Computers and Geotechnics 122 (2020) 103502

Contents lists available at ScienceDirect

Computers and Geotechnics

journal homepage: www.elsevier.com/locate/compgeo

Research Paper

Finite particle method for static deformation problems solved using JFNK T
method
Ding Chena, Wenxiong Huanga, , Andrei Lyaminb
⁎

a
College of Mechanics & Materials, Hohai University, China
b
PRRCGSE, The University of Newcastle, Australia

ARTICLE INFO ABSTRACT

Keywords: In this work, the Finite Particle Method (FPM) which is a typical meshfree particle method is formulated for
Finite particle method (FPM) static deformation and plastic flow problems. The novel formulation utilizes the Lagrangian-type kernel, which
Equilibrium problem guarantees the method being free from the tension instability. A pseudo-viscosity is introduced into the particle
Implicit integration system to prevent the hourglass phenomenon in the numerical procedure. The nonlinear system of governing
Jacobian-free Newton-Krylov (JFNK)
equations is solved with implicit time integration scheme. For numerical efficiency the Jacobian-Free Newton-
Krylov (JFNK) method is applied to obtain the solution for each time step. Numerical examples are provided to
demonstrate the validity and the effectiveness of the method in capturing finite deformation with moderate
distortion.

1. Introduction address a particular numerical difficulty known as tension instability,

which occurs when tensile stresses develop in the system (Swegle et al.
The Finite Particle Method (FPM) discussed in this work is a [17]; Belytschko et al. [18]). To overcome this issue, Monaghan [19]
meshless particle method, which was developed on the basis of the proposed to include an artificial stress in the numerical model. Bonet
Smoothed Particle Hydrodynamics (SPH). The SPH method was ori- and Kulasegaram [20] demonstrated that the pathetic phenomenon can
ginally invented for solving astronomical and hydrodynamic problems be avoided by using the total Lagrangian method. Recently, Chen et al.
[1,2]. Recent years the method has also been applied to solid mechanics [21] formulated an updated Lagrangian method with a Lagrange-type
problems. Without special treatments, the conventional SPH solutions kernel, which allows the support domain of particles to deform with the
suffer from lower accuracy near boundaries, which is due to incomplete configuration such that the original particle interactions are kept. The
support domain of the near-boundary particles. To deal with this pro- method has been demonstrated to be free from tension instability. In
blem, Chen et al. [3] proposed a modified formulation termed Cor- contrast, Vidal et al. [22] showed that tension instability problem can
rected Smoothed Particle Method (CSPM) by using a Taylor series ex- also be avoided if the particle interactions as well as the reference
pansion. Bonet et al. [4] also presented a modified SPH formulation that configuration are updated only sporadically.
can ensure conservation of linear and angular momentum. Also based Conventionally, the SPH-based particle methods are formulated for
on the Taylor series expansion, Liu and his co-workers [5,6] proposed dynamic problems (e.g. Monaghan [19]; Wang & Chan [23]). For nu-
an alternative formulation termed the Finite Particle Method (FPM), in merical efficiency, the governing equations of the particle system are
which the values of a function and its partial derivatives are calculated usually integrated by an explicit scheme. In order to achieve solutions
simultaneously. The FPM may be regarded as an improved smoothed with sufficient accuracy and stability, small time increment steps are
particle method with better performance than SPH and CSPM near required. Such dynamic model is, however, not suitable for solving
boundaries and for using non-uniform particle-discretization [7–9]. static deformation problems, as the numerical procedure usually re-
The SPH-based particle methods are predominantly suitable for quires too many increment steps, and often renders certain oscillations
solving fluid mechanics problems or simulating fluid-like material be- to the solutions with reduced accuracy.
haviors, such as metal forging [10], landsliding [11,12], dam breaking This work extends the finite particle method (FPM) for solving static
[13], debris flow [14] and other complex fluid flows [15,16], etc. For deformation and plastic flow problems of solids. For this purpose, the
solving solid mechanics problems, the methods need be revised to governing equations of the particle system are formulated for

⁎
Corresponding author.
E-mail addresses: [email protected] (D. Chen), [email protected] (W. Huang), [email protected] (A. Lyamin).

