1 s2.0 S0013794423006203 Main

Engineering Fracture Mechanics 292 (2023) 109662
Contents lists available at ScienceDirect
Engineering Fracture Mechanics

journal homepage: www.elsevier.com/locate/engfracmech
An implicit level set algorithm for hydraulic fracturing with a

stress-layer asymptote
A.V. Valov a,b ,∗, E.V. Dontsov c , A.N. Baykin a,b , S.V. Golovin a,b
a Lavrentyev Institute of Hydrodynamics SB RAS, Novosibirsk, 630090, Russia
b Novosibirsk State University, Novosibirsk, 630090, Russia
c
ResFrac Corporation, Palo Alto, CA 94301, USA
ARTICLE INFO ABSTRACT
Keywords: The capability to simulate a hydraulic fracturing process is an essential tool that can be used
Hydraulic fracturing to optimize treatment design and increase the efficiency of field operations. In most practical
Asymptotic solution cases, hydraulic fractures propagate in a multi-layered rock formation. As a result, there is
Level set algorithm
a need to incorporate the effect of such heterogeneities in fracturing models to achieve an
Contact algorithm
accurate prediction. To capture the layered structure of rocks, a hydraulic fracture simulator
typically requires a fine mesh, which leads to a drastic reduction in computational performance.
An alternative is to use more sophisticated models that are capable of providing reasonably
accurate predictions even on a relatively coarse mesh. In the case of fracture growth modeling,
the pivotal component of the simulation is a fracture front tracking algorithm that accounts for
the layered structure of the formation. Consequently, this paper aims to extend the established
Implicit Level Set Algorithm (ILSA) to account for the effect of multiple stress layers within the
tip asymptote. The enhanced front tracking algorithm involves the stress-corrected asymptote
that incorporates the influence of stress layers within the near-tip region. To further increase
the validity region of the stress-corrected asymptote, the stress relaxation factor is introduced,
and its accuracy is examined. The numerical algorithm is validated against the reference
semi-analytical solutions as well as experimental observations. In addition, we investigate the
sensitivity of the fracture geometry to mesh size to demonstrate that the front tracking algorithm
based on the stress-corrected asymptote retains its accuracy on a coarse mesh.
1. Introduction
Hydraulic fractures are tensile cracks that are often created in underground rock formations due to the injection of pressurized
fluid. The most common technological application of hydraulic fracturing is production enhancement in low and moderate-
permeability reservoirs by the creation of high-permeable pathways [1]. Other engineering applications of hydraulic fracturing
include preconditioning the rock mass with the aim of improving caveability and fragmentation for block-caving mining opera-
tions [2,3], enhancing the injectivity and capacity of geologic carbon storage reservoirs [4,5], and heat production from geothermal
reservoirs [6–8]. Hydraulic fractures also occur naturally as kilometer-long dikes that are driven by magma from deep underground
chambers to the Earth’s surface [9–11]. Many of the aforementioned applications require at least some level of understanding of
fracture behavior. Therefore, it is important to have accurate models to have the ability to predict the propagation of hydraulic
fractures.
∗ Corresponding author at: Lavrentyev Institute of Hydrodynamics SB RAS, Novosibirsk, 630090, Russia.
E-mail addresses: a.valov@g.nsu.ru (A.V. Valov), egor@resfrac.com (E.V. Dontsov), alexey.baykin@gmail.com (A.N. Baykin), s.golovin@g.nsu.ru
(S.V. Golovin).
https://doi.org/10.1016/j.engfracmech.2023.109662
Received 24 May 2023; Received in revised form 23 September 2023; Accepted 30 September 2023
Available online 11 October 2023
0013-7944/© 2023 Elsevier Ltd. All rights reserved.
A.V. Valov et al. Engineering Fracture Mechanics 292 (2023) 109662
The history of mathematical models for hydraulic fracturing has its origin in the middle of the last century. One of the pioneering
studies include Khristianovich–Zheltov–Geertsma–De Kler (KGD) model [12,13], Perkins–Kern–Nordgren (PKN) model [14,15], and
the radial or penny-shaped model [16,17]. The KGD model assumes that the fracture propagates under plane strain elastic conditions.
The PKN model considers a vertical planar fracture with fixed height propagating horizontally and assumes a state of plane strain in
each vertical cross-section. It also assumes that the fracture length is much larger than its height. The radial model is applicable only
in a homogeneous formation when the wellbore is along the direction of the minimum principal stress and has mostly mathematical
importance. In order to allow fracture height growth, Pseudo-3D (P3D) models have been developed [18,19], which are essentially
extensions of the PKN model that allow height growth. The P3D models have been extended to the enhanced P3D model [20], the
stacked-height model [21], and the layered P3D models [22,23].
With an increase in computing power, further developments have shifted towards more accurate, fully planar 3D models
(PL3D) [24–27]. Such models assume that the fracture footprint is planar and has an arbitrary shape. PL3D models involve coupling
between two-dimensional fluid flow and the fully 3D elasticity equation representing the non-local elastic response. Planar 3D
models are able to handle specific types of fractures that pseudo-3D models are unable to model. Planar 3D models have gained
wide popularity in engineering applications due to their high accuracy compared to pseudo-3D models and computational efficiency
compared to fully 3D models [28–30].
By further following the path of increasing complexity, fully 3D hydraulic fracturing models have shifted towards incorporating
poroelastic effects as well as simultaneous propagation and interaction of multiple fractures [31]. A phase field variational approach
to model fracturing is employed in [32]. The result was also coupled to a reservoir simulator and the effect of natural fractures
is investigated. A fully coupled three-dimensional finite element model for examining the growth of multiple hydraulic fractures
in a permeable poroelastic formation is presented in [33]. This model simultaneously accounts for laminar fluid flow within the
fracture, Darcy flow in the rock matrix, and poroelastic deformation of the rock. Finally, it is worth mentioning the study [34], in
which the effect of initial fracture orientation is investigated using the Generalized Finite Element Method (GFEM) for the case of
simultaneously propagating multiple hydraulic fractures.
Most hydraulic fracturing simulators utilize the propagation condition that is based on Linear Elastic Fracture Mechanics
(LEFM) [35,36], while some use the cohesive zone model [30,37]. However, several studies have employed the universal tip
asymptotic solution to locate the fracture front for the purpose of increasing the overall accuracy of the model [38]. One of the
pioneering works in this direction is [26], which introduced a robust and accurate approach to tracking the fracture front using an
Implicit Level Set Algorithm (ILSA). Extensions as well as alternative implementations can be found in [27,39–45]. In particular,
the study [44] incorporates the universal tip asymptotic solution [46] into the numerical scheme as a propagation condition. This
universal solution accounts for the simultaneous interplay of the effects of fracture toughness, fluid viscosity, and leak-off. The use
of the universal asymptotic solution leads to a more accurate numerical solution on a relatively coarse mesh, as was shown in [38]
for the case of a radial fracture.
Despite the fact that the tip asymptote accounts for multiple physical processes, its use introduces significant errors when the
fracture front crosses a stress layer. As shown in [47], the geologic stress layers have a dominant impact on the accuracy of fracture
geometry, and ILSA combined with the universal asymptote demonstrates a notable reduction in accuracy once the stress layers are
considered. Therefore, to obtain more accurate results, the effect of layers should be incorporated into the tip asymptotic solution.
The paper [48] considers such the propagation condition for a plane-strain fracture, where the tip solution that accounts for layers
is calculated numerically using a non-singular integral formulation. Recently, an alternative to ILSA for front tracking was proposed
in [49,50]. The Multi Layer Tip Element (MuLTipEl) [49,50] algorithm introduces the concept of the additional fictitious stress to
track the fracture front in lieu of the level set methodology and also accounts for the effect of layers within a fracture element.
To do the latter, MuLTipEl approach uses the concept of an apparent toughness in combination with the universal tip asymptotic
solution. The effect of thin layers is incorporated by calculating the effective values for stress and toughness within the tip element.
As is evident from the previous paragraph, two options have been employed to account for the effect of stress layers in the
propagation condition. The first option is to numerically solve for the tip asymptote, while the second method relies on the concept
of apparent toughness. The paper [51] proposes an alternative third method, in which the ordinary differential equation (ODE)
approximation of the original non-singular integral equation is considered. This option is more accurate compared to the apparent
toughness approach, and yet much more computationally efficient compared with the numerical solution. Therefore, the goal of this
paper is to incorporate this ODE-based stress-corrected asymptotic solution into the ILSA algorithm and to evaluate the resultant
improvement for the cases of multiple stress layers in the rock formation.
The paper is organized as follows. Section 2 outlines the governing equations for a planar hydraulic fracture under consideration.
Section 3 provides motivation for using the tip asymptotic solution with stress layers, briefly summarizes the approaches to solving
the layer crossing problem, and also contains details of the implementation into the ILSA algorithm. Then, Section 4 evaluates the
accuracy of the tip asymptotic solution and provides corrections to increase the validity region of the tip asymptotes. Section 5
presents a series of numerical examples validating the developed algorithm against reference solutions and experimental data.
Section 6 provides a sensitivity analysis of the hydraulic fracture geometry to mesh size and illustrates the importance of using
the stress-corrected tip asymptote. The details of the numerical scheme and tip leak-off calculation are presented in Appendices A
and B.
2
Fig. 1. Scheme of a planar hydraulic fracture with a footprint 𝐴(𝑡) bounded by the fracture front 𝐶(𝑡) in a layered media. The point source 𝑄(𝑡) is located at
the origin. The coordinate indicating the distance from a point that lies within the fracture boundary curve 𝐶(𝑡) to the fracture front is denoted by 𝜁 and the
normal front velocity is 𝑉 .
2. Mathematical formulation
We consider the problem of a planar 3D hydraulic fracture propagating in the (𝑥, 𝑦) plane due to the injection of an incompressible
viscous fluid with the rate 𝑄(𝑡) through a point source located at the origin, see Fig. 1. The minimum in situ compressive stress
field 𝜎ℎ (𝑦) is oriented perpendicular to the fracture plane. It is assumed that the fluid front coincides with the fracture front since
the fluid lag is negligible for typical depths encountered in field hydraulic fracturing operations [52,53]. Fluid leak-off into the
surrounding rock is described by Carter’s model [54], fluid rheology is Newtonian, and the effect of hydrostatic pressure is neglected
for simplicity. The formation is assumed to be elastically homogeneous, while the stress is allowed to vary discontinuously from
one layer to another. Other parameters, such as toughness and leak-off are assumed to be homogeneous as well for simplicity. We
refer to [26,44,55,56] for a more comprehensive description of the underlying assumptions that are typically made in such models.
We are looking for the solution for the fracture footprint 𝐴(𝑡) and its boundary or front 𝐶(𝑡), the fracture width 𝑤, and the fluid
pressure 𝑝 as a function of the spatial coordinates (𝑥, 𝑦) and injection time 𝑡.
