013) Well recovery, drainage area,
and future drillwell inventory: Empirical study of the Barnett Shale gas play. SPE
Reservoir Eval Eng, in press.
22. Cui X, Bustin AMM, Bustin RM (2009) Measurements of gas permeability and diffusivity of tight reservoir rocks: Different approaches and their application. Geouids
9:208233.
23. Al-Hussainy R, Ramey HJJ, Crawford PB (1966) The ow of real gases through porous
media. AIME Petrol Transactions 237:624636.

Patzek et al.

Supporting Information
Patzek et al. 10.1073/pnas.1313380110
SI Text
Here, we describe the role of natural gas in the US economy, describe the details of our data-tting procedures, provide additional
evidence concerning the correspondence of the dimensionless recovery factor (RF) and well production data, provide upper and
lower bounds on gas production from wells not yet showing interference, demonstrate why production declines as the square root
of time early on and exponentially later, and tabulate the coefcient
for a variety of reservoir and well owing pressures to demonstrate how little it varies. In a separate spreadsheet (Dataset S1), we
provide tabulations of the dimensionless RF.
Impacts of Natural Gas Production in the United States
As shown in Fig. S1, the United States has managed to maintain
gas production at an essentially at rate for 40 y after a 1974 peak
of gas production that closely followed the 1971 peak in oil
production. No other country has done the same. Novel technology,
most recently the massively hydrofractured horizontal wells in
shale plays, has played a crucial role in maintaining US gas
production at or slightly above its 1974 level. As a result of
plentiful gas production, US gas prices have recently been a
fraction of the typical world gas prices, injecting over half a
trillion current dollars into the US economy (Fig. S2). This
second stimulus package in the United States has been almost
invisible to the public.
Data Analysis and Fitting Procedure
We analyzed the 16,533 wells in our dataset for the Barnett Shale
through the following steps:
i) We eliminate all wells that have been recompleted, all vertical wells, and all months from each wells time history with
production of zero. At this point, 11,566 wells remain.
ii) We eliminate all wells with less than 18 mo of total production. Now, 8,807 wells remain.
iii) For each well, we have a measurement of production per
day for a sample of days each month. We convert to production per month by multiplying by 30.4.
iv) The rst 3 to 4 mo of production are typically noisy and
sporadic, particularly because hydrofracturing water is still
being back-produced. Therefore, from the time series for
each well, we construct a slightly modied one. We label the
starting time of this new series 2.5 (mo), and we assign to it
the cumulative production of the rst 4 mo. There is no
further processing: For each new nonzero gas volume produced in a given month, time increases by 1 mo and cumulative production increases by the production of that month
until the data end.
v) We use the LevenbergMarquardt least-squares minimization (lmt) Python package to nd the values of the interference time and gas in place M that best t our scaling
curve to the measured cumulative production. In particular,
we minimize the objective function mt MRFt=, where
mt is the measured cumulative production data. Although
the reservoir pressure pi is not the same for all wells in the
Barnett Shale, we make use only of the curve corresponding
to pi = 3;500psi, pf = 500psi in this paper. We also conducted the analysis allowing pi to vary according to measured
pressure variations, but the difference was negligible.
vi) We had to guard against a number of artifacts that could
produce spurious agreement between well histories and the
scaling function. For wells with very short histories, uctuations
Patzek et al. www.pnas.org/cgi/content/short/1313380110

in production could produce shapes that mimicked large

segments or small portions of the scaling curve. We eliminated these matches by requiring at least 18 mo of production and requiring well histories to traverse a considerable
portion of the scaling function rather than running tangent to
it over a brief interval. In total, 513 wells were eliminated in
this way, leaving 8,294 wells. In no case did we eliminate any
well history because its t to the scaling function was poor.
Our plots include all wells except for those whose history was
too short to include for the reasons we describe.
vii) An advantage of the lmt package is that it includes careful
estimates of the uncertainty of parameters. We used the routine conf_interval to improve the estimates. Our scaling curve
RF~t is practically indistinguishable from a square root until
the argument ~t approaches 1. Thus, it is impossible to obtain
a useful estimate of the interference time and gas in place
unless interference has become visible. To select a quantitative
criterion, we found that the average uncertainty of the parameters and M was more than 20% unless the scaled age of the
well, tmax =, was greater than around 0.64. Accordingly, we
used the condition ~tmax > 0:64 to divide wells into two groups.
Careful analysis of the magnitude of the objective function
as a function of the tting parameters M and indicates
that these uncertainty estimates are somewhat too tight.
There is a narrow valley in -M space, where the function
varies very slowly on large scales but has relatively rapidly
varying local minima on shorter scales. The uncertainty estimates could probably be improved and would become somewhat larger by taking into account these adjacent local
minima, but we have not yet done this.
We found a small number of wells with interference times
less than 1 y, probably because of the high-permeability channels leading to interactions of hydrofractures or hydrofracture branches that are very close to each other.
viii) For wells with ~tmax > 0:64 and ~tmin < 0:25, we estimate and
M. The result appears in Fig. 4. The interference times are
short as they must be; because such horizontal wells are, at
most, 14 y old and most are younger, it is impossible to detect
interference times of much more than 10 y, and the typical
measured interference time is only 5 y.
ix) For wells with ~tmax < 0:64, the estimates of and M are
too uncertain to be useful, but we can provide bounds.
The lower bound on is obtained from the observation
that interference would be visible if ~tmax were greater than
0.64. Because it is not visible, one must have ~tmax = tmax =
< 0:64 > tmax =0:64.
x) For wells that do not show interference, the constant K from
Eq. 6 can be determined from the data and the lower bound
on can be converted into a lower bound on M.
xi) A reasonable upper bound on M can be obtained from data
on the size of each well and the thickness of the mudrock
layer (1). The upper bound on M can be converted into an
upper bound on using Eq. 6.
Additional Checks on RF
We provide two additional checks on the scaling function formalism. First, we check whether production rates for wells showing
evidence of interference do indeed decline exponentially. Evidence
is provided in Fig. S3. Although rates are very noisy, with many
months where production drops by a large factor and recovers,
as well as occasional excursions above the predicted curve, overall
rates decline in accord with the predicted exponential.
1 of 6