https://doi.org/10.1016/j.compgeo.2020.103502
Received 29 August 2019; Received in revised form 30 January 2020; Accepted 18 February 2020
Available online 27 February 2020
0266-352X/ © 2020 Elsevier Ltd. All rights reserved.
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

equilibrium problems, and integrated implicitly, so that the size of in- represents the boundaries where surface traction and displacement are
crement loading step is less restricted. The updated Lagrangian for- prescribed respectively; u is the prescribed boundary displacement on
mulation that as proposed in Chen et al. [21] which is tension-in- t , and
u is a penalty factor enforcing the displacement boundary
stability free is adopted. To prevent the spurious zero-energy condition.
deformation mode from entering numerical solutions, pseudo viscosity With particle discretization of the continuum, Eq. (3) can approxi-
is introduced into the nonlinear system of equations. The emphasis will mately be written as
be put on the numerical technique for solving the nonlinear system of np np1 np2 np
governing equations. As a fair number of particles is usually included in aT aV a = uc Tt cAc + uc T (u c uc ) Ac + uaT b aV a.
the support domain of each particle, the FPM governing equations often a=1 c=1 c=1 a=1
render a broad bandwidth in the Jacobian matrix. Conventional (4)
Newton-Raphson method then becomes too costly as forming and
Here np represents the total number of particles, np1 and np2 re-
saving the Jacobian matrix is required on each iteration. In this work,
present the particles for the boundaries t t and t u respectively.
the Jacobian-Free Newton-Krylov (JFNK) method [24,25] is employed
Following Eq. (1), the particle approximation of the displacement and
for numerical efficiency. On the basis of Newton method, the Krylov
the rate of deformation are given by
subspace iteration with a GMRES (Generalized Minimal Residual)
ns
[26–28] scheme is applied for solving the linearized system of equa- a
ua = V bha, b ub = Had ;
tions. The method can avoid forming and saving the Jacobian matrix for (5)
b=1
Newton iterations, which is a prominent advantage. Finite deformation
problems of elastoplastic materials are considered. The stress is updated a
ns
a a
= V bBa, bu b = H d .
using a semi-implicit integration scheme [29]. Numerical examples are (6)
b= 1
presented for the validity of the method.
a
In these two expressions, d = [u11, u 21, u31, u12, u 22 , u32, …u1ns , u2ns , u3ns]T
2. Basic formulae of FPM is a column matrix containing displacement of particles in the influence
a
domain of particle a; Ha and H are matrices associated with particle a
2.1. Particle approximation in FPM and given by:
Ha = [V 1h0a,1 I , V 2h 0a,2 I, … , V nsh0a, ns I]; (7)
In the FPM, following Liu [5,30] & Chen et al. [21], the particle
approximation of a function of position such as displacement of a a
H = [V 1Ba,1, V 2Ba,2, … , V ns Ba, ns] (8)
continuum body ui (x j ) and its partial derivatives are calculated si-
multaneously, according to where I represents a 3 × 3 unit matrix, Ba, b (b = 1, 2, …, ns ) are sub-
matrices given by
ns ns
uia = h 0a, b uib V b; uia, j = hja, b (uib uia) V b . a, b
b=1 b=1 (1) h1 0 0
a, b
Herein the right superscript represents the particle the quantity is 0 h2 0
a, b
associated to, as uia = ui (xka) ; uia, j = ui x j xk = xka ; summation is 0 0 h3
carried out through all particles (1 to ns) in the influence domain Ba, b = a, b a, b
h2 h1 0
(support domain) of particle a; V b is the volume associated with particle
a, b a, b
b ; h0a, b and hja, b are auxiliary kernels given by: 0 h3 h2
a, b a, b
h 0a, b
ns b a, b ns b a, b b, a ns b a, b b, a ns b a, b b, a
1
W a, b
h3 0 h1 (9)
b=1 V W b = 1 V W x1 b = 1 V W x2 b=1 V W x3
ns b a, b ns a, b b, a ns a, b b, a ns a, b b, a a, b
h1a, b b b b W ,1
b = 1 V W ,1 b = 1 V W ,1 x1 b = 1 V W ,1 x2 b = 1 V W ,1 x 3 with
ns ns ns ns a, b
h2a, b b a, b
b = 1 V W ,2
b a, b b, a
b = 1 V W ,2 x1
b a, b b, a
b = 1 V W ,2 x2
b a, b b, a
b = 1 V W ,2 x 3
W ,2
ns
h3a, b ns b a, b
b = 1 V W ,3
ns b a, b b, a
b = 1 V W ,3 x1
ns b a, b b, a
b = 1 V W ,3 x2
ns b a, b b, a
b = 1 V W ,3 x 3
a, b
W ,3 a, b hja, a b=1
(V b V a) hja, b b = a,
hj = for (j = 1, 2, 3).
(2) hja, b b a,
(10)
where xkb, a = xkb xka (k = 1, 2, 3) and W a, b = W (xk xkb , ls ) a a
Similarly, we have u = d , = H d . By substituting them
Ha a
W (xi xib, ls )
xk = xka = W (xka xkb , ls ) , W ,aj , b = (j = 1, 2, 3) with into Eq. (4), the following governing equations for the particle system
xj
xi = xia are consequently obtained:
W (xk xk , ls ) being the kernel function and ls the smoothing length.
F (d, t ) = F int F ext = 0 (11)
2.2. FPM governing equation for quasi-static problems Herein
np aT a a
The present work concerns with the quasi-static deformation and F int =
a=1
ca TH V ,
plastic flow problems of continuum bodies. A loading or deformation F ext =
np1 c T c T c c
c H t A +
np
c aTHaT ba a V a
process is traced with a fictitious time t. For a continuum body in c=1 a=1
c
np2 cT c cT c c
equilibrium, the virtual work principle expressed in the current con- + c=1 c H (u H d ); (12)
figuration t at time t reads:
where d represents the vector of particle displacements arrayed in the
t
T dV = t uTt dA + t vT (u u) dA + t uT b dV . global order of sequence; c a is the selective matrix for assembly of the
t u
sub-matrices from the local order to the global order.
(3)
Here the matrix representations are used for vectors and symmetric 3. Numerical implementation
tensors; b and t represent the body force and surface traction applying
on the continuum body, with being the current mass density; is the 3.1. Newton method for solving governing equations
Cauchy stress and u the displacement; u represents the virtual dis-
placement with being the virtual strain derived from u ; t t and t u For a loading procedure described by the fictitious time t [0, T ],