To simplify mathematical expressions, it is convenient to introduce the following scaled parameters
( )1∕2
𝐸 2
𝐸′ = , 𝐾′ = 4 𝐾Ic , 𝐶 ′ = 2𝐶𝐿 , 𝜇′ = 12𝜇,
1 − 𝜈2 𝜋
where 𝐸 is the Young’s modulus, 𝜈 is the Poisson’s ratio, 𝐾Ic is the fracture toughness, 𝐶𝐿 is Carter’s leak-off coefficient, and 𝜇 is
the fluid viscosity.
2.1. Elasticity equation
The elasticity equation relates the width 𝑤 of a planar fracture to the normal component of the traction vector 𝑇𝑛 applied to
the fracture surface. According to the displacement discontinuity method [57,58], for the single planar fracture the result can be
written in terms of the hypersingular integral equation
𝐸′ 𝑤(𝑥′ , 𝑦′ , 𝑡) d𝑥′ d𝑦′
𝑇𝑛 (𝑥, 𝑦, 𝑡) = 𝜎ℎ (𝑦) − . (1)
8𝜋 ∫𝐴(𝑡) (𝑥′ − 𝑥)2 + (𝑦′ − 𝑦)2 ]3∕2
[
Here, 𝐴(𝑡) is the fracture footprint enclosed by the fracture front 𝐶(𝑡), and 𝜎ℎ is the in situ compressive stress field that for the layered
formation under consideration varies only with respect to the vertical coordinate 𝑦. By considering 𝑛 stress layers in the formation,
we can have the following piece-wise constant representation
∑
𝑛−1
[ ]
𝜎ℎ (𝑦) = 𝜎ℎ1 (1 − (𝑦 − 𝑦1 )) + 𝜎ℎ𝑖 (𝑦 − 𝑦𝑖−1 ) − (𝑦 − 𝑦𝑖 ) + 𝜎ℎ𝑛 (𝑦 − 𝑦𝑛−1 ),
𝑖=2
where (⋅) denotes Heaviside step function, 𝜎ℎ𝑖 is the stress in the 𝑖th layer that is located within the interval (𝑦𝑖−1 , 𝑦𝑖 ), see Fig. 1.
One of the possible scenarios of fracture evolution can be its partial or complete closure. The crack can close, for instance, due
to shut-in or decreasing flow rate. Thus, it is necessary to account for the additional constraint for the fracture width. To this end,
the fracture width 𝑤 is restricted to some minimum value 𝑤min as (see e.g. [27])
𝑤(𝑥, 𝑦, 𝑡) ≥ 𝑤min (𝑥, 𝑦, 𝑡). (2)
If the fracture is opened 𝑤(𝑥, 𝑦, 𝑡) > 𝑤min (𝑥, 𝑦, 𝑡), the normal traction on the fracture surface 𝑇𝑛 is equal to the fluid pressure 𝑝 inside
the fracture. Otherwise, the normal traction 𝑇𝑛 also accounts for the contact force. Even though the contact algorithm works for any
distribution of 𝑤min , this study exclusively focuses on the case without proppant and roughness, for which 𝑤min = 0.
3
2.2. Fluid flow
Assuming laminar flow of a Newtonian fluid, the fluid flux inside the fracture is reduced to Poiseuille’s law
𝑤3
𝒒=− ∇𝑝, (3)
𝜇′
where 𝑝 is fluid pressure. By utilizing the assumption that fluid is incompressible, we obtain the continuity equation
𝜕𝑤 𝐶′
+ div 𝒒 = 𝑄(𝑡)𝛿(𝑥, 𝑦) − √ , (4)
𝜕𝑡 𝑡−𝑡0 (𝑥, 𝑦)
where fluid leak-off is calculated according to Carter’s model [54], and 𝑡0 (𝑥, 𝑦) denotes the time instant at which the fracture front
is located at the point (𝑥, 𝑦). Substituting Poiseuille’s law (3) into the mass balance equation (4) results in the Reynolds lubrication
equation
( 3 )
𝜕𝑤 𝑤 𝐶′
− div ∇𝑝 = 𝑄(𝑡)𝛿(𝑥, 𝑦) − √ . (5)
𝜕𝑡 𝜇′ 𝑡−𝑡0 (𝑥, 𝑦)
This is a standard equation that describes the evolution of fracture width inside a hydraulic fracture.
2.3. Conditions at the fracture front
We consider the following boundary conditions at the fracture front 𝐶(𝑡): zero crack aperture 𝑤 = 0, the propagation criterion
𝐾𝐼 = 𝐾Ic and zero flux condition 𝒒 ⋅ 𝒏 = 0, where 𝒏 is an outward normal to the fracture front in the (𝑥, 𝑦) plane. According to
linear elastic fracture mechanics solution for the pure mode I crack propagation [59,60], the fracture aperture in the vicinity of the
fracture front behaves as
𝐾 ′ 1∕2
𝑤 ∼ 𝜁 , (6)
𝑠→0 𝐸′
where 𝜁 is the distance to the fracture font 𝐶(𝑡). It is assumed that the fracture propagates in the limit of quasi-static equilibrium,
so the normal front velocity is positive 𝑉 > 0. If the fracture does not propagate (𝑉 = 0), then the scaled fracture toughness 𝐾 ′
in Eq. (6) should be replaced by the actual scaled stress intensity factor 𝐾𝐼′ .
3. Numerical scheme
The hydraulic fracture domain is discretized by a fixed rectangular mesh as shown in Fig. 2. The cell (𝑖, 𝑗) is centered at (𝑥𝑖 , 𝑦𝑗 )
and has dimensions 𝛥𝑥 and 𝛥𝑦. All mesh elements are divided into three types: channel, tip, and survey [26,42–44]. The channel
elements are completely filled with fluid and are contained within the fracture footprint 𝐴(𝑡). The tip elements are partially filled
with fluid and intersected by the fracture front. The survey elements are the subset of the channel elements that have at least one
adjacent tip element.
Backward Euler time stepping is used alongside central differences for spatial discretization. This is a relatively standard approach
for such problems [26,44]. Therefore, details of the numerical scheme and the contact algorithm are given in Appendix A. At the
same time, the primary focus of this section is to address the issue of fracture front tracking, as is discussed further.
3.1. The layer crossing problem
The ILSA approach invokes the use of the tip asymptotic solution to determine the distance to the fracture front and the fluid
volume in the tip elements. One such solution is presented in [46], and was used in [44]. This solution simultaneously captures the
effects of rock toughness, fluid viscosity, and leak-off in the tip region and can be expressed as
( ′ 1∕2 )
𝑠2 𝑉 𝜇 ′ 𝐾𝑠 2𝑠1∕2 𝐶 ′
= 𝑔𝛿 , , (7)
𝐸 ′ 𝑤3𝑎 𝐸 ′ 𝑤𝑎 𝑤𝑎 𝑉 1∕2
where 𝑉 is the instantaneous fracture front velocity, 𝑤𝑎 (𝑠) is the asymptotic solution, 𝑠 is the distance to the tip, while the definition
of the function 𝑔𝛿 is given in [44].
For the given fracture width 𝑤𝑠 at the center of a survey element, the distance to the fracture front 𝑠 can be calculated by solving
the nonlinear equation (7) with 𝑤𝑎 = 𝑤𝑠 . The fracture front velocity is approximated as 𝑉 = (𝑠 − 𝑠0 )∕𝛥𝑡, where 𝑠0 is the distance
from the center of the survey element to the fracture front at the previous time step. The distance 𝑠 is computed for each survey
element, and it is used to locate the fracture front as shown on the right panel in Fig. 2. In addition, the fluid in the tip element
occupies a polygonal region highlighted by the dark red color. In contrast, the unfilled part of the tip element is highlighted by
the light red color. For the purpose of the numerical scheme, the width of the partially filled tip element 𝑤𝑡 is calculated from the
volume of fluid contained in this element as
1
𝑤𝑡 = 𝑤 (𝑥, 𝑦) d𝑥 d𝑦, (8)
𝛥𝑥𝛥𝑦 ∫𝑃 𝑅 𝑎
4
Fig. 2. Left panel: Discretization for a planar fracture using a fixed rectangular mesh. Classification of the fracture elements into channel, tip, and survey. The
black line shows the fracture front and the arrows schematically indicate the direction of propagation. Right panel: schematic of the fracture front tracking and
tip volume calculation, where 𝑠 is the distance from the center of the survey element to the fracture front. The fluid-filled part of the tip element is highlighted
by the dark red color, while the remaining unfilled regions of the tip element are shown by the light red color.
Fig. 3. Example numerical calculation of a hydraulic fracture for the case of strong symmetric stress barriers. Dashed white lines show the boundaries of
adjacent geological layers, while solid yellow segments indicate the fracture front. Green circular and red crossing markers show the survey and the tip elements,
respectively.
where 𝑃 𝑅 is the polygonal region of the tip element occupied by fluid. An effective procedure for integrating the asymptotic solution
𝑤𝑎 over the tip element is described in [44].
Despite the fact that the aforementioned tip asymptote accounts for multiple physical effects, it may lead to significant errors in
situations when the fracture front is crossing a stress layer since the effects associated with layering are not included. The propagation
condition is effectively enforced over two numerical elements, from a survey to a tip element. If a geologic layer is located between
these two elements, its effect is completely ignored. To illustrate this effect, consider an example with three stress layers where
the outer layers have much higher confining stress than the middle layer. This situation corresponds to the propagation of a nearly
constant height hydraulic fracture. The results of the calculations are shown in Fig. 3. The color filling shows the fracture aperture,
while the yellow line outlines the fracture front. Dashed white lines indicate the stress layer boundaries. Green circular markers
and red crosses signify the survey and the tip elements, respectively. As can be seen from the figure, there is a numerical artifact
located near the transition from horizontal to vertical propagation. There are tip elements in the surrounding layers whose width
is significantly larger than the width of neighboring elements located in the same layer. This happens because the propagation
condition effectively spans two elements, from a survey to a tip element. And, this propagation condition does not account for the
effect of a layer boundary that is located between the two elements.
It is important to emphasize that the error associated with the location of the fracture front is on the order of one element
size. Consequently, the problem of layer crossing can always be negated by using the universal asymptotic solution (7) with a
sufficiently fine mesh. However, such an approach can result in a significant decrease in computational efficiency and forgoes the
primary advantages of the ILSA algorithm, which can capture the multiscale tip behavior on a relatively coarse mesh.
To address the problem, the tip asymptotic solution needs to account for the effect of stress layers, or the so-called layer
crossing problem needs to be solved. Such a problem is solved in [51] using several approaches, each having its own advantages
5
and disadvantages. For completeness, this paper provides a brief description of these approaches and primarily focuses on the
implementation into the Planar3D ILSA framework.
3.2. Tip asymptote from integral equation
To address the layer crossing problem, we need to solve the problem of a steadily propagating semi-infinite hydraulic fracture
in a stress-layered medium. The problem can be mathematically formulated in terms of a non-singular integral equation [46,48] as
follows
∞ [ ]
𝐾′ 4 𝜇′ 𝑥′ 1∕2
𝑤𝑎 (𝑠) = ′ 𝑠1∕2 − 𝐹 (𝑠, 𝑥′ ) 𝑉 + 2𝐶 ′ 𝑉 1∕2 d𝑥′
𝐸 𝜋𝐸 ∫0
′ 𝑤𝑎 (𝑥 ) ′ 2 ′
𝑤𝑎 (𝑥 )
(9)
4 ∑ 𝑗
𝑛−1
− 𝛥𝜎 𝐹 (𝑠, 𝑠 𝑗 ),
𝜋𝐸 ′ 𝑗=1 ℎ
where 𝑠𝑗 is the distance from the fracture tip to 𝑗th stress layer, 𝛥𝜎ℎ𝑗 = 𝜎ℎ𝑗 − 𝜎ℎ𝑗+1 is 𝑗th stress jump and the kernel 𝐹 (𝑠, 𝑥′ ) is given
by
| 𝑠1∕2 + 𝑥′ 1∕2 |
| | 1∕2
𝐹 (𝑠, 𝑥′ ) = (𝑥′ − 𝑠) log | | − 2𝑠1∕2 𝑥′ . (10)
| 𝑠1∕2 − 𝑥′ 1∕2 |
| |
In order to obtain the solution, the integral in Eq. (9) is discretized using Simpson’s rule and the resulting system of nonlinear
equations is solved numerically using Newton’s method. The infinite upper limit in integral is replaced by a sufficiently large finite
number.
This approach has high accuracy, but at the same time, it is computationally inefficient since it requires the solution of the
system of nonlinear equations on the grid approximating the infinite upper integration limit.
3.3. Toughness corrected asymptote
An alternative approach attempts to combine the universal asymptotic solution (7) and the integral equation (9) for the case of
′ as
toughness-dominated propagation by introducing the effective toughness 𝐾ef f
4 ∑ 𝑗
𝑛−1
′ 𝑤ef f (𝑠)𝐸 ′ 𝐾 ′ 1∕2
𝐾ef f
(𝑠) = , 𝑤ef f (𝑠) = ′
𝑠 − 𝛥𝜎 𝐹 (𝑠, 𝑠𝑗 ), (11)
𝑠1∕2 𝐸 𝜋𝐸 𝑗=1 ℎ
′
where 𝑤ef f (𝑠) is obtained from the integral equation (9) by excluding the terms responsible for viscous dissipation and fluid leak-off.
The effective toughness (11) is substituted into (7) to yield the asymptotic solution that approximately accounts for stress layers
( ′ )
𝑠2 𝑉 𝜇 ′ 𝐾ef f (𝑠)𝑠1∕2 2𝑠1∕2 𝐶 ′
= 𝑔𝛿 , . (12)
𝐸 ′ 𝑤3𝑎 𝐸 ′ 𝑤𝑎 𝑤𝑎 𝑉 1∕2
The primary advantage of the toughness-corrected approach is the simplicity of implementation and computational efficiency.
However, as the influence of viscous dissipation increases, the accuracy of the solution in the presence of stress layers is significantly
reduced. In addition, the effective toughness (11) should be non-negative, which imposes restrictions on the values of 𝛥𝜎ℎ𝑗 .
3.4. ODE approximation for the asymptote
The last solution is based on an ODE approximation of the non-singular integral formulation (9). By applying the following
scaling to the integral equation in (9)
( )2
𝐸 ′ 𝑤𝑎 2𝐶 ′ 𝐸 ′ 𝐾 ′3
𝑤̃ 𝑎 = , 𝜒= , 𝑙= ,
𝐾 ′ 𝑠1∕2 𝑉 1∕2 𝐾 ′ 𝜇 ′ 𝐸 ′2 𝑉
( ) (13)
𝑗 1∕2
𝛥𝜎ℎ 𝑙 ( )1∕2 1∕2
𝑠 ′ 𝑥′
𝛥𝛴𝑗 = , 𝑠
̃ = , 𝑥
̃ = ,
𝐾′ 𝑙 𝑙
and differentiating the resulting dimensionless integral formulation, one can substitute the approximation 𝑤̃ 𝑎 ≈ 𝑠̃𝛿 (as was done
in [46]) to obtain the following initial value problem
d𝑤̂𝑎 𝛽𝑚3 𝜒𝛽𝑚4̃