Second, we check in Fig. S4 whether the measurement of the

original gas in place M obtained from the scaling formalism is
bounded above by estimates obtained from data on the extent of
the well. The two measurements are coming from separate data
sources, so the comparison is a strong test of both data integrity
and the sense of our formalism. The estimated gas in place of
virtually every well lies below the upper bound. In addition, the
information coming from the scaling formalism is not redundant
because the measurement of M obtained in this way is usually
considerably less than the upper bound.
Upper and Lower Bounds on Gas Production from Wells in
Square Root Phase
Fig. S5 provides four additional pieces of information for the wells
that show no evidence of interference. In the Fig. S5 (Upper Left),
we provide a lower bound on the interference time . This lower
bound is obtained by noting that interference becomes evident
when ~t reaches 0.64, so if interference is not evident, the interference time must be at least 1.6 times larger than the
current life of the well. From this estimate one obtains a lower
bound on the gas in place for each well, since Eq. 6 and the
known value of K for each well turns a lower bound on into
a lower bound on M (Fig. S5, Upper Right). Fig. S5 (Lower Right)
displays an upper bound on the original gas in place M obtained
by using the measured thickness of the mudstone source rock of
each well, and the length of the well. From the upper bound on
M one obtains through Eq. 6 an upper bound on , shown in the
lower left. This bound on is not very tight. There is a peak at
around 30 y, but a long tail stretching into the hundreds of years.
We think it is impossible that wells will last this long before
beginning to interfere, but they are simply too young to provide
evidence that interference will occur any sooner.
Asymptotic Analysis of Gas Production at Early Times
For a domain bounded on two sides, as in Fig. 1, the diffusion
equation (Eq. 21) must be solved numerically. However, if the
~ vanishes at x = 0 and
boundary conditions are changed so that m
~ m
~ i as x , then as pointed out by Crank and Henry (2),
m
this equation possesses an exact similarity solution.
To nd this solution, let
x
p :
t

[S1]

Then Eq. 21 becomes

~
~
1 m
2 m
=
:
2
i 2

[S2]

~
This equation is of the rst order in m=,
and the solution is
obtained in a straightforward manner. Dene F by
Z
F =

d exp

Z
0

d
:

[S3]

Then

An important result that can nevertheless be obtained from

it is that

~ 
m
1 mi
= p
:
[S5]
~x 0
~t F
Thus, mass transport due to this similarity
solution has the exact
p
property of decaying in time as 1= t. The coefcient of the decay
can only be determined through integrals over the complete spatial
solution.pCumulative
production is given by the time integral, and

goes as t.
For any given initial condition, solutions of the diffusion
equation (Eq. 21) in a semiinnite space tend toward the solution given in Eq. S4. This is why,
p after an early transient period,
decline of p
production
as 1= t and growth of cumulative pro
duction as t are universal for a time. This solution persists until
the onset of interference between consecutive hydrofractures.
Asymptotic Analysis of Gas Production at Late Times
When one waits sufciently long, pressure drops everywhere in
the reservoir until it hovers just above the well owing pressure. In
this late-time regime, the hydraulic diffusivity p can be replaced by the constant pf . With the simplication that the
hydraulic diffusivity is constant and the interference time and
scaled time ~t are dened in terms of , Eq. 21 can be solved
exactly with the same boundary conditions as before, and the
result is