2
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

the nonlinear system Eq. (11) will be solved incrementally. For a typical 3.3. JFNK method in combination with GMRES
increment step [tn, tn + 1], assuming that displacement dn have been ob-
tained for t = tn , equilibrium of the particle system at time tn + 1 is de- With the pseudo-viscous force term included, Eq. (11) becomes
scribed by
F (dn + 1) + F v (dn + 1, d) = F (dn) + J (dn) d + kv (dn) d + o ( d||2 )
F (dn + 1) = F (dn) + J (dn) d + o ( d||2 ) = 0, (13)
= 0. (21)
where d = dn + 1 dn is the displacement increment to be solved,
The Newton iteration scheme (14) is then modified to
· represents the Euclidean norm of a vector, and J = F d is the
Jacobian matrix. Notice here that the fixed time t = tn + 1 is omitted in J (dkn+ 1) dk + kv (dnk + 1) dk = Fnk + 1 Fnvk+ 1;
the expressions of F for simplicity. Based on Eq. (11), a Newton itera-
dk + 1 = dk + dk , dnk++11 = dn + dk + 1. (22)
tion scheme can be formulated:

J (dkn+ 1) dk = F (dkn+ 1); Herein Fkn+ 1 = F (dnk + 1) , Fnvk+ 1 = F v (dnk + 1, dk) . This scheme works as
the outer (nonlinear) iteration, which stops at meeting the convergence
dk + 1 = dk + dk , dnk++11 = dn + dk + 1. (14) conditions:
The iteration may be started with d0 = 0 and dn0 + 1 = dn .
Fnk++11 + Fnvk+ 1|| dk||
< F & < u,
Fn0 + 1|| dnk + 1|| (23)
3.2. Hourglass control
where F and u are two predefined error tolerances.
In most collocational meshless methods including SPH and FPM, For each fixed k, the large linear system (22)1 is then solved for
spurious deformation mode may enter to numerical solutions, leading unknown vector x : = dk using the Krylov subspace iteration combined
to incorrect results. A frequently used approach to control this hour- with the GMRES method (linear iteration). Starting with an initial guess
glass phenomenon is by introducing an artificial viscosity [2,31]. Other x 0 , the solution is corrected in the Krylov subspace Km , which is
effective techniques include using specialized algorithms [32,33] for or ~ ~ ~
spanned by vectors {r0, J r0, ( J )2r0, …, ( J )m 1r0} , where
using additional stress-integration points [34]. In this work, the first
~
approach is employed and adapted to the static problems. For a dy- J = J (dkn+ 1) + kv (dnk + 1) (24)
namic particle system, application of artificial viscosity is a natural
is the enhanced Jacobian, and
option. Between two particles in interaction, Monaghan [2] introduced
an artificial viscosity proportional to the relative velocity in the fol- r0 =
~
Jx 0 Fnk + 1 Fnvk+ 1 (25)
lowing form:
is the initial residual of Eq. (22)1. At the m-th step, the approximate
1 ls x abTv ab
ab = c ab := ab abT ab
x v , solution can be written as
ab
x abTx ab + (0.1ls )2 (15)
m (m )
xm = x 0 + j = 1 y j zj = x 0 + Zm y (m) , (26)
where x ab = x a x b , v ab = v a v b are the difference in position and
velocity of the particle pair; is a constant parameter, which is typi- where zj (j = 1, 2, …, m) are the orthonormal vectors of Km ,which can be
cally set to 1.0; ls is the mean smoothing length, and obtained by the Arnoldi algorithm; Zm = [z1, z2, …, zm] is a n × m ma-
ab 1 a b ), 1 trix.
= ( + c ab = ( E a + E b ).
2 2 (16) Let
The corresponding artificial viscous force applied to each particle is ~
Hm = ZTm J Zm. (27)
then given by
ns Hm is a m × m upper Hessenberg matrix. From Eqs. (22)1 and (26) it
fiva = V a Vb( abh a, b
i
bah b, a).
i
can be derived that the residual rm satisfies the following condition:
b=1 (17) ~
T
Zm T
+ 1rm = Zm + 1r0
T
Zm + 1 J Zm y
(m ) : = r e
0 1 Hm y (m) . (28)
To adapt to the static problems in this work, the difference of dis-
placement increment divided by the time increment is deployed to re- Here, e1 = [1, 0, …, 0]T(m + 1) × 1 is a unit vector in m + 1; Hm is a
place the relative velocity in Eq. (15): (m + 1) × m matrix containing matrix Hm . Vector y (m) is then de-
termined by minimizing the residual:
v ab ( ua ub) t. (18)
Thus, for hourglass control, a pseudo-viscous force vector in the y (m) = arg min ||r0 ||e1 Hm y .
following form will be added to the global particle system: y (29)
np np
a,1 a,2 a, ns a This GMRES scheme requires the action of the Jacobians matrix
F v (dn + , d) = c aTf va = c aTV a [k n + , k n + , …k n + ] d :
a=1 a=1
only in the matrix–vector products, which are calculated by a direct
approximation without forming and saving the Jacobian itself, as per
=kv (dn + ) d . (19)
~
J zj = Jzj + kvzj (F (dnk + 1 + z j) F (d kn + 1)) + F v (d kn + 1, zj). (30)
Herein