= + , ̂𝑎 (0) = 1,
𝑤 (14)
d𝑠̃ ̂𝑎 + 𝐺𝛴 )2 4(𝑤
3(𝑤 ̂𝑎 + 𝐺𝛴 )3
where
( )
4∑
𝑛
𝑠̃𝑗
̂𝑎 = 𝑤̃ 𝑎 − 𝐺𝛴 (𝑠),
𝑤 ̃ 𝐺𝛴 (𝑠)
̃ = 𝛥𝛴𝑗 𝑠̃𝑗 𝐺 , (15)
𝜋 𝑗=1 𝑠̃
6
Fig. 4. Panel (𝑎): Comparison of the asymptotic solutions for the case of two stress drop layers located ahead of a survey element. The value of the stress drop
is 𝛥𝜎ℎ1 = −0.3 MPa and 𝛥𝜎ℎ2 = −0.2 MPa. Panel (𝑏): Distances in the top, bottom, and horizontal directions from the center of the survey element to the fracture
front that are used as an initial condition for the eikonal equation.
√
coefficients 𝛽𝑚 = 21∕3 35∕6 , 𝛽𝑚̃ = 4∕(15( 2 − 1))1∕4 , and the function 𝐺(𝑡) is defined as
1 − 𝑡2 || 1 + 𝑡 ||
𝐺(𝑡) = ln | | + 2. (16)
𝑡 |1 − 𝑡|
The initial value problem (14) is solved numerically using explicit Dormand-Prince method [61] with an adaptive step size
control [62]. The first and the second terms on the right side of Eq. (14) allow us to capture the influence of stress layers in
the case of non-negligible viscous dissipation and leak-off. Also, there is no need to impose additional restrictions on the magnitude
of 𝛥𝛴𝑗 to solve equation (14). Thus, the ODE approximation is more accurate than the toughness-corrected approach. At the same
time, the method is computationally more efficient than the integral equation approach.
Based on the advantages and disadvantages of the approaches described above, we select the ODE approximation as the primary
method for fracture front tracking. See also [51] for the detailed comparison of the accuracy and performance of the different
approaches.
3.5. Locating the fracture front
To determine distance 𝑠 to the fracture front, it is necessary to invert the asymptotic solution 𝑤𝑎 (𝑠, 𝑉 ) for the given fracture width
𝑤𝑠 of the survey element. Since there is no stress variation in the horizontal direction, the solution without the stress correction is
used. At the same time, since the stress varies in the vertical or 𝑦 direction, the stress-corrected asymptote needs to be used.
The fracture front velocity 𝑉 is unknown at the current time step, and it is therefore approximated by the backward Euler scheme
as:
𝑠 − 𝑠0
𝑉 = ,
𝛥𝑡
where 𝑠0 is the distance to the fracture front from the previous time step. For the case of vertical propagation, the distance from
the 𝑗th stress layer to the fracture tip 𝑠𝑗 is also unknown at the current time step. To overcome this, we introduce an auxiliary
parameter 𝜌𝑗layer = 𝑦𝑗 − 𝑦𝑠 representing the signed vertical distance from the center of the survey element to the location of the 𝑗th
stress layer, where 𝑦𝑗 is the 𝑦 coordinate of the 𝑗th stress layer boundary and 𝑦𝑠 is the 𝑦 coordinate of the center of survey element.
Therefore, the distance from the fracture tip to the 𝑗th stress layer is defined as 𝑠𝑗 = 𝑠 − 𝜌𝑗layer .
Fig. 4(a) shows the comparison of various asymptotic solutions 𝑤𝑎 (𝑠, 𝑉 ) for the case of two stress drops located at 𝜌1layer = 3 m
and 𝜌2layer = 5 m from the center of the survey element. It is important to point out that the presence of at least one stress drop can
make the asymptotic solution non-monotonic. As a result, the equation 𝑤𝑎 (𝑠, 𝑉 (𝑠)) = 𝑤𝑠 can have several roots 𝑠𝑖 corresponding to
different distances to the fracture front for the given width 𝑤𝑠 of the survey element. The smallest root is chosen in the algorithm
because other roots correspond to the situation in which the fracture crosses multiple layers. But if there is a solution before the
layer, it means that the fracture is unable to cross the layer and therefore such roots are not physical. In the case of stress jumps or
the absence of layers, the asymptotic solution 𝑤𝑎 (𝑠, 𝑉 (𝑠)) is monotonic and there is only one root.
According to the ILSA approach [26,42–44], the location of the fracture front is defined by the level set  = 0 of the signed
distance function  (𝑥, 𝑦), which is the solution of the eikonal equation [44]. To account for the horizontal stress layers in the plane,
we distinguish the top, bottom, and horizontal directions of the fracture propagation. Equation 𝑤𝑎 (𝑠, 𝑉 (𝑠)) = 𝑤𝑠 is solved three times
7
Fig. 5. Scheme to integrate the asymptotic solution over the partially filled tip element.
top
for each survey element. First, we need to find a solution assuming that the fracture tip propagates upward to yield distance 𝑠𝑘 .
Then the equation is solved for the distance 𝑠bot 𝑘
, assuming that the fracture front propagates downward. And finally, we have to
find the distance 𝑠h𝑘 to the fracture front in the horizontal direction, i.e. without layers. The obtained distances to the fracture front
are used as an initial condition to solve the eikonal equation for the signed distance function  as
{ }
top
 (𝑥𝑘 , 𝑦𝑘 ) = −𝑠𝑘 , −𝑠bot h
𝑘 , −𝑠𝑘 , (17)
where (𝑥𝑘 , 𝑦𝑘 ) is the center of the 𝑘th survey element. For the remaining elements, the signed distance  takes only one value. To
compute the values of  in the remaining grid cells, the eikonal equation is solved using the fast marching method [63,64] and the
following finite difference scheme [65]
( )2
𝑖,𝑗 − 𝑖−1,𝑗 𝑖,𝑗 − 𝑖+1,𝑗
max , ,0 +
𝛥𝑥 𝛥𝑥
( )2
𝑖,𝑗 − 𝑖,𝑗−1 𝑖,𝑗 − 𝑖,𝑗+1
max , , 0 = 1. (18)
𝛥𝑦 𝛥𝑦
In order to account for multiple initial values in each survey element, we modify the finite differences in (18) as follows:
top
• If element (𝑖, 𝑗 − 1) is the 𝑘th survey, then 𝑖,𝑗−1 = −𝑠𝑘 ;
• If element (𝑖, 𝑗 + 1) is the 𝑘th survey, then 𝑖,𝑗+1 = −𝑠bot
𝑘
;
• If element (𝑖 ± 1, 𝑗) is the 𝑘th survey, then 𝑖±1,𝑗 = −𝑠h𝑘 .
To illustrate the procedure of the finite difference computation, let us consider Fig. 4(b). For cell ‘‘1’’ the horizontal difference takes
the form
( )
𝑖,𝑗 − 𝑖−1,𝑗 𝑖,𝑗 − 𝑖+1,𝑗 𝑖,𝑗 + 𝑠h𝑘
max , ,0 = ,
𝛥𝑥 𝛥𝑥 𝛥𝑥
top
while for cell ‘‘2’’ the vertical difference is calculated using 𝑠𝑘 value from the neighboring survey element
( ) top
𝑖,𝑗 − 𝑖,𝑗−1 𝑖,𝑗 − 𝑖,𝑗+1 𝑖,𝑗 + 𝑠𝑘
max , ,0 = .
𝛥𝑦 𝛥𝑦 𝛥𝑦
A finite difference accounting for the distance 𝑠bot
𝑘
is constructed similarly.
3.6. Tip volume calculation
As was mentioned above, the tip width 𝑤𝑡 is computed from the volume of fluid occupying the tip element by integrating the
asymptotic solution 𝑤𝑎 over the partially filled tip element. As can be seen from Fig. 5, the fluid volume occupying the polygonal
region 𝐴𝐷𝐸𝐹 𝐺 can be calculated similarly as in [44], namely
𝑉𝐴𝐷𝐸𝐹 𝐺 = 𝑉𝐴𝐵𝐶 − (𝓁 − 𝛥𝑦 sin 𝛼)𝑉𝐺𝐹 𝐶 − (𝓁 − 𝛥𝑥 cos 𝛼)𝑉𝐷𝐵𝐸 , (19)
where (⋅) is the Heaviside step function, 𝓁 is the distance from the farthest interior corner to the front, and 𝛼 is the front orientation.
Note that the triangle 𝐺𝐹 𝐶 disappears when 𝓁 < 𝛥𝑦 sin 𝛼, while the triangle 𝐷𝐵𝐸 disappears when 𝓁 < 𝛥𝑥 cos 𝛼. The volume of
fluid occupying a triangular region can be computed as
𝑠 𝑠𝑀0 (𝑠) − 𝑀1 (𝑠)
1
𝑉▵ (𝑠) = 𝑠′ 𝑤𝑎 (𝑠 − 𝑠′ ) d𝑠′ = , (20)
sin 𝛼 cos 𝛼 ∫0 sin 𝛼 cos 𝛼
8
Fig. 6. Fracture height versus time for the example case of symmetric stress barriers computed for the different meshes using the universal asymptotic solution
(left) and the stress-corrected asymptote (right). The horizontal black solid line indicates the location of the stress layer.
where 𝑀0 (𝑠) and 𝑀1 (𝑠) are zeroth and first moments of the asymptotic solution and are defined as
𝑠 𝑠
𝑀0 (𝑠) = 𝑤𝑎 (𝑠′ ) d𝑠′ , 𝑀1 (𝑠) = 𝑠′ 𝑤𝑎 (𝑠′ ) d𝑠′ . (21)
∫0 ∫0
Therefore, to calculate the tip volume, it is sufficient to calculate the two moments of the asymptotic solution 𝑤𝑎 .
Since the stress layers are arranged vertically, they should not affect the fracture tip propagation in the horizontal direction.
Thus, in the case of horizontal front orientation or 𝛼 = 0, the fracture tip width is determined by the universal asymptotic solution
that we denote as 𝑤𝑢𝑎 (𝑠) in this section. On the other hand, when the fracture propagates in the vertical direction or 𝛼 = 𝜋∕2, then the
influence of the stress layers needs to be included, and the fracture tip width is determined by the asymptotic solution accounting
for stress layers that we denote as 𝑤𝜎𝑎 (𝑠) in this section. To account for the arbitrary orientation of the fracture front, an interpolation
scheme is implemented as follows
𝑤𝑎 (𝑠) = 𝑤𝑢𝑎 (𝑠) cos2 𝛼 + 𝑤𝜎𝑎 (𝑠) sin2 𝛼. (22)
The zeroth and first moments of the interpolated asymptotic solution (22) can be expressed as
𝑠
𝑀0 (𝑠) = 𝑤𝑎 (𝑠′ ) d𝑠′ = 𝑀0𝑢 (𝑠) cos2 𝛼 + 𝑀0𝜎 (𝑠) sin2 𝛼,
∫0
𝑠 (23)
𝑀1 (𝑠) = 𝑠′ 𝑤𝑎 (𝑠′ ) d𝑠′ = 𝑀1𝑢 (𝑠) cos2 𝛼 + 𝑀1𝜎 (𝑠) sin2 𝛼,
∫0
where 𝑀0𝑢 and 𝑀1𝑢 (𝑠) are zeroth and first moments of the universal asymptotic solution [44], while 𝑀0𝜎 and 𝑀1𝜎 are zeroth and
first moments of the asymptotic solution accounting for stress layers. The computation of the moments in the case of the universal
asymptotic solution is given in [44]. The moments 𝑀0𝜎 and 𝑀1𝜎 are calculated by numerical integration of the solution with stress
layers.
4. Stress relaxation for the asymptote
As shown in [51], the tip asymptote with stress layers (or stress-corrected asymptote) has a relatively small validity region,
which may lead to significant errors in quantifying the fracture front location, especially for the case of a coarse mesh. Therefore,
the purpose of this section is to evaluate the accuracy of the tip asymptote in the case of a planar fracture. In addition, we consider
suitable corrections to increase the validity region of the tip asymptotic solution.
In particular, we focus on the case when the reservoir layer is surrounded by two identical layers with higher confining
stresses (i.e. symmetric stress barriers). The problem parameters that are used for calculations are the following: Young’s modulus
𝐸 = 10 GPa, Poisson’s ratio 𝜈 = 0.3, fracture toughness 𝐾Ic = 1 MPa m1∕2 , leak-off coefficient 𝐶𝐿 = 0, fluid viscosity 𝜇 = 0.1 Pa s,
injection rate 𝑄 = 0.2 m3 /s, height of the central layer 𝐻 = 15 m, and compressive stresses 𝜎ℎ1 = 𝜎ℎ3 = 28 MPa, and 𝜎ℎ2 = 25 MPa.
The computations are performed for the different number of elements in the central reservoir layer 𝑛𝑐 ∈ {3, 5, 9, 15}, and the mesh
is such that 𝛥𝑥 = 𝛥𝑦 = 𝐻∕𝑛𝑐 .
Fig. 6 shows the fracture height computed for the different meshes for the universal asymptotic solution (7) and the stress-
corrected asymptote calculated via ODE approximation (14). The horizontal black solid line indicates the location of the high stress
layer. As can be seen from the left panel of Fig. 6, the accuracy of the universal asymptote is compromised near the layer boundary,
especially for coarse meshes. However, as the fracture height increases, the mesh size dependence reduces. In contrast, the accuracy
of the stress-corrected asymptote is great near the layer intersection, but it is reduced further away from the layer boundary. This
example clearly shows that the stress-corrected ODE approximation requires a modification to mitigate the inaccuracy associated
with the fracture front being away from the layer boundary.
9
Fig. 7. Comparison of the error 𝜀2 defined in (27) for different values of the stress relaxation distance 𝑘 ∈ {2, 2.5, 3, 3.5}.
The paper [51] proposed a stress relaxation factor 𝜆 that allows us to significantly extend the validity region of the stress-corrected
asymptote. The stress relaxation factor is incorporated into the solution by modifying the stress terms in (15) as follows
( )
4∑
𝑛
𝑠̃𝑗
𝐺𝛴 (𝑠)
̃ = 𝜆(𝑠̃𝑗 )𝛥𝛴𝑗 𝑠̃𝑗 𝐺 , (24)
𝜋 𝑗=1 𝑠̃
where the stress relaxation factor 𝜆 has values from 0 to 1. Thus, 𝜆 effectively reduces the magnitude of the stress barrier. For a
uniformly pressurized finite plane-strain fracture with symmetric stress barriers, the optimal relaxation factor has the following form
(see [51])
0.7
𝜆ps (𝜉) = , (25)
0.7 + 𝜉∕𝐻
where 𝜉 is the distance from the fracture tip to the stress layer and 𝐻 is the height of the central layer. Unfortunately, such a
relaxation factor depends on the height of the central layer, which is not well-defined for an arbitrary configuration of the reservoir
layers. In addition, for the case of a finite planar fracture, the expression for the stress relaxation factor (25) might not lead to the
optimal result. Therefore, we aim to determine an alternative expression for the stress relaxation factor that does not depend on the
geometry of the reservoir layers.
Results shown in Fig. 6 demonstrate that the stress-corrected asymptote leads to more accurate results near the layer boundary,
while the universal asymptote excels far away from the layer boundary. In view of this observation, we are looking for the stress
relaxation factor in the following form
{
1 − 𝜉∕(𝑘𝛥𝑦), 𝜉 < 𝑘𝛥𝑦,
𝜆𝑘 (𝜉) = (26)
0, 𝜉 ≥ 𝑘𝛥𝑦,
where 𝑘 represents the relaxation distance measured in the number of elements.