X
2 2~
~ 
m
=2
e 2n + 1 t=4

~x 0
n=0

 p
p
erfc 3 2~t=4
p
4 2~t=4
|{z}

Hyperbolic or square root of time decline

1
2~
+ e t=4 + e9 t=4 :
2
|{z}
2~

Exponential decline

[S6]
The relative importance of the three terms in Eq. S6 is plotted in
Fig. S5. We make two points about the result.
i) If the only goal is to provide an accurate account of the longtime behavior, this computation shows self-consistently that
the decline rate is exponential just so long as the limit of as
p pf is well dened.
ii) One can instead use the computation as an approximate analytical description of the entire decline process. In this case,
instead of using = pf , one should use = p; that is, one
should use a hydraulic diffusivity characteristic of the average pressure in the reservoir. The resulting approximation
haspthe property of leading to a decline curve that goes as
1= t at early times, and declining exponentially at late times
just like the exact solution (Fig. S6). However, this approximation leads to errors on the order of 50%. No matter how
one tunes a constant
, one cannot get both the coefcient of
p
the original 1= t decline and the total gas recovered right.

This equation appears at rst to be an explicit expression in closed

form, but because is, in fact, m, it is actually an integral
equation whose solution must be determined self-consistently.

Tabulation of
Table S1 extracts the coefcient
p , describing the initial rise in
cumulative production as ~t from the dimensionless RF for a
variety of reservoir and well owing pressures. The main lesson is
that it varies rather little and can safely be taken to assume a
nominal value of 0.645 across the Barnett Shale.

1. Fu Q, et al. (2013) Log-based thickness and porosity mapping of the Barnett Shale play,
Fort Worth Basin, Texas: A proxy for reservoir quality assessment. AAPG Bull, in press.

2. Crank J, Henry ME (1949) Diffusion in media with variable properties. Transactions of

the Faraday Society 45:636650.

~
m
= mi

F
:
F

Patzek et al. www.pnas.org/cgi/content/short/1313380110

[S4]

2 of 6

Fig. S1. In the United States, gas production doubled about 10-fold, growing by a factor of 1,000. Natural gas production, associated with the fundamental
Hubbert oil peak in the United States, peaked in 1974. However, a second fundamental gas cycle has been created by producing new gas (increasingly deep
offshore, from Alaska, tight gas sands, coal-bed methane, and now shales) resources in the United States. As a result, after growing at 7% per year until 1974
(red line), gas production has remained at over the past 40 y (blue line). Recently, gas production in the United States has actually been increasing, mostly due
to shale gas production. Note that 1 exaJoule (EJ) 1 trillion standard cubic feet (Tscf) of gas. Data source: US Department of Energy, Energy Information
Administration.

Fig. S2. Because of shale gas production, natural gas prices in the United States have plummeted to one-fourth of the Russian gas price. This plot subtracts
the cumulative amount the United States has actually spent on natural gas from the amount it would have spent at world (Russian) prices. Since 1991, domestic
gas production in the United States has delivered $560 billion US dollars (blue line) ($250 billion in constant 1983 US dollars, red line) to the US economy. Cheap
natural gas has led to a decrease of US CO2 emissions, increased employment, and a renaissance of steel production and manufacturing. Cheap reliable energy
is fundamental to a healthy economy, but very few people recognize this truism. Natural gas prices are from the Index Mundi (www.indexmundi.com/
commodities/). The consumer price index is from the US Census Bureau.

Patzek et al. www.pnas.org/cgi/content/short/1313380110

3 of 6

Fig. S3.

Production rate for all qualifying horizontal wells in the Barnett Shale shows evidence of interference.

Fig. S4. Original gas in place is computed by tting M to well production data compared for each well, with an upper bound obtained from data on the
dimensions of the well.

Patzek et al. www.pnas.org/cgi/content/short/1313380110

4 of 6

Fig. S5.

Bounds on the interference time and the original mass of gas in place M for the wells from Fig. 3B in the square root phase.

Fig. S6. Relative importance of the three-term rate approximation in Eq. S6, summed to form the black curve. Note that for ~t 0:2, production declines as the
square root of time (red curve). For ~t 1, the production rate decline is exponential (pink and blue curves). For 0:2 ~t < 1, there is a transition from the square
root of time decline to the exponential decline. This plot was obtained from linear analysis. For nonlinear cases, it turns out that when the time to interference,
, is referenced to the initial reservoir conditions, similar values of ~t govern transitions between production decline regimes.

Patzek et al. www.pnas.org/cgi/content/short/1313380110

5 of 6

Table S1.

Scaling factor

pi
2,000
2,500
3,000
3,500
4,000
4,500

pf = 500

pf = 400

0.650
0.663
0.661
0.645
0.634
0.633

0.671
0.672
0.671
0.653
0.642
0.641

p
For scaled times p~t <
0:2, cumulative production increases as t . The dimensionless RF ~t and this table display values of for various initial
and well owing pressures for a reservoir temperature of 190 F. There is
no simple expression for , but it does not vary much.

Other Supporting Information Files

Dataset S1 (XLS)

Patzek et al. www.pnas.org/cgi/content/short/1313380110

6 of 6