Vnb+
ab
(hna,+b hbn,+a ) x nab+T / t b a
Herein defines a small perturbation, which can be set to
a, b
(1 + zj ) c d kn + 1 with c being the computer precision.
n+
k n+ = ab ns
=
Vnb+ (hna,+b hnb,+a )(x ab T
(V b V a) x nab+T )/ t b = a
n+ n+ b=1 As the Newton iteration technique does not require the exact solu-
(20) tion of the linear system Eq. (22)1, the Krylov iteration will be termi-
T nated on meeting the following condition:
with han,+b = [h1a, b, h2a, b, h3a, b] n + , [0, 1].
Expression (19) indicates that F v (dn + 1, d) is linear in d . It should ~
J x m + Fnk + 1 + Fnvk+ 1 Fkn+ 1 + Fnvk+ 1 , (31)
also be noticed that F v (dn + 1, d) is globally a self-equilibrium system.
Therefore, by adding the pseudo- viscous force, the global equilibrium where is a small error control factor. A detailed scheme of the JFNK is
of the particle system is not changed. presented in Appendix A.

3
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

3.4. Elimination of tension instability =

1
[( n + 1 2 u) + ( n + 1 2 u) T],
2
In conventional Updated Lagrangian formulation of meshless par- 1
= [( n + 1 2 u) ( n + 1 2 u) T].
ticle methods, the support domain of each particle moves with the 2 (39)
particle as a rigid spatial domain. The interactive particles in this spatial Let
domain is updated with the reference configuration through a re-search
1
scheme while the kernel functions are kept the same. Tension instability 1 1
Q= I · I+ .
occurs when the number of interactive particles reduces noticeably as a 2 2 (40)
consequence of finite tensile deformation. To eliminate the tension in- represent the increment rotation associated with the effective spin, the
stability, the update Lagrangian formulation by Chen et al. [21] is stress is then updated according to
adopted, which permits particle support domains deform with the
continuum body such that the initial particle interactions are kept. For n+1 = Q · n·QT + D : D : m. (41)
the deformed support domain, the kernels need to be updated with the
Herein is determined by an iterative scheme to enforce the
reference configuration. In the context of the FPM formulation, this is
consistency condition.
realized by updating the auxiliary kernels according to the following
updating rules:
4. Numerical examples
3
h0,a,nb+ 1 = bh a, b ;
0, n hna,+b1 = ab hna, b ( F an+ 1 2) 1. In this section, three numerical examples are presented to demon-
(32)
strate the validity and applicability of the formulated FPM method in
k=1