To determine the optimal value of the relaxation distance 𝑘, we consider the case of symmetric stress barriers for the parameters
specified at the beginning of this section. Then, we independently vary fluid viscosity 𝜇 ∈ [0.01, 0.5] Pa s, fracture toughness
𝐾Ic ∈ [0.001, 5]𝑀𝑃 𝑎m 1∕2 , and magnitude of the stress barrier 𝛥𝜎 ∈ [1.7, 5] MPa, where 𝛥𝜎 = 𝜎ℎ3 − 𝜎ℎ2 . To compare the performance of
the relaxation factors, for each fixed set of parameters 𝜇, 𝐾Ic , and 𝛥𝜎, we vary the relaxation distance 𝑘 ∈ {2, 2.5, 3, 3.5}. As the error
measure, we take the average difference between the solution for the fine mesh 𝑛𝑐 = 15 and the solutions for the coarser meshes
𝑛𝑐 < 15 as follows
‖ ‖
‖ℎ𝑛𝑐 (𝑡) − ℎ15 (𝑡)‖ 𝑡max
‖ ‖2
𝜀2 = max , ‖𝑓 (𝑡)‖22 = 𝑓 2 (𝑡) d𝑡, (27)
𝑛𝑐 ‖ℎ15 (𝑡)‖ ∫0
‖ ‖2
where the subscript 𝑛𝑐 denotes the number of elements in the central layer and 𝑡max = 200 s is the calculation time. The values of
the fracture height ℎ𝑛𝑐 in (27) for different meshes are computed for the same fluid viscosity 𝜇, fracture toughness 𝐾Ic , magnitude
of the stress barrier 𝛥𝜎, and relaxation distance 𝑘.
Fig. 7 shows the comparison of the error 𝜀2 defined in (27) for different values of the stress relaxation distance 𝑘. Different values
of fluid viscosity 𝜇, fracture toughness 𝐾Ic , and magnitude of the stress barrier 𝛥𝜎 are considered. For most of the parameters, the
minimum of the error 𝜀2 is reached for the case 𝑘 = 3. However, it is worth noting that the difference between the results for 𝑘 = 2,
𝑘 = 2.5, and 𝑘 = 3 is not very significant. Thus, for simplicity, we choose 𝜆3 (𝜉) as the optimal stress relaxation factor.
In addition, Fig. 8 shows the comparison of the corresponding errors for 𝜆3 (𝜉) and 𝜆ps (𝜉) defined in (25) in the case of symmetric
stress barriers for which the height of the central layer 𝐻 is well-defined. We also consider the universal asymptotic solution (𝜆 = 0)
and the stress-corrected asymptote without relaxation (𝜆 = 1). The case with 𝜆3 (𝜉) shows the best performance for almost the
entire range of parameters, and it is close to the result calculated with the stress relaxation factor 𝜆ps (𝜉) for a plane-strain fracture
with symmetric stress barriers. Thus, for further calculations, we use the stress-corrected asymptote with optimal relaxation, which
includes solving the initial value problem (14) with the modified stress term (24) with 𝜆 = 𝜆3 (𝜉).
10
Fig. 8. Comparison of the error 𝜀2 for the universal asymptote (𝜆 = 0), the stress-corrected asymptote without relaxation (𝜆 = 1), the stress-corrected asymptote
with optimal relaxation 𝜆3 (𝜉), and 𝜆ps defined in (25).
Table 1
Input parameters for the case of a penny-shaped hydraulic fracture.
Regime 𝜇, Pa s 𝐾Ic , MPa m1∕2 𝐶𝐿 , m/s1∕2
𝐾 10−4 3 0
𝑀 0.1 0 0
𝑀̃ 0.1 0 5 ⋅ 10−4
𝐾̃ 10−4 5 1 ⋅ 10−4
5. Numerical examples
This section presents the comparison of the numerical results with the semi-analytical solution for a penny-shaped hydraulic
fracture, validation of the numerical algorithm against experimental observations, and illustration of the fracture closure algorithm.
For visualization purposes, the discontinuous fracture front is post-processed. The centers of fracture front segments in each tip
element are connected to their closest neighbors to construct the continuous fracture front.
5.1. Penny-shaped fracture
To verify the numerical algorithm, we first consider the case of a penny-shaped hydraulic fracture with no fluid lag [66,67].
The problem has four vertex or limiting solutions [68,69]: storage-viscosity (𝑀), storage-toughness (𝐾), leak-off-viscosity (𝑀),̃ and
leak-off-toughness (𝐾).̃ This section compares the developed numerical method to the semi-analytical solution for a penny-shaped
hydraulic fracture [67]. In order to probe the limiting regimes, we consider parameters corresponding to the validity zone of the
corresponding vertex solution. The common problem parameters for all propagation regimes are the following: Young’s modulus
𝐸 = 32 GPa, and Poisson’s ratio 𝜈 = 0.3. The remaining parameters for the corresponding regimes are summarized in Table 1. The
fluid is injected at a constant flow rate 𝑄0 = 0.01 m3 /s for 1000 s.
The left panel in Fig. 9 shows the comparison of the fracture radius evolution for the limiting propagation regimes of the penny-
shaped hydraulic fracture. Here the radius of fracture is calculated as the fracture extension along the 𝑥-axis at 𝑦 = 0. For the 𝐾
regime, the numerical solution features step-wise behavior. This propagation pattern is associated with discretization by rectangular
elements and nearly constant fluid pressure in the fracture. It is worth noting that a similar behavior of the fracture radius has been
observed in [70,71] for the toughness-dominated regime. As can be seen from the figure, the fracture radius overall closely matches
the vertex solutions for all limiting regimes.
The right panel in Fig. 9 shows the comparison of the fracture width for the different regimes at the final time 𝑡 = 1000 s. The
agreement between the numerical results and the semi-analytical solutions is excellent.
5.2. Comparison with experimental observations
In order to validate the numerical algorithm, we compare computations using the developed numerical algorithm with the
experimental data from [72]. In the experiment, the fracture propagates between two separate blocks of polymethylmethacrylate
(PMMA). Hence, the fracture toughness in the contact surface is zero, and also the blocks are impermeable. For comparison with
the experiment, we use the following parameters: Young’s modulus 𝐸 = 3.3 GPa, Poisson’s ratio 𝜈 = 0.4, fracture toughness 𝐾Ic = 0,
and leak-off coefficient 𝐶𝐿 = 0. The surface of one of the blocks is profiled so that after compression a piece-wise constant stress
distribution over the contact plane is effectively imposed. The layer boundaries and the corresponding compressive stresses are:
𝑦1 = −25 mm, 𝑦2 = 25 mm, 𝜎ℎ1 = 11.2 MPa, 𝜎ℎ2 = 7 MPa, 𝜎ℎ3 = 5 MPa.
11
Fig. 9. Results of simulations for a penny-shaped hydraulic fracture. Numerical results are shown by the circular markers, while the vertex solutions are shown
by the solid lines. The left panel shows the evolution of the fracture radius, while the right panel depicts the fracture width at 𝑡 = 1000 s.
Fig. 10. Left panel: numerically computed fracture width (color filling) and the fracture footprint obtained in the experiment (solid white line). Right panel:
temporal evolution of the fracture width. The dashed red and blue lines depict the experimental measurements at (30, 0) mm and (−30, 0) mm, respectively. The
solid green line shows the numerical result.
A fluid with viscosity 𝜇 = 30 Pa s is pumped with a piece-wise constant flow rate

