10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023].

See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
Received: 12 October 2020 Revised: 25 November 2022 Accepted: 6 December 2022

DOI: 10.1002/num.22997

RESEARCH ARTICLE

A novel local Hermite radial basis function-based

differential quadrature method for solving
two-dimensional variable-order time fractional
advection–diffusion equation with Neumann
boundary conditions

Jianming Liu1 Xin Kai Li2 Xiuling Hu1

1 School of Mathematics and Statistics, Jiangsu

Normal University, Xuzhou, China Abstract

2 Faculty of Technology, De Montfort University,
A novel Hermite radial basis function-based differential
Leicester, UK
quadrature (H-RBF-DQ) method is presented in this
Correspondence
Xin Kai Li, Faculty of Technology, De Montfort paper based on 2D variable order time fractional
University, Leicester LE1 9BH, UK. advection–diffusion equations with Neumann boundary
Email: [email protected] conditions. The proposed method is designed to treat
Funding information accurately for derivative boundary conditions, which con-
National Numerical Wind Tunnel Project,
Grant/Award Number: NNW2018-ZT1A02;
siderably improve the approximation results and extend
Natural Science Foundation of the Jiangsu Higher the range of applicability for the method of RBF-DQ.
Education Institutions of China, Grant/Award The advantage of the present method is that the Hermite
Number: 18KJA110001.
interpolation coefficients are only dependent of the point
distribution yielding a substantially better imposition
of boundary conditions, even for time evolution. The
proposed algorithm is thoroughly validated and is demon-
strated to handle the fractional calculus problems with
both Dirichlet and Neumann boundaries very well.

KEYWORDS

differential quadrature method, Hermite RBF-DQ,

meshfree method, Neumann boundary condition,
variable-order time fractional

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in
any medium, provided the original work is properly cited.
© 2023 The Authors. Numerical Methods for Partial Differential Equations published by Wiley Periodicals LLC

Numer Methods Partial Differential Eq. 2023;1–22. wileyonlinelibrary.com/journal/num 1

 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
2 LIU ET AL .

1 INTRODUCTION

Nowadays, the field of fractional calculus has attracted interest of researchers in several areas
including mathematics, physics, chemistry, engineering, and even finance and social sciences. Main
reason for this development is that a fractional model could describe simply and elegantly the com-
plex characteristics of physical models. Therefore, the classical integer order models, for example,
Maxwell model and Oldroyd-B model in non-Newtonian fluids, have been updated to the fractional
order ones for complex viscoelastic materials [1–3]. Due to the growing number of applications of
fractional calculus in science and engineering, numerical methods are being developed to provide tools
for solving such problems [4]. Nevertheless, the intrinsic complexity of fractional calculus, caused
partially by non-local properties of fractional derivatives and integrals, makes it rather difficult to find
efficient numerical methods in this field. Despite this fact, however, the literature exhibits a growing
interest and improving achievements in numerical methods for fractional calculus in general [5–14].
It seems enough to mention here that currently the numerical methods commonly used in fractional
calculus included finite difference method and finite element method. The finite element method has
now become the most widely used numerical method in research and application, and has played a
great role in scientific research and engineering analysis. The finite element method discretized the
complex region into elements, and a node based interpolation function approximation was established
on the element for each variable. The governing equation was established by the variational principle
or the weighted residual method. It effectively overcame the limitation of the finite difference method
for the complex geometrical domain, and can handle the different boundaries flexibly. However, when
the finite element method was used to deal with the problems with large deformation, dynamic crack
propagation, and fluid solid interaction, the grids can be distorted. Therefore, mesh reconstruction for
large deformation was unavoidable during the process of numerical simulation, which not only limited
the computational efficiency, but also led to the damage of the numerical precision. In order to get
rid of the solution dependence on the grid, the meshless method, only based on nodes, has made great
progress since 1990s [15–18]. The meshless method with the structure of the test function is built on
a series of discrete nodes. The connection between the field point and the node is no longer realized
by the element. Hence, it gets rid of the constraints of the grid or element, and no longer needs the
process of the grid reconfiguration when it involves the grid distortion and the grid movement.
In order to simplify the mesh generation for the complex computational domains, the mesh-
less/meshfree methods attracted more attention in recent years. With the effort of scientists, many kinds
of effective meshless methods appeared, including smooth particle hydrodynamics method (SPH),
element-free Galerkin method (EFG), meshless local Petrov–Galerkin method (MLPG), reproducing
kernel particle method (RKPM), radial bases functions method (RBFs), finite point method (FPM),
moving least square method (MLS), and so on [16–18]. Their main difference lies in the use of different
test functions or different equivalent forms of differential equations. Generally, at the present, mesh-
less methods can be divided into two categories, that is, weak-form method and collocation method
[16]. In the former, the meshless method gives the discrete form of the problem based on the weak
form controlled by variational principle of partial differential equations (PDEs). The characteristic of
this method is that the solution is of high accuracy and good stability, but the amount of calculation is
high, and the numerical integration is demanded. In the later, the collocation method directly satisfies
the differential equation or boundary condition at discrete points. In addition, the collocation method
is simple and does not require numerical integration. Furthermore, the computational efficiency is
higher than that of the weak form method.
In 1990, Kansa first applied the collocation method and RBFs method with application to compu-
tational fluid dynamics [19, 20]. After that, there are lots of research works on the applications of the
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 3