where Fna + 1 2 = I + n + 1 2 ( ua) is the increment deformation gradient solving static deformation and plastic flow problems. In these examples,
associated to particle a, and the materials are modelled either as elastic or elastic-perfectly-plastic.
For the elastic-perfectly-plastic model, the Drucker-Prager failure cri-
b Vnb ab Vnb xnbaT a ba
+1 Fn +1 2x n terion (which is matched to the Mohr-Coulomb criterion for plane
= , = .
Vnb+ 1 b
Vn + 1 baT ba
xn + 1xn (33) strain problems) is used as the yield condition with a non-associated
flow rule. The constitutive model contains 5 parameters, namely, the
Young’s modulus E, the Poisson’s ratio , the Mohr-Coulomb strength
parameters c, , and a dilation angle . The values of these parameters
3.5. Constitutive integration
used in the following numerical examples are presented in Table 1.
For each increment time step, the update of stress and other state
variables according to a specific constitutive relation is to be im- 4.1. Example 1: Simply supported elastic beam subjected to a concentrated
plemented. As an illustration, the hypoelastic–plastic type model is force
briefly discussed here.
The constitutive relation considered is in the following form: The first numerical example is designed to show the effectiveness of
the artificial viscosity in preventing spurious deformation mode coming
°
= D: ( p).
(34) into solutions. For this purpose, bending of a simply supported elastic
beam subjected to concentrated force is modelled. The geometry of the
Here the objective rate of the Cauchy stress is related to the elastic beam and the load applying on it are presented in Fig. 1, where x
part of the deformation rate e = p
; D is an isotropic fourth-order represents the particle interval and ls the smoothing length. Numerical
tensor containing 2 constant parameters. The model involves a yield results and the analytical solutions are compared in Figs. 2 and 3. It can
condition be seen that without applying the artificial viscosity ( = 0 ), the pre-
dicted displacement and stress exhibit a strong oscillation. With the
f( , ) 0. (35) artificial viscosity ( = 1.0 ) applied, the numerical results match very
well with the corresponding analytical solutions.
and a flow rule and a hardening law:
p = m ( , ); = r ( , ). (36) 4.2. Example 2: Shear failure in plane strain uniaxial compression test

Here represents a set of hardening variables. During plastic In this example, a rectangular elastic-perfectly-plastic material
loading, the stress remains on the yield surface, which is stated by the sample with a size of 0.1 m× 0.2 m in plane strain uniaxial compression
consistency condition: test is simulated (Fig. 4). Smooth surface conditions are assumed for the
top and the bottom platens. The vertical displacement-controlled test is
0, f 0, f = 0. (37)
proceeded up to a mean axial strain (top displacement/initial height) of
On a typical incremental loading step t [tn, tn+ 1] with the known 15%. Main results are presented in Fig. 4 and in Fig. 5. For comparison,
state variables n and n determined at t = tn , for a given particle dis- the results from FEM simulation using Abaqus software are presented as
placement ua from the he global equilibrium iteration, stress is updated well. A good agreement is observed for the FPM and FEM predictions of
using an incrementally objective integration scheme in the sense of the mean axial stress (reactive force/initial sample width) – mean strain
Hughes and Winget [35], which is described as follows.
The effective incremental displacement gradient is calculated at Table 1
Material and numerical parameters for numerical examples.
mid-point as
Parameters
ns
( n+ 1 2 u)ij = hja,,nb+ 1 2 ui Vnb+ 1 2, (°) (°)
E (MPa) (kg m3) c (kPa)
b=1 (38)
Example 1 100 0.25 5000 – – –
where the subscript “n + 1/2” refers to coordinates
Example 2 5 0.4 1850 15 20 0
xn + 1 2 = xn + u 2 . Following that, the effective rate of deformation Example 3 50 0.4 1850 2.5 20 10
and spin are calculated by

4
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

Fig. 1. Simply supported elastic beam subjected to a concentrated load.

45

40

35

30

Mean Stress (kPa)

25

20

15

10
FEM
5 FPM

0
0.0 0.02 0.04 0.06 0.08 0.10
Mean Strain
Fig. 2. The deflection curve of the beam. Fig. 4. The mean stress-mean strain relations in uniaxial compression test.

4.3. Example 3: Bearing capacity of strip footing

As the third example, the FPM method is applied to find the bearing
capacity of a strip footing of soil foundation. The soil is idealized as an
elastic-perfectly plastic material with parameters listed in Table 1. The
soil foundation is model as a symmetric plane strain problem, with a
rigid plate sized of B × h = 0.22 m× 0.05 m (half-width by thickness)
setting on it. FPM simulation is carried out by assigning a vertical
settlement to the rigid plate. In order to test the capability of the nu-
merical method in simulating large distortion, the settlement d was
purposely set to 15%B. The FPM solution procedure did not complete
due to unconvergence occurred as the settlement was proceeded to
8.7%B ,which corresponds to a moderately large local distortion next to
the corner of the rigid plate. To promote the solution, the FPM method
is combined with a local re-search of the interactive particles in the
plastic zone. The local research was activated for a tentative value of
2
the equivalent plastic strain p = p p
3 ij ij
0.15. With this special
treatment the solution procedure has been completed without diffi-
culties.
For comparison, the FEM analysis using Abaqus and the upper and
lower bound (UB&LB) limit analyses [37–39] using OptumG2
Fig. 3. Maximum axial stress on beam cross sections.
(Op + umG2) have also been conducted. In the limit analyses, the same
x

strength parameters are utilized.