⎧0.9 ⋅ 10−9 m3 ∕s, 0 s < 𝑡 ⩽ 31 s,
⎪
𝑄(𝑡) = ⎨6.5 ⋅ 10−9 m3 ∕s, 31 s < 𝑡 ⩽ 151 s,
⎪ −9 3
⎩2.3 ⋅ 10 m ∕s, 151 s < 𝑡.
We refer to [72] for a more detailed description of the experimental procedure.
The left panel in Fig. 10 shows the numerical results for the fracture width 𝑤 at 𝑡 = 665 s. The solid white line shows the fracture
footprint obtained from the experimental data. The horizontal dashed white lines depict the boundaries of the stress layers. The
comparison shows good agreement between the numerical results and experimental data for the fracture footprint. The right panel
in Fig. 10 displays the comparison of the fracture width evolution near the injection point. In the experiment, the fracture width is
measured at (30, 0) mm and (−30, 0) mm. The fracture width computed numerically is evaluated only at (30, 0) mm due to symmetry.
As can be seen from Fig. 10, the numerical calculations of the fracture width closely match the data measured in the experiment
within the measurement accuracy.
5.3. Fracture closure
The next example illustrates the fracture closure algorithm. In particular, the three-layered geometry of compressive stress 𝜎ℎ
and the leak-off coefficient 𝐶𝐿 is considered, such that there are two disconnected fracture segments after closure. The problem
12
Fig. 11. The fracture width distribution for the example of fracture closure evaluated at 𝑡 ∈ {350, 550, 750, 2000} s. The horizontal dashed lines show layer
boundaries.
parameters are the following: Young’s modulus 𝐸 = 9.5 GPa, Poisson’s ratio 𝜈 = 0.2, fracture toughness 𝐾Ic = 1 MPa m1∕2 , and fluid
viscosity 𝜇 = 0.1 Pa s. The central layer is surrounded by two layers with lower compressive stress and zero leak-off coefficient:
𝜎ℎ1 = 𝜎ℎ3 = 6.5 MPa, 𝜎ℎ2 = 7.5 MPa, 𝐶𝐿1 = 𝐶𝐿3 = 0, 𝐶𝐿2 = 5 ⋅ 10−5 m/s1∕2 .
The lower value of stress in the surrounding layers allows the fracture front to grow quicker once the boundary of the layers is
reached. Fluid is injected in the middle of the central layer with the piece-wise constant flow rate
{
0.02 m3 ∕s, 0 s < 𝑡 ⩽ 550 s,
𝑄(𝑡) =
0 m3 ∕s, 550 s < 𝑡.
Fig. 11 illustrates the calculated fracture width for different time instants. The fracture front reaches the boundaries of the central
layer and starts to grow quicker due to the lower value of stress at 𝑡 = 350 s. Fluid injection stops at 𝑡 = 550 s. Then, the fracture in
the central layer starts to close due to fluid leak-off into the reservoir. At the same time, the fracture portions in the top and bottom
layers continue to propagate because the leak-off coefficient in these layers is equal to zero. At 𝑡 = 2000 s, the fracture in the central
layer is completely closed, and the contact condition 𝑤(𝑥, 𝑦) = 𝑤min (𝑥, 𝑦) is satisfied. The two fracture regions become separated by
the zone where fracture walls are in contact.
6. Mesh sensitivity
In this section, we investigate the sensitivity of the hydraulic fracture geometry to mesh size. The mesh that is used for all cases
is square, i.e. 𝛥𝑥 = 𝛥𝑦. Since the injection point is located in the middle of the central layer, it is assumed that the number of
elements in the central layer 𝑛𝑐 is odd. This ensures correct location of the injection point and captures fracture symmetry until the
fracture front intersects the stress barriers. An adaptive time stepping is applied to ensure convergence of the numerical algorithm.
As was stated at the beginning of this paper, the primary purpose of using the stress-corrected asymptote is to improve accuracy
for coarse meshes, which in turn allows using such coarse meshes for practical problems to obtain the numerical result quicker.
6.1. Constant height hydraulic fracture
We first consider the propagation of a height-contained fracture or Perkins–Kern–Nordgren (PKN) fracture [14,15] in which two
layers with high stress surround the reservoir and restrict vertical height growth. The problem of PKN fracture has four limiting
regimes of propagation [1,73] that are determined by the dominance of fluid viscosity or rock toughness as well as a feature either
small or large fluid leak-off in the surrounding rock. In this section, we consider only the storage toughness (𝐾) and storage viscosity
(𝑀) regimes corresponding to no leak-off.
To check the sensitivity of the fracture geometry to mesh size, we perform simulations for a different number of elements 𝑛𝑐 in the
central layer. The height of the reservoir layer is 𝐻 = 20 m, and the surrounding layers have compressive stress 𝜎ℎ1 = 𝜎ℎ3 = 30 MPa,
while the compressive stress in the middle layer is 𝜎ℎ2 = 20 MPa. The parameters corresponding to the 𝐾 and 𝑀 regimes are outlined
in Table 2. The simulations are performed for 700 s of fluid injection with a constant flow rate.
The left panel in Fig. 12 plots the fracture height versus the number of elements 𝑛𝑐 in the reservoir layer. Blue diamonds
correspond to the 𝐾-regime and the universal asymptote, red circles to 𝐾-regime and the stress-corrected asymptote, green squares
to the 𝑀-regime and the universal asymptote, and yellow hexagons to the 𝑀-regime and the stress-corrected asymptote. The gray
horizontal line shows the height of the reservoir layer. In the case of the universal asymptotic solution, the discrepancy between
13
Table 2
Parameters for the storage toughness (𝐾) and storage viscosity (𝑀) regimes of a constant height hydraulic fracture.
Regime 𝐸, GPa 𝜈 𝐾Ic , MPa m1∕2 𝐶𝐿 , m/s1∕2 𝜇, Pa s 𝑄, m3 /s
𝐾 1 0.3 1.5 0 0.01 0.05
𝑀 3 0.3 0.5 0 0.1 0.02
Fig. 12. Fracture height (left), fracture length (center), and the error of the fracture length compared to the reference solution (right) for a constant height
fracture versus the number of elements 𝑛𝑐 in the central layer. Blue diamonds correspond to 𝐾-regime and the universal asymptote, red circles to 𝐾-regime
and the stress-corrected asymptote, green squares to the 𝑀-regime and the universal asymptote, and yellow hexagons to the 𝑀-regime and the stress-corrected
asymptote. Gray horizontal lines denote the reference length computed by EPKN model [74].
the fracture height and the reservoir layer height is proportional to double the element height, i.e. 2𝛥𝑦, for both 𝐾 and 𝑀 regimes.
This error corresponds to the growth of approximately one element beyond the layer boundary and is an unwanted feature of the
original ILSA implementation. In the case of stress-corrected asymptote, the fracture height almost precisely coincides with the
reservoir layer height for both of the regimes and all meshes.
The central panel in Fig. 12 shows the final fracture length versus the number of elements 𝑛𝑐 in the reservoir layer. The gray
horizontal lines signify the reference fracture length computed by EPKN model with non-local elasticity [74]. For the 𝐾-regime,
both the universal asymptotic solution and the stress-corrected asymptote predict a considerably smaller length than the reference
solution. The discrepancy in length between the coarsest mesh (𝑛𝑐 = 3) and the finest mesh (𝑛𝑐 = 31) is approximately 25 percent.
For the 𝑀-regime, this difference reduces to approximately 10 percent. It is interesting to observe that the improvement in the
height calculation does not noticeably impact the fracture length. The reason for this lies in the fact that the elements that are
located outside the central layer are practically closed and do not have any fluid volume. Therefore, the fracture volume in each
vertical cross-section is approximately the same no matter what asymptote is used, which ultimately results in similar fracture length
predictions.
The right panel in Fig. 12 depicts the error of the fracture length relative to the reference solution. One can observe a convergence
rate of approximately 1.2 for both 𝐾 and 𝑀 regimes. It is worth noting that the fracture length converges at the same rate for both
the universal asymptote and the stress-corrected asymptote.
It is important to note that a similar issue of length convergence for a constant height fracture was mentioned in Dontsov
et al. [47]. In the latter work, inaccurate computation of the fracture height was cited as the primary reason for length error.
However, in the case of the stress-corrected asymptote, the fracture growth is arrested exactly at the location of the stress barrier, but
the length accuracy remains practically unchanged. One possible explanation is that the asymptotic solution becomes less accurate
for high fracture front curvatures, where the validity zone of the plane strain assumption for the tip asymptote is reduced. This
problem requires further investigation and is beyond the scope of this study.
6.2. Symmetric stress barriers
This section examines the sensitivity of the fracture geometry to mesh size for the case of moderate symmetric stress barriers,
i.e. when some fracture height growth is permitted. Similarly to the previous section, we consider the two sets of parameters
presented in Table 2. To promote the fracture height growth, we reduce the compressive stress in the surrounding layers to
𝜎ℎ1 = 𝜎ℎ3 = 20.15 MPa for the 𝐾 regime and to 𝜎ℎ1 = 𝜎ℎ3 = 20.35 MPa for the 𝑀 regime. The confining stress in the central layer
remains the same, i.e. 𝜎ℎ2 = 20 MPa.
Fig. 13 shows the fracture height (left) and fracture length (right) versus the number of elements 𝑛𝑐 in the reservoir layer. As
for the constant height case, the blue diamonds correspond to the 𝐾 regime and the universal asymptote, red circles to the 𝐾
regime and the stress-corrected asymptote, green squares to the 𝑀 regime and the universal asymptote, and yellow hexagons to the
𝑀 regime and the stress-corrected asymptote. The non-monotonic behavior of the fracture height and length for the 𝐾 regime is
related to the step-wise fracture propagation, as was discussed in Section 5.1. The universal asymptote considerably overestimates
the fracture height on a coarse grid, which also leads to significant errors in length since the fracture volume remains the same. The
stress-corrected asymptote, on the other hand, estimates the fracture height more accurately, resulting in accurate computation of
the fracture length as well.
14
Fig. 13. Fracture height (left) and fracture length (right) at 𝑡 = 300 s for a symmetric stress barrier versus the number of elements 𝑛𝑐 in the central layer. Blue
diamonds correspond to the 𝐾 regime and the universal asymptote, red circles to the 𝐾 regime and the stress-corrected asymptote, green squares to the 𝑀
regime and the universal asymptote, and yellow hexagons to the 𝑀 regime and the stress-corrected asymptote.
Fig. 14. Fracture height versus time for 𝐾 regime (𝑎) and 𝑀 regime (𝑏). The left panel for each row shows the fracture height for the universal asymptote,
while the right panel depicts the result for the stress-corrected asymptote. The black horizontal line shows the height of the reservoir layer, while 𝑛𝑐 denotes
the number of elements in the central layer.
Fig. 14 shows the fracture height versus time for the 𝐾 regime (𝑎) and the 𝑀 regime (𝑏). The left panel for each row shows the
fracture height for the universal asymptote, while the right panel shows the results for the stress-corrected asymptote. The black
horizontal lines signify the height of the reservoir layer. For the case of the universal asymptote, the fracture height prediction on
15
Fig. 15. Fracture footprint for the case of thin high stress layer at 𝑡 = 150 s for the universal asymptote (left) and the stress-corrected asymptote (right). The
number of elements in the thin layer is denoted by 𝑛. The horizontal black lines show the layer boundaries, while the black circular marker highlights the
injection point.
a coarse mesh has a significant error relative to the fine mesh result. In contrast with the universal asymptote, the stress-corrected
asymptote provides a more accurate prediction of layer crossing. The fracture height practically does not depend on the mesh size
when the fracture crosses the layer boundary. Note that the 𝐾 regime case features oscillatory behavior, as was also previously
observed for the radial fracture case. It is important to note that the fracture height at 𝑡 = 0 varies for different numbers of elements
in the central layer 𝑛𝑐 since the initial fracture size varies proportionally to mesh size, see Appendix A. The initial fracture consists
of a single survey element at the origin, which is surrounded by the adjacent tip elements. Thus, the height of the initial fracture is
exactly equal to 2𝛥𝑦, which causes the difference in the initial fracture height for different meshes. However, the volume of the initial
fracture is set to zero, which significantly reduces the effect of the initial condition on further fracture propagation. Furthermore, as
indicated in [75], the effect of initial condition is dissipates very quickly once fracture starts to propagate. Therefore, the difference
of the fracture height for different meshes is not a significant source of error, especially for large times.
6.3. Thin high stress layer
This section aims to examine the fracture propagation for the case of a configuration with a thin layer that has high compressive
stress. The parameters used in the simulation are the following: Young’s modulus 𝐸 = 20 GPa, Poisson’s ratio 𝜈 = 0.3, fracture
toughness 𝐾Ic = 0.5 MPa m1∕2 , leak-off coefficient 𝐶𝐿 = 0, fluid viscosity 𝜇 = 0.1 Pa s, and injection flowrate 𝑄 = 0.1 m3 /s. The
layer boundaries and compressive stresses are defined as follows:
𝑦1 = −7.5 m, 𝑦2 = 7.5 m, 𝑦3 = 12.5 m, 𝑦4 = 37.5 m,