curve. It may be noted that in this example of finite deformation, ma- Unconvergence also occurred to the FEM solution procedure at a
terial distortion is still limited. Therefore, FEM simulation can also be settle about 7.6%B . The FPM and FEM predictions of the mean pressure
carried out without difficulty [36]. (reactive force/B)-settlement (normalized as d B ) curves are presented
in Fig. 6. It can be seen that the turning point on either of the pressure-
settlement curve, which signify the predicted loading capacity, coin-
cides very well with the other. They fall reasonably between the lower

5
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

0.4e-3 1.2e-3 0.06 0.17 0.29 0.88 0.43 1.28

0.00 0.8e-3 1.6e-3 0.00 0.12 0.23 0.00 0.59 1.17 0.00 0.85 1.70

(a) (b) (c) (d)

0.2e-3 0.5e-3 0.05 0.13 0.23 0.67 0.30 0.91

0.00 0.4e-3 0.7e-3 0.00 0.09 0.17 0.00 0.45 0.89 0.00 0.61 1.21

(e) (f) (g) (h)

Fig. 5. Contour of equivalent plastic strain in four representative instants: FPM, at mean axial strain of (a) 0.7% (b) 3% (c) 10% (d) 15%. FEM, at mean axial strain of
(e) 0.7% (f) 3% (g) 10% (h) 15%

and the upper bounds of the limit analyses. For continuing application 70
of the plate settlement, the resistance increases as the soil block next to
the plate heaves up (Fig. 7). FEM predicts a higher soil resistance than
the FPM. Referring to Fig. 7(a), we note that application of the local re- 60
search for the particle interaction in the FPM procedure did not induce
tension instability. The section of the predicted mean pressure-settle-
50
ment curves by FPM with and without local re-search coincides very
Pressure (F/B:kPa)

well.
40
5. Concluding remarks
30
The FPM method for solving static deformation and plastic flow OptumG2-LB
problems has been formulated. The method uses an updated Lagrangian OptumG2-UB
formulation combined with a Lagrangian kernel, which has been de- 20
FEM
monstrated in our previous work being free from tension instability FPM with re-search
problem. For the static problems, the system of governing equations, 10 FPM
which is basically nonlinear in displacements, is integrated implicitly
with respect to a fictitious time. To prevent the hourglass phenomenon
in the numerical procedure, an artificial pseudo-viscosity is introduced 0
0.0 0.05 0.10 0.15
into the particle system. Concerning the broader bandwidth of the
Jacobian in the FPM method, the JFNK method is employed for efficient Displacement/half width (d/B)
solution of the nonlinear system of equations. The details of the Fig. 6. The mean pressure-settlement curves obtained from numerical analysis.

6
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

0.00 0.24 0.48 0.71 0.95 0.00 0.11 0.22 0.32 0.43
(a) (b)

0.00 0.10 0.20 0.30 0.40

(c)
Fig. 7. Contour of equivalent plastic strain fat: (a) d B = 15%, FPM with local re-search; (b) d B = 8.7% , FPM without local re-search; (c) d B = 7.6% , FEM.

numerical algorithms for the JFNK method are presented as well. The Investigation, Writing - original draft, Visualization. Wenxiong Huang:
provided numerical examples have demonstrated the validity of the Conceptualization, Methodology, Investigation, Resources, Writing -
method. review & editing, Supervision, Project administration, Funding acqui-
To summarize, this study has revealed that sition. Andrei Lyamin: Software, Investigation, Resources, Writing -
review & editing, Supervision.
(i) The pseudo-viscosity introduced into the particle system is effec-
tive in preventing the spurious deformation appearing in the so-
lutions of the static problems; Declaration of Competing Interest
(ii) The JFNK method is numerically efficient in solving nonlinear
system of FPM equations. The authors declare that they have no known competing financial
(iii) The FPM described in this paper (without local re-search) is sui- interests or personal relationships that could have appeared to influ-
table only for finite deformation problems with moderate distor- ence the work reported in this paper.
tion. Combined with local re-search of the particle interactions in
the plastic zone, the method is capable to capture deformation
problems with large distortion. Further research on this aspect will Acknowledgement
be carried out.
This work is financially supported by the China National Natural
CRediT authorship contribution statement Science Foundation (No. 11772117/A020311), which is gratefully ac-
knowledged.
Ding Chen: Conceptualization, Methodology, Software, Validation,

Appendix A

In this work, the system of nonlinear Eq. (21) is solved for unknown d using the JFNK method. The details of the algorithm are presented below
(see Tables A1 and A2).

7
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

Table A1
Algorithm of JFNK.
Initialization:
d0: =0; dn0 + 1: =dn + d0 ;
DO (while dk dnk + 1|| u and F kn+ 1 + Fnvk+ 1|| F 0n+ 1|| F)
Set d(k ) :
=0
Step 1:
Apply Krylov iteration described in Table A2 to obtain x m .
dk : =xm
Step 2:
Set : =1.0
DO (while F kn++11 + Fnvk+ 1+ 1|| > F nk + 1 + Fnvk+ 1||)
dk + 1: = dk + dk , dnk+ 1
+ 1: =dn + d
k+1

: =0.8
END
k=k+1
END
d: = dk + 1

Table A2
Krylov iteration with GMRES.
Initialization
SET : =0.01, x 0: =0
Step1:
~
r0: = J x 0 Fnk + 1 Fnvk+ 1
Set z1: = r0 r0 , m: =1; mmax : =20 *
Step2:
~
Calculatewm: = J (dnk + 1) zm
Step3:
FOR l = 1, m
T
hlm : =wm zl ; wm : =wm hlm zl
END
T 1 2
hm + 1, m: =(wm wm)
Step4,
zm + 1: =wm wm
IFm < mmax
m=m+1
Go to Step2
END
Step5:

Solve the minimax y (m) = arg (min r0 e1 Hm y (QR factorization is used

y
[24])
x m: =x 0 + Zm y (m)
Step6:
~
IF J x m + Fnk + 1 + Fnvk+ 1 Fnk + 1 + Fnvk+ 1
Setx 0 = x m
Repeat Step1 to Step6.
END

* Note that, the maximum iteration at a time mmax may set to a value between 20 and 1000, depending on
the problem to be solved.

References free surface flows of viscous fluids. Appl Numer Math 2009;59(2):251–71.
[8] Huang C, Zhang DH, Shi YX, Si YL, Huang B. Coupled finite particle method with a
modified particle shifting technology. Int J Numer Meth Eng 2018;113(2):179–207.
[1] Lucy LB. A numerical approach to the testing of the fission hypothesis. Astronom J [9] He F, Zhang H, Huang C, Liu M. Numerical investigation of the solitary wave
1977;82:1013. breaking over a slope by using the finite particle method. Coast Eng 2020;156.
[2] Monaghan JJ. Simulating free surface flows with SPH. J Comput Phys https://doi.org/10.1016/j.coastaleng.2019.103617.
1994;110(2):399–406. [10] Cleary PW, Prakash M, Das R, Ha J. Modelling of metal forging using SPH. Appl
[3] Chen JK, Beraun JE, Carney TC. A corrective smoothed particle method for Math Model 2012;36(8):3836–55.
boundary value problems in heat conduction. Int J Numer Meth Eng [11] Bui HH, Fukagawa R. An improved SPH method for saturated soils and its appli-
1999;46(2):231–52. cation to investigate the mechanisms of embankment failure: Case of hydrostatic
[4] Bonet J, Lok T-SL. Variational and momentum preservation aspects of Smooth pore-water pressure. Int J Numer Anal Meth Geomech 2013;37(1):31–50.
Particle Hydrodynamic formulations. Comput Methods Appl Mech Eng [12] Peng C, Xu G, Wu W, Yu H-S, Wang C. Multiphase SPH modeling of free surface flow
1999;180(1–2):97–115. in porous media with variable porosity. Comput Geotech 2017;81:239–48.
[5] Liu MB, Liu GR. Restoring particle consistency in smoothed particle hydrodynamics. [13] Adami S, Hu XY, Adams NA. A generalized wall boundary condition for smoothed
Appl Numer Math 2006;56(1):19–36. particle hydrodynamics. J Comput Phys 2012;231(21):7057–75.
[6] Huang C, Lei JM, Liu MB, Peng XY. A kernel gradient free (KGF) SPH method. Int J [14] Huang Y, Cheng H, Dai Z, Xu Q, Liu F, Sawada K, et al. SPH-based numerical si-
Numer Meth Fluids 2015;78:691–707. mulation of catastrophic debris flows after the 2008 Wenchuan earthquake. Bull
[7] Fang J, Parriaux A, Rentschler M, Ancey C. Improved SPH methods for simulating Eng Geol Environ 2015;74(4):1137–51.