𝜎ℎ1 = 35 MPa, 𝜎ℎ2 = 26 MPa, 𝜎ℎ3 = 30 MPa, 𝜎ℎ4 = 26 MPa, 𝜎ℎ5 = 35 MPa.
The thin layer is located between 𝑦2 = 7.5 m and 𝑦3 = 12.5 m and is subject to the confining stress of 𝜎ℎ3 = 30 MPa. The
computations are performed for different numbers of elements in the thin layer 𝑛 ∈ {1, 2, 3, 5, 10}, and the mesh is such that
𝛥𝑥 = 𝛥𝑦 = (𝑦3 − 𝑦2 )∕𝑛.
Fig. 15 shows the fracture footprint at 𝑡 = 150 s for the universal asymptote (left) and the stress-corrected asymptote (right).
The blue dash-dotted lines depict the footprint for coarse mesh with 𝑛 = 1, the red dashed lines show the result for mesh with
𝑛 = 2, and the solid green lines correspond to the footprint for fine mesh with 𝑛 = 10. As can be seen from the figure, the stress-
corrected asymptote gives a very accurate height approximation, which is in contrast to the universal asymptote. The reduced
accuracy of the fracture height prediction for the universal asymptote is related to the issue illustrated on the left panel in Fig. 12.
Since the tip elements do not account for stress layers, the fracture always grows at least one element beyond the layer boundary.
The fracture length depends on the mesh size for both the universal asymptote and the stress-corrected asymptote. However, the
length discrepancy between the coarse and fine meshes does not exceed 10 percent for the universal asymptote and 6 percent for
the stress-corrected asymptote, which is acceptable for most practical cases. Note that such a dependence of length on the mesh
size can be related to the fact that the fracture in the layers above and below the thin layer resembles PKN fracture, for which a
relatively strong length discrepancy is observed, see discussion in Section 6.1.
In order to quantify accuracy associated with fracture crossing the thin layer, Fig. 16 plots the fracture height versus time. The left
panel corresponds to the universal asymptote, while the results for the stress-corrected asymptote are shown on the right. Different
numbers of elements in the thin layer 𝑛 ∈ {1, 2, 3, 5, 10} are considered. Circular markers indicate the time instances corresponding
16
Fig. 16. Fracture height versus time for the case with high stress thin layer. Different lines correspond to different numbers of elements 𝑛 ∈ {1, 2, 3, 5, 10} in
the thin layer. The left panel shows the results of the universal asymptote, while the results of the stress-corrected asymptote are shown in the right picture.
Circular markers indicate the crossing time of the thin layer.
Fig. 17. The left panel shows the total computation time versus the number of elements in the central layer 𝑛𝑐 for the universal and the ODE asymptotes for
an example case. The right panel depicts the fraction of the computation time spent on evaluating tip asymptote relative to the total time for a single ILSA
iteration.
to the crossing of the thin layer. In the case of the stress-corrected asymptote, the crossing occurs when a tip element is created
above the thin layer. For the universal asymptote, the thin layer is considered crossed when the survey element appears above the
thin layer since the universal asymptote does not ‘‘feel’’ the stress layer. As can be seen from the figure, for the case of universal
asymptote, the time of crossing varies significantly versus mesh. On the other hand, in the case of the stress-corrected asymptote, the
evolution of fracture height and the thin layer crossing time are practically independent of the mesh size. Thus, the stress-corrected
asymptote can significantly increase the accuracy of simulations when thin layers are present in the model.
6.4. Comparison of computational performance
The application of ODE asymptote for the fracture front tracking requires solving an ordinary differential equation numerically
for each survey and tip element for every time step and for every iteration. This is computationally more demanding than evaluating
the universal asymptote for the same occasions. To ensure that the added complexity does not outweigh the accuracy benefits, it
is worth estimating computational performance of ILSA algorithm that is utilizing the tip asymptote with stress layers. The overall
computation time of the algorithm depends on many factors, such as mesh, boundary conditions, configuration of the reservoir
layers, etc. To address the aforementioned problem of computational efficiency, we consider the case of the moderate symmetric
stress barriers for the 𝑀 regime, which was also considered in Section 6.2.
The left panel in Fig. 17 shows the total computation time for the considered example for the universal asymptote and the ODE
approach versus the number of elements 𝑛𝑐 in the central layer. The computation time for the algorithm with the ODE asymptote is
somewhat higher than for its counterpart with the universal asymptote. However, as the number of elements increases, the difference
in the computation time is practically vanished. To further understand the computational performance, the right panel in Fig. 17
depicts the fraction of the asymptote computation time relative to the total time for a single ILSA iteration. For a coarse mesh the
number of the tip elements is relatively large relative to the total number of elements. The former scales with the perimeter of the
fracture, while the latter scales with its area. As a result, the computational cost associated with the asymptote evaluation dominates
17
over the rest of the calculation for coarse meshes. For finer meshes, on the other hand, the computation time associated with solving
the system of equations becomes the bottle neck and the fraction of the time spent on the asymptote evaluation becomes negligible.
Finally, it is important to note that the use of the ODE asymptote increases accuracy, as is shown in the previous section.
Therefore, the increased computation time can be balanced by the possibility of using a coarser mesh.
7. Conclusions
This paper presents a fracture front tracking algorithm for an implicit level set algorithm (ILSA) that accounts for the effect of
multiple stress layers. The front tracking procedure employs the stress-corrected asymptotic solution that accounts for the effect
of stress layers in the multiscale tip asymptote [51]. To apply the latter stress-corrected asymptote, the core components of the
original ILSA approach, such as locating the fracture front and tip volume calculation, have been extended to include the effect
of multiple horizontal stress layers. Additionally, we have determined the stress relaxation factor that does not depend on the
geometry of the reservoir layers to increase the validity region of the stress-corrected asymptote in the case of a finite planar
fracture. These modifications significantly increase the accuracy of ILSA algorithm for a layered formation, especially in the case
of a coarse computational mesh. The accuracy of the presented solver is verified by comparing the results to reference numerical
calculations for a radial hydraulic fracture in different propagation regimes and experimental results conducted for a transparent
PMMA block. In order to illustrate the fracture closure algorithm, we consider the case of three-layered geometry, such that there
are two disconnected segments after closure.
In addition, we have investigated the sensitivity of fracture geometry to mesh size since the primary goal of using the stress-
corrected asymptote is to improve accuracy for coarse meshes. For a height-contained fracture, the height error corresponds to the
growth of approximately one element outside the layer boundary for the case of the universal asymptote. While in the algorithm with
the stress-corrected asymptote, the fracture growth is arrested almost precisely at the location of the stress barrier for all meshes.
Additionally, both algorithms with the universal asymptote and the stress-corrected asymptote predict a notably smaller length than
the reference solution. This discrepancy can be related to the reduction of accuracy of the asymptotic solution for high fracture front
curvatures. In the case of a moderate stress barrier, the stress-corrected asymptote provides a more accurate prediction of height and
length compared to the universal asymptote. Furthermore, the accuracy of the developed ILSA algorithm with the stress-corrected
asymptote has been examined for the case of a thin layer with high confining stress. It is concluded that the developed algorithm
provides a very accurate height prediction for all meshes, which is in contrast to the standard ILSA implementation with the universal
asymptotic solution that shows strong mesh sensitivity.
CRediT authorship contribution statement
A.V. Valov: Writing – review & editing, Writing – original draft, Visualization, Validation, Software, Methodology, Investigation,
Conceptualization. E.V. Dontsov: Writing – review & editing, Writing – original draft, Supervision, Methodology, Investigation,
Conceptualization. A.N. Baykin: Writing – review & editing, Project administration, Methodology, Conceptualization. S.V. Golovin:
Supervision, Funding acquisition, Conceptualization.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared
to influence the work reported in this paper.
Data availability
Data will be made available on request.
Appendix A. Discretized governing equations
In order to obtain discretized equations, a piece-wise constant approximation of the fracture width 𝑤(𝑥, 𝑦, 𝑡) and the fluid pressure
𝑝(𝑥, 𝑦, 𝑡) is employed. It is assumed that the fracture width 𝑤, the fluid pressure 𝑝, and the fracture footprint are known at the previous
time step 𝑡 − 𝛥𝑡.
Using the displacement discontinuity method [57] and assuming a piece-wise constant fracture width approximation, the
elasticity equation (1) in the case of non-constrained fracture (𝑤(𝑥, 𝑦, 𝑡) > 𝑤min (𝑥, 𝑦, 𝑡) and 𝑇𝑛 = 𝑝) takes the following form
∑
𝑝𝑖,𝑗 (𝑡) = 𝜎ℎ𝑖,𝑗 + 𝐶𝑖,𝑗;𝑘,𝑙 𝑤𝑘,𝑙 (𝑡), (A.1)
𝑘, 𝑙
where 𝐶𝑖,𝑗;𝑘,𝑙 are the coefficients of the elasticity matrix representing the influence of element (𝑘, 𝑙) on element (𝑖, 𝑗). For the case of
a homogeneous linear elastic medium, as considered in this study, the elasticity matrix can be computed as [26]
√ 𝑥=𝑥 +𝛥𝑥∕2, 𝑦=𝑦𝑙 +𝛥𝑦∕2
⎡ ⎤ 𝑘
′ ⎢ (𝑥𝑖 − 𝑥)2 + (𝑦𝑗 − 𝑦)2 ⎥
𝐸
(A.2)
8𝜋 ⎢⎢ ⎥
𝐶𝑖,𝑗;𝑘,𝑙 = − ,
(𝑥𝑖 − 𝑥)(𝑦𝑗 − 𝑦) ⎥
⎣ ⎦𝑥=𝑥𝑘 −𝛥𝑥∕2, 𝑦=𝑦𝑙 −𝛥𝑦∕2
18
𝑥=𝑥 ,𝑦=𝑦
where [𝑓 ]𝑥=𝑥21 ,𝑦=𝑦21 = 𝑓 (𝑥1 , 𝑦1 ) + 𝑓 (𝑥2 , 𝑦2 ) − 𝑓 (𝑥1 , 𝑦2 ) − 𝑓 (𝑥2 , 𝑦1 ).
The lubrication equation (5) is discretized using the finite volume method, as in [44]. By using backward Euler’s scheme for the
time integration from 𝑡 − 𝛥𝑡 to 𝑡, we obtain the following system of equations
𝑤𝑖,𝑗 (𝑡) − 𝑤𝑖,𝑗 (𝑡 − 𝛥𝑡) = 𝛥𝑡 [𝐴𝑝]𝑖,𝑗 + 𝛥𝑡 𝑄𝑖,𝑗 − 𝛥𝑖,𝑗 , (A.3)