8
D. Chen, et al. Computers and Geotechnics 122 (2020) 103502

[15] Liu M, Zhang Z. Smoothed particle hydrodynamics (SPH) for modeling fluid- linear systems. J Comput Appl Math 2015;283:71–8.
structure interactions. Sci China Phys Mech Astron 2019;62(8):984701. [29] Simo JC, Ortiz M. A unified approach to finite deformation elastoplastic analysis
[16] Ye T, Pan D, Huang C, Liu M. Smoothed particle hydrodynamics (SPH) for complex based on the use of hyperelastic constitutive equations. Comput Meth Appl Mech
fluid flows: Recent developments in methodology and applications. Phys Fluids Eng 1985;49(2):221–45.
2019;31(1):011301. [30] Liu MB, Xie WP, Liu GR. Modeling incompressible flows using a finite particle
[17] Swegle JW, Hicks DL, Attaway SW. Smoothed particle hydrodynamics stability method. Appl Math Model 2005;29(12):1252–70.
analysis. J Comput Phys 1995;116(1):123–34. [31] Bui HH, Fukagawa R, Sako K, Ohno S. Lagrangian meshfree particles method (SPH)
[18] Belytschko T, Guo Y, Liu WK, Xiao SP. A unified stability analysis of meshless for large deformation and failure flows of geomaterial using elastic-plastic soil
particle methods. Int J Numer Meth Eng 2000;48(9):1359–400. constitutive model. Int J Numer Anal Meth Geomech 2008;32(12):1537–70.
[19] Monaghan J. SPH without a Tensile Instability. J Comput Phys [32] Lee CH, Gil AJ, Greto G, Kulasegaram S, Bonet J. A new Jameson–Schmidt–Turkel
2000;159(2):290–311. Smooth Particle Hydrodynamics algorithm for large strain explicit fast dynamics.
[20] Bonet J, Kulasegaram S. Remarks on tension instability of Eulerian and Lagrangian Comput Meth Appl Mech Eng 2016;311:71–111.
corrected smooth particle hydrodynamics (CSPH) methods. Int J Numer Meth Eng [33] Ganzenmüller GC, Sauer M, May M, Hiermaier S. Hourglass control for Smooth
2001;52(11):1203–20. Particle Hydrodynamics removes tensile and rank-deficiency instabilities: Hourglass
[21] Chen D, Huang WX, Sloan SW. An alternative updated Lagrangian formulation for control for SPH. Eur Phys J Spec Top 2016;225(2):385–95.
finite particle method. Comput Methods Appl Mech Eng 2019;343:490–505. [34] Dyka CT, Randles PW, Ingel RP. Stress points for tension instability in SPH. Int J
[22] Vidal Y, Bonet J, Huerta A. Stabilized updated Lagrangian corrected SPH for explicit Numer Meth Eng 1997;40(13):2325–41.
dynamic problems. Int J Numer Meth Eng 2007;69(13):2687–710. [35] Hughes TJR, Winget J. Finite rotation effects in numerical integration of rate
[23] Wang J, Chan D. Frictional contact algorithms in SPH for the simulation of soil- constitutive equations arising in large-deformation analysis. Int J Numer Meth Eng
structure interaction. Int J Numer Anal Meth Geomech 2014;38(7):747–70. 1980;15(12):1862–7.
[24] Knoll DA, Keyes DE. Jacobian-free Newton-Krylov methods: A survey of approaches [36] Guo N, Zhao J. Multiscale insights into classical geomechanics problems. Int J
and applications. J Comput Phys 2004;193(2):357–97. Numer Anal Meth Geomech 2016;40(3):367–90.
[25] Zhou D, Lu Z, Guo T. Improvement of computational efficiency of circular function- [37] Lyamin AV, Sloan SW. Lower bound limit analysis using non-linear programming.
based gas kinetic scheme by using Jacobian-free Newton-Krylov method. Comput Int J Numer Meth Eng 2002;55(5):573–611.
Fluids 2018;161:121–35. [38] Lyamin AV, Sloan SW. Upper bound limit analysis using linear finite elements and
[26] Saad Y, Schultz MH. GMRES: a generalized minimal residual algorithm for solving non-linear programming. Int J Numer Anal Meth Geomech 2002;26(2):181–216.
nonsymmetric linear systems. SIAM J Sci Stat Comput 1986;7(3):856–69. [39] Krabbenhoft K, Karim MR, Lyamin AV, Sloan SW. Associated computational plas-
[27] Saad Y. A flexible inner-outer preconditioned GMRES algorithm. SIAM J Sci ticity schemes for nonassociated frictional materials. Int J Numer Meth Eng
Comput 1993;14(2):461–9. 2012;90(9):1089–117.
[28] Bai ZZ. Motivations and realizations of Krylov subspace methods for large sparse