where 𝑄𝑖,𝑗 denotes the source term, 𝛥𝑖,𝑗 captures the fluid leak-off for the element (𝑖, 𝑗) over time [𝑡−𝛥𝑡, 𝑡]. For the channel elements,
the leak-off term can be calculated as
(√ √ )
𝛥𝑖,𝑗 = 2𝐶 ′ 𝑡 − 𝑡0𝑖,𝑗 − 𝑡 − 𝛥𝑡 − 𝑡0𝑖,𝑗 , (A.4)
where 𝑡0𝑖,𝑗 is the time instance at which the fracture front was located at the center of the channel element (𝑥𝑖 , 𝑦𝑗 ). The fluid fluxes
across the element edges are calculated using the central difference
[ ]
1 𝑝𝑖+1,𝑗 − 𝑝𝑖,𝑗 𝑝𝑖,𝑗 − 𝑝𝑖−1,𝑗
[𝐴𝑝]𝑖,𝑗 = 𝑀𝑖+ 1 ,𝑗 − 𝑀𝑖− 1 ,𝑗
𝛥𝑥 2 𝛥𝑥 2 𝛥𝑥
[ ]
1 𝑝𝑖,𝑗+1 − 𝑝𝑖,𝑗 𝑝𝑖,𝑗 − 𝑝𝑖,𝑗−1
+ 𝑀𝑖,𝑗+ 1 − 𝑀𝑖,𝑗− 1 , (A.5)
𝛥𝑦 2 𝛥𝑦 2 𝛥𝑦
where the corresponding fluid mobilities are defined as
( 3 )
1 𝑤𝑖±1,𝑗 𝑤3𝑖,𝑗
𝑀𝑖± 1 ,𝑗 = ′
+ ′ . (A.6)
2 2 𝜇𝑖±1,𝑗 𝜇𝑖,𝑗
In order to incorporate the asymptotic solution into the front tracking algorithm it is convenient to split variables into the channel
and tip values. In the following expressions, we will use the superscript ‘‘𝑐’’ for the channel variables and the superscript ‘‘𝑡’’ for the
tip variables.
The elasticity matrix C is calculated using Eq. (A.2), the flux matrix 𝐴 in Eq. (A.5) is computed as A = 𝐴(𝒘). The tip widths
𝒘𝑡 are determined by integrating the asymptotic solution. Thus, in the tip elements, the fluid pressure vector 𝒑𝑡 is the primary
unknown. In the channel elements, the fluid pressure vector 𝒑𝑐 is determined using Eq. (A.1) and the fracture width vector 𝒘𝑐 is
unknown. Substituting (A.1) for the channel elements in Eq. (A.3) we obtain the following system of equations
[ ][ ] [ ]
I − 𝛥𝑡A𝑐𝑐 C𝑐𝑐 −𝛥𝑡A𝑐𝑡 𝒘𝑐 𝑹𝑐
𝑡𝑐 𝑐𝑐 𝑡𝑡 𝑡 = 𝑡 𝑡 , (A.7)
− 𝛥𝑡A C −𝛥𝑡A 𝒑 𝑹 −𝒘
where I is the identity matrix, the double superscripts denote corresponding tip-channel sub-matrices, and the right-hand side is
defined as follows
( )
𝑹𝐽 = 𝒘𝐽0 + 𝛥𝑡 A𝐽 𝑐 C𝑐𝑡 𝒘𝑡 + 𝑸𝐽 + 𝜟𝝈 𝐽 − 𝜟𝐽 , 𝐽 = 𝑐, 𝑡. (A.8)
The impact of stress layers in the above equation is calculated as 𝜟𝝈 𝐽 = A𝐽 𝑐 𝝈 𝑐ℎ + A𝐽 𝑡 𝝈 𝑡ℎ . The vector 𝒘0 denotes the known fracture
width from the previous time step 𝑡 − 𝛥𝑡. The fluid leak-off term for the channel elements 𝜟𝑐 is calculated using (A.4). Calculation
of the fluid leak-off 𝜟𝑡 is described in Appendix B.
It is worth noting that the elasticity matrix C is dense, while the flux matrix A is sparse. The system (A.7) is the nonlinear system
because the matrix A = A(𝒘) depends on the current solution. Note that that system of Eqs. (A.7) is suitable only for the case of a
non-constrained fracture i.e. 𝑤𝑖,𝑗 > 𝑤min𝑖,𝑗 and 𝑇𝑛 = 𝑝.
If a contact occurs in the (𝑖, 𝑗) cell, then the conditions 𝑤𝑖,𝑗 = 𝑤min𝑖,𝑗 and 𝑇𝑛 ≠ 𝑝 are satisfied. Hence, the fracture width becomes a
known parameter (is equal to the prescribed minimal width 𝑤min𝑖,𝑗 ), while the normal traction 𝑇𝑛𝑖,𝑗 becomes unknown. Here one can
consider an analogy with the tip elements, where the fracture width for the tip elements is determined by the asymptotic solution.
Hence, contact elements can be handled similarly, where the fracture width at the constrained elements is defined by the contact
condition (2). Thus, we need to add one more group of elements in addition to the channel and tip variables. This group contains
the variables in the elements with active constraint condition (2). The values corresponding to the contact cells are indicated by the
superscript ‘‘d’’ (deactivated). Then, given the presence of contact elements, the system of Eqs. (A.7) can be rewritten as follows
−𝛥𝑡A𝑐𝑑 ⎤ ⎡𝒘𝑐 ⎤ ⎡ ̃𝑐
𝑹 ⎤
⎡I − 𝛥𝑡A𝑐𝑐 C𝑐𝑐 −𝛥𝑡A𝑐𝑡
⎢ −𝛥𝑡A𝑡𝑐 C𝑐𝑐 ⎥ ⎢ ⎥ ⎢ ̃ 𝑡 ⎥
−𝛥𝑡A𝑡𝑡 −𝛥𝑡A𝑡𝑑 𝑡
𝒑 = ⎢ 𝑹 − 𝒘 ⎥, 𝑡 (A.9)
⎢ ⎥ ⎢ ⎥
⎣ −𝛥𝑡A𝑑𝑐 C𝑐𝑐 −𝛥𝑡A𝑑𝑡 −𝛥𝑡A𝑑𝑑 ⎦ ⎣𝑻 𝑛 ⎦ ⎢⎣𝑹
𝑑
̃ − 𝒘𝑑 ⎥⎦
𝑑
min
where the right-hand-side is defined as
̃ 𝐽 = 𝑹𝐽 + 𝛥𝑡 A𝐽 𝑐 C𝑐𝑑 𝒘𝑑 ,
𝑹 𝐽 = 𝑐, 𝑡, 𝑑. (A.10)
min
Despite the fact that the initial fracture does not affect the solution after a short period of time [75], it is important to specify
an initial condition to initialize ILSA algorithm. Fig. A.18 shows the scheme of the initial fracture that is used for initialization. At
least one survey element is required to initialize ILSA algorithm. This element is located at the origin and is surrounded by eight tip
elements. The fracture front crosses the midpoints of the tip elements located immediately above, below, or on the side from the
survey element. This results in an octagonal initial fracture shape. The width of the initial fracture is assumed to be zero for both
survey and tip elements. Thus, the initial fracture volume does not introduce an error to the overall fracture volume, regardless of
the mesh size.
19
Fig. A.18. Scheme of the initial fracture that is used to initialize ILSA algorithm.
Appendix B. Tip leak-off calculation
This section aims to describe a simplified approach for calculating the tip leak-off volume based on the midpoint rule. Since a tip
element is partially filled, see Fig. 5, the leak-off term should be integrated only over the internal part of the tip element. According
to (5) and (A.3), the full leak-off is obtained by integration using the midpoint rule
𝑆 √
(𝑡) = 2𝐶 ′ 𝑡 − 𝑡0𝑚 , (B.1)
𝛥𝑥𝛥𝑦
where 𝑆 is the area of the 𝐴𝐷𝐸𝐹 𝐺 polygon, 𝑡0𝑚 is the time at which the fracture front was located at the center of 𝐴𝐷𝐸𝐹 𝐺 polygon.
In the case of 𝛼 ≠ 0 and 𝛼 ≠ 𝜋∕2 the value of 𝑆 is calculated as 𝑆𝐴𝐵𝐶 − 𝑆𝐷𝐵𝐸 − 𝑆𝐺𝐹 𝐶 . If 𝑙 ≤ 𝛥𝑥 cos 𝛼 then triangle 𝐷𝐵𝐸 disappears,
while in the case of 𝑙 ≤ 𝛥𝑦 sin 𝛼 the triangle 𝐺𝐹 𝐶 disappears. In the case of 𝛼 = 0 or 𝛼 = 𝜋∕2, the area becomes rectangular. Finally,
the value of 𝑆 is calculated as follows
⎧𝑆▵ (𝑙) − 𝑆▵ (𝑙 − 𝛥𝑥 cos 𝛼) − 𝑆▵ (𝑙 − 𝛥𝑦 sin 𝛼), 𝛼 ≠ 0, 𝛼 ≠ 𝜋∕2
⎪
𝑆 = ⎨𝑙𝛥𝑦, 𝛼=0 (B.2)
⎪
⎩𝑙𝛥𝑥, 𝛼 = 𝜋∕2,
where 𝑆▵ (𝑙) = 𝑙2 ∕ sin 2𝛼 for 𝑙 > 0 and 𝑆▵ (𝑙) = 0 otherwise. The time 𝑡0𝑚 is implicitly determined by the following equation
𝑡
𝑙
𝑉 (𝜏) d𝜏 = . (B.3)
∫𝑡0𝑚 2
References
[1] Economides MJ, Nolte KG, et al. Reservoir stimulation, vol. 2. Prentice Hall Englewood Cliffs, NJ; 1989.
[2] Jeffrey R, Mills K. Hydraulic fracturing applied to inducing longwall coal mine goaf falls. In: North America rock mechanics symposium. 2000, All Days,
ARMA-2000-0423.
[3] Katsaga T, Riahi A, DeGagne D, Valley B, Damjanac B. Hydraulic fracturing operations in mining: conceptual approach and dfn modeling example. Min
Technol 2015;124(4):255–66.
[4] Fu P, Settgast RR, Hao Y, Morris JP, Ryerson FJ. The influence of hydraulic fracturing on carbon storage performance. J Geophys Res 2017;122(12):9931–49.
[5] Huerta NJ, Cantrell KJ, White SK, Brown CF. Hydraulic fracturing to enhance injectivity and storage capacity of CO2 storage reservoirs: Benefits and risks.
Int J Greenhouse Gas Control 2020;100:103105.
[6] Kumari W, Ranjith P, Perera M, Li X, Li L, Chen B, et al. Hydraulic fracturing under high temperature and pressure conditions with micro CT applications:
Geothermal energy from hot dry rocks. Fuel 2018;230:138–54.
[7] Cui G, Wang W, Dou B, Liu Y, Tian H, Zheng J, et al. Geothermal energy exploitation and power generation via a single vertical well combined with
hydraulic fracturing. J Energy Eng 2022;148(1):04021058.
[8] Akdas SB, Onur M. Analytical solutions for predicting and optimizing geothermal energy extraction from an enhanced geothermal system with a multiple
hydraulically fractured horizontal-well doublet. Renew Energy 2022;181:567–80.
[9] Rubin AM. Propagation of magma-filled cracks. Ann Rev Earth Planet Sci 1995;23(1):287–336.
20
[10] Roper S, Lister J. Buoyancy-driven crack propagation: the limit of large fracture toughness. J Fluid Mech 2007;580:359–80.
[11] Rivalta E, Taisne B, Bunger A, Katz R. A review of mechanical models of dike propagation: Schools of thought, results and future directions. Tectonophysics
2015;638:1–42.
[12] Khristianovic S, Zheltov YP. Formation of vertical fractures by means of highly viscous liquid. In: World petroleum congress proceedings. 1955, p. 579–86.
[13] Geertsma J, De Klerk F. A rapid method of predicting width and extent of hydraulically induced fractures. J Petrol Technol 1969;21(12):1571–81.
[14] Nordgren R. Propagation of a vertical hydraulic fracture. Soc Petrol Eng J 1972;12(04):306–14.
[15] Perkins T, Kern L, et al. Widths of hydraulic fractures. J Pet Technol 1961;13(09):937–49.
[16] Geertsma J, De Klerk F. A rapid method of predicting width and extent of hydraulically induced fractures. J Petrol Technol 1969;21(12):1571–81.
[17] Abe H, Keer L, Mura T. Growth rate of a penny-shaped crack in hydraulic fracturing of rocks, 2. J Geophys Res 1976;81(35):6292–8.
[18] Palmer ID, Carroll H. Three-dimensional hydraulic fracture propagation in the presence of stress variations. Soc Petrol Eng J 1983;23(06):870–8.
[19] Settari A, Cleary MP. Development and testing of a pseudo-three-dimensional model of hydraulic fracture geometry. SPE Prod Eng 1986;1(06):449–66.
[20] Dontsov E, Peirce A. An enhanced pseudo-3D model for hydraulic fracturing accounting for viscous height growth, non-local elasticity, and lateral toughness.
Eng Fract Mech 2015;142:116–39.
[21] Cohen C-E, Kresse O, Weng X. A new stacked height growth model for hydraulic fracturing simulation. In: 49th US rock mechanics/geomechanics
symposium. OnePetro; 2015.
[22] Zhang X, Wu B, Jeffrey RG, Connell LD, Zhang G. A pseudo-3D model for hydraulic fracture growth in a layered rock. Int J Solids Struct 2017;115:208–23.
[23] Zhang X, Wu B, Connell L, Han Y, Jeffrey R. A model for hydraulic fracture growth across multiple elastic layers. J Pet Sci Eng 2018;167:918–28.
[24] Barree RD. A practical numerical simulator for three-dimensional fracture propagation in heterogeneous media. In: SPE reservoir simulation symposium.
OnePetro; 1983.
[25] Clifton RJ, Abou-Sayed AS. A variational approach to the prediction of the three-dimensional geometry of hydraulic fractures. In: SPE/DOE low permeability
gas reservoirs symposium. OnePetro; 1981.
[26] Peirce A, Detournay E. An implicit level set method for modeling hydraulically driven fractures. Comput Methods Appl Mech Engrg
2008;197(33–40):2858–85.
[27] Zia H, Lecampion B. PyFrac: A planar 3D hydraulic fracture simulator. Comput Phys Comm 2020;255:107368. http://dx.doi.org/10.1016/j.cpc.2020.107368.
[28] Cherny S, Lapin V, Esipov D, Kuranakov D, Avdyushenko A, Lyutov A, Karnakov P. Simulating fully 3D non-planar evolution of hydraulic fractures. Int
J Fract 2016;201(2):181–211.
[29] Paul B, Faivre M, Massin P, Giot R, Colombo D, Golfier F, et al. 3D coupled HM–XFEM modeling with cohesive zone model and applications to non
planar hydraulic fracture propagation and multiple hydraulic fractures interference. Comput Methods Appl Mech Engrg 2018;342:321–53.
[30] Baykin A, Golovin S. Application of the fully coupled planar 3D poroelastic hydraulic fracturing model to the analysis of the permeability contrast impact
on fracture propagation. Rock Mech Rock Eng 2018;51:3205–17.
[31] Kumar D, Ghassemi A. A three-dimensional analysis of simultaneous and sequential fracturing of horizontal wells. J Pet Sci Eng 2016;146:1006–25.
[32] Yoshioka K, Bourdin B. A variational hydraulic fracturing model coupled to a reservoir simulator. Int J Rock Mech Min Sci 2016;88:137–50.
[33] Salimzadeh S, Usui T, Paluszny A, Zimmerman RW. Finite element simulations of interactions between multiple hydraulic fractures in a poroelastic rock.
Int J Rock Mech Min Sci 2017;99:9–20.
[34] Shauer N, Duarte CA. A three-dimensional generalized finite element method for simultaneous propagation of multiple hydraulic fractures from a wellbore.
Eng Fract Mech 2022;265:108360.
[35] Rezaei A, Siddiqui F, Bornia G, Soliman M. Applications of the fast multipole fully coupled poroelastic displacement discontinuity method to hydraulic
fracturing problems. J Comput Phys 2019;399:108955.
[36] Chen M, Zhang S, Li S, Ma X, Zhang X, Zou Y. An explicit algorithm for modeling planar 3D hydraulic fracture growth based on a super-time-stepping
method. Int J Solids Struct 2020;191:370–89.
[37] Wang H, Liu H, Wu H, Zhang G, Wang X. A 3D nonlinear fluid-solid coupling model of hydraulic fracturing for multilayered reservoirs. Petrol Sci Technol
2012;30(21):2273–83.
[38] Lecampion B, Peirce A, Detournay E, Zhang X, Chen Z, Bunger A, et al. The impact of the near-tip logic on the accuracy and convergence rate of hydraulic
fracture simulators compared to reference solutions. In: Bunger A, McLennan J, Jeffrey R, editors. Effective and sustainable hydraulic fracturing. 2013, p.
855–73.
[39] Gordeliy E, Peirce A. Coupling schemes for modeling hydraulic fracture propagation using the XFEM. Comput Methods Appl Mech Engrg 2013;253:305–22.
[40] Gordeliy E, Peirce A. Implicit level set schemes for modeling hydraulic fractures using the XFEM. Comput Methods Appl Mech Engrg 2013;266:125–43.
[41] Gordeliy E, Peirce A. Enrichment strategies and convergence properties of the XFEM for hydraulic fracture problems. Comput Methods Appl Mech Engrg
2015;283:474–502.
[42] Peirce A. Modeling multi-scale processes in hydraulic fracture propagation using the implicit level set algorithm. Comput Methods Appl Mech Engrg
2015;283:881–908.
[43] Peirce A. Implicit level set algorithms for modelling hydraulic fracture propagation. Phil Trans R Soc A 2016;374(2078):20150423.
[44] Dontsov E, Peirce A. A multiscale implicit level set algorithm (ILSA) to model hydraulic fracture propagation incorporating combined viscous, toughness,
and leak-off asymptotics. Comput Methods Appl Mech Engrg 2017;313:53–84.
[45] Linkov AM. Modern theory of hydraulic fracture modeling with using explicit and implicit schemes. 2019, arXiv preprint arXiv:1905.06811.
[46] Dontsov E, Peirce A. A non-singular integral equation formulation to analyse multiscale behaviour in semi-infinite hydraulic fractures. J Fluid Mech
2015;781.
[47] Dontsov E, Hewson C, McClure M. A comparison of hydraulic fracture front tracking algorithms. In: 56th US rock mechanics/geomechanics symposium.
OnePetro; 2022.
[48] Dontsov EV. A homogenization approach for modeling a propagating hydraulic fracture in a layered material. Geophysics 2017;82(6):MR153–62.
[49] Dontsov E. A continuous fracture front tracking algorithm with multi layer tip elements (MuLTipEl) for a plane strain hydraulic fracture. J Pet Sci Eng
2022;217:110841.
[50] Dontsov E, Hewson C, McClure M. A new crack propagation algorithm that enables accurate simulation of propagation across thin layers in a practical
field-scale fracturing model. In: Proceedings of the SPE hydraulic fracturing technology conference. 2022.
[51] Valov AV, Dontsov EV. On the layer crossing problem for a semi-infinite hydraulic fracture. 2023, arXiv:2305.13176.
[52] Garagash D, Detournay E. The tip region of a fluid-driven fracture in an elastic medium. J Appl Mech 2000;67(1):183–92.
[53] Lecampion B, Detournay E. An implicit algorithm for the propagation of a hydraulic fracture with a fluid lag. Comput Methods Appl Mech Engrg
2007;196(49–52):4863–80.
[54] Carter E. Optimum fluid characteristics for fracture extension. In: C. HG, R. FC, editors. Drilling and production practices. 1957, p. 261–70.
[55] Detournay E. Mechanics of hydraulic fractures. Annu Rev Fluid Mech 2016;48:311–39.
[56] Lecampion B, Bunger A, Zhang X. Numerical methods for hydraulic fracture propagation: A review of recent trends. J Nat Gas Sci Eng 2018;49:66–83.
[57] Crouch S, Starfield AM. Boundary element methods in solid mechanics: With applications in rock mechanics and geological engineering. Allen & Unwin;
1983.
[58] Hills DA, Kelly P, Dai D, Korsunsky A. Solution of crack problems: The distributed dislocation technique, vol. 44. Springer Science & Business Media;
2013.
21
[59] Rice JR. Mathematical analysis in the mechanics of fracture. Fract Adv Treatise 1968;2:191–311.
[60] Detournay E, Peirce A. On the moving boundary conditions for a hydraulic fracture. Internat J Engrg Sci 2014;84:147–55.
[61] Dormand J, Prince P. A family of embedded Runge-Kutta formulae. J Comput Appl Math 1980;6(1):19–26. http://dx.doi.org/10.1016/0771-050X(80)90013-
3.
[62] Macdonald CB. The predicted sequential regularization method for differential-algebraic equations [Ph.D. thesis], Acadia University; 2001.
[63] Sethian JA. A fast marching level set method for monotonically advancing fronts. Proc Natl Acad Sci 1996;93(4):1591–5.
[64] Bærentzen JA. On the implementation of fast marching methods for 3D lattices. Tech. rep. informatics and mathematical modelling, IMM-TR-2001-13,
Richard Petersens Plads, Building 321, DK-2800 Kgs. Lyngby: Technical University of Denmark, DTU; 2001.
[65] Malladi R, Sethian JA. Level set and fast marching methods in image processing and computer vision. In: Proceedings of 3rd IEEE international conference
on image processing, vol. 1. IEEE; 1996, p. 489–92.
[66] Madyarova MV. Fluid-driven penny-shaped fracture in elastic medium [Ph.D. thesis], University of Minnesota; 2003.
[67] Dontsov E. An approximate solution for a penny-shaped hydraulic fracture that accounts for fracture toughness, fluid viscosity and leak-off. Royal Soc
Open Sci 2016;3(12):160737.
[68] Savitski A, Detournay E. Propagation of a penny-shaped fluid-driven fracture in an impermeable rock: asymptotic solutions. Int J Solids Struct
2002;39(26):6311–37.
[69] Detournay E. Propagation regimes of fluid-driven fractures in impermeable rocks. Int J Geomech 2004;4(1):35–45.
[70] Chen X, Zhao J, Yan W, Zhang X. Numerical investigation into the simultaneous growth of two closely spaced fluid-driven fractures. SPE J
2019;24(01):274–89.
[71] Shovkun I, Espinoza DN. Propagation of toughness-dominated fluid-driven fractures in reactive porous media. Int J Rock Mech Min Sci 2019;118:42–51.
[72] Wu R, Bunger A, Jeffrey R, Siebrits E, et al. A comparison of numerical and experimental results of hydraulic fracture growth into a zone of lower
confining stress. In: The 42nd US rock mechanics symposium. American Rock Mechanics Association; 2008.
[73] Dontsov E. Analysis of a constant height hydraulic fracture. In: Mechanics of hydraulic fracturing: Experiment, model, and monitoring. Wiley Online
Library; 2022, p. 127–39.
[74] Dontsov E, Peirce A. Comparison of toughness propagation criteria for blade-like and pseudo-3D hydraulic fractures. Eng Fract Mech 2016;160:238–47.
[75] Zia H, Lecampion B, Zhang W. Impact of the anisotropy of fracture toughness on the propagation of planar 3D hydraulic fracture. Int J Fract
2018;211:103–23.
22

1 s2.0 S0013794423006203 Main

Uploaded by

Copyright:

Available Formats

1 s2.0 S0013794423006203 Main

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

1 s2.0 S0013794423006203 Main

Uploaded by

Copyright:

Available Formats

Engineering Fracture Mechanics 292 (2023) 109662

Contents lists available at ScienceDirect

Engineering Fracture Mechanics

An implicit level set algorithm for hydraulic fracturing with a

ARTICLE INFO ABSTRACT

2.1. Elasticity equation

𝑤(𝑥, 𝑦, 𝑡) ≥ 𝑤min (𝑥, 𝑦, 𝑡). (2)

2.2. Fluid flow

2.3. Conditions at the fracture front

3.1. The layer crossing problem

3.2. Tip asymptote from integral equation

3.3. Toughness corrected asymptote

3.4. ODE approximation for the asymptote

d𝑤̂𝑎 𝛽𝑚3 𝜒𝛽𝑚4̃

3.5. Locating the fracture front

3.6. Tip volume calculation

𝑉𝐴𝐷𝐸𝐹 𝐺 = 𝑉𝐴𝐵𝐶 − (𝓁 − 𝛥𝑦 sin 𝛼)𝑉𝐺𝐹 𝐶 − (𝓁 − 𝛥𝑥 cos 𝛼)𝑉𝐷𝐵𝐸 , (19)

𝑤𝑎 (𝑠) = 𝑤𝑢𝑎 (𝑠) cos2 𝛼 + 𝑤𝜎𝑎 (𝑠) sin2 𝛼. (22)

4. Stress relaxation for the asymptote

where 𝑘 represents the relaxation distance measured in the number of elements.

5.1. Penny-shaped fracture

5.2. Comparison with experimental observations

A fluid with viscosity 𝜇 = 30 Pa s is pumped with a piece-wise constant flow rate

5.3. Fracture closure

6.1. Constant height hydraulic fracture

6.2. Symmetric stress barriers

6.3. Thin high stress layer

𝑦1 = −7.5 m, 𝑦2 = 7.5 m, 𝑦3 = 12.5 m, 𝑦4 = 37.5 m,

6.4. Comparison of computational performance

CRediT authorship contribution statement

Declaration of competing interest

Data will be made available on request.

Appendix A. Discretized governing equations

𝑤𝑖,𝑗 (𝑡) − 𝑤𝑖,𝑗 (𝑡 − 𝛥𝑡) = 𝛥𝑡 [𝐴𝑝]𝑖,𝑗 + 𝛥𝑡 𝑄𝑖,𝑗 − 𝛥𝑖,𝑗 , (A.3)

Appendix B. Tip leak-off calculation

You might also like