RBFs collocation methods (see [11, 21–31] and their references). The collocation method with RBFs
is a real meshless method for solving PDEs, and it is independent of the spatial dimensions [17, 32].
As we all know, for stationary grid, the coefficients of finite difference scheme for spatial derivative
is always fixed. But the common collocation method with RBFs for temporal problem, the interpola-
tion coefficients should be calculated in every step during the time advancing process. To enhance the
computational efficiency and treat complex geometries, Shu et al. developed a meshfree local radial
basis function-based differential quadrature (RBF-DQ) approach to solve the two-dimensional incom-
pressible Navier–Stokes equations [21]. RBF-DQ method originated from the concept of differential
quadrature (DQ) [33, 34], however, it cannot directly be applied to the problems with irregular geome-
tries. Due to the vast flexibility, RBF-DQ method has been successfully used to study many scientific
and engineering problems with irregular geometries [22, 32, 35–41]. In this method, the interpola-
tion coefficients are only dependent on the distributions of the points and independent of the solution.
However, the accuracy achieved by direct meshless collocation scheme is a bit poor especially on the
boundaries [42, 43]. Furthermore, the collocation scheme has difficulties in dealing with Neumann
boundary conditions. In order to improve the accuracy on the derivative boundaries, recently, Hermite
RBFs method has been developed by many authors [16, 17, 43–45]. It should be noted that, for the most
of the Hermite RBFs methods, the interpolation coefficients depend on the solutions. In this paper, we
developed a novel Hermite RBF-DQ (H-RBF-DQ) method for solving PDEs based on a variable-order
time fractional advection–diffusion equation with both Dirichlet and Neumann boundary conditions.
The advantage of the present method is that the interpolation coefficients are only dependent of the
point distributions, therefore, it can greatly reduce the cost of computation. The present method can be
seen as an extension of local RBF-DQ meshless approach for Neumann boundary problems in [46], but
the accuracy for approximating to the Neumann boundary conditions has been much improved. With-
out loss of generality, in this paper, the method is validated by a special variable-order time fractional
advection–diffusion equation with Neumann boundary conditions.
There are some researches on meshless method for the numerical simulations of time fractional
advection–diffusion equations. For example, MLS method is developed to solve the constant-order
time fractional advection–diffusion equation in the papers [47, 48]. For the variable-order time frac-
tional advection–diffusion equation, it can be referred to the paper [49]. However, the shape function
in MLS method is not satisfied with Kronecker 𝛿 function character and the boundary conditions
cannot be directly enforced, which increase the difficulty for the treatment of boundaries. Further-
more, the above works were all on the problem with Dirichlet boundary conditions. In [11], a global
Kansa method is developed to solve a variable-order time-fractional advection–diffusion equation
arising in anomalous transport. And also, in [50], a novel global RBF-based meshless method with
particular boundary treatment for solving time-fractional transport equations is presented. In [46],
a meshless local RBF-DQ approach to solve a variable-order time fractional advection–diffusion
equation subjected to Neumann boundary conditions. In [51–53], boundary collocation and local-
ized collocation methods for anomalous heat conduction and diffusion on surfaces are presented. In
order to improve the approximating accuracy on the Neumann boundary conditions, in this paper, a
novel H-RBF-DQ method is proposed to deal with both Dirichlet and Neumann boundary conditions
based on a two-dimensional variable order time fractional advection diffusion equation. The numerical
results demonstrate the robustness and the versatility of the proposed method.
The paper is organized as follows. In Section 2, the modeling equations and their time discretiza-
tion are presented. The review of RBF-DQ method is presented in Section 3. For the treatment of the
Neumann boundary condition, in Section 3, based on the Hermite RBF interpolation, a novel
H-RBF-DQ method is proposed to improve the approximating accuracy on the Neumann boundaries.
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
4 LIU ET AL .

Numerical results with applications are presented in Section 4 to demonstrate the accuracy of the
proposed method. Finally, the conclusion is presented in Section 5.

2 THE M ODELING EQUATION AND TIME DISCRETIZATION

APPROXIMATION

The variable-order time fractional advection–diffusion equation can be formulated as [49]

c 𝛼(x,t)
0 Dt u(x, t) = 𝜅(x, t)Δu(x, t) − v(x, t) ⋅ ∇u(x, t) + f (x, t), x = (x, y) ∈ Ω ⊂ R2 , t > 0, (1)
subject to the following initial and boundary conditions
u(x, 0) = g(x), x ∈ Ω,
𝜕u(x, t)
= h(x, t), x ∈ 𝜕𝛺, t > 0, (2)
𝜕n
( )
where Δu = 𝜕𝜕xu2 + 𝜕𝜕yu2 , and ∇u = 𝜕u 𝜕u
2 2
,
𝜕x 𝜕y
. Ω is a bounded domain in R2 , 𝜕Ω is the boundary of Ω,
𝜅(x, t) > 0 is the diffusion coefficient function, v(x,t) is advection velocity, and f (x, t), g(x), h(x, t) are
given functions.
In Equation (1), the variable-order time fractional derivative c0 Dt𝛼(x,t) u(x, t) with 0 < 𝛼(x, t) ≤ 1
follows Caputo definition, which is given by
{
∫0 (t−𝜉)1𝛼(x,t)) 𝜕u(x,𝜉)
1 t
c 𝛼(x,t) Γ(1−𝛼(x,t)) 𝜕𝜉
d𝜉, 0 < 𝛼(x, t) < 1,
0 Dt u(x, t) = 𝜕u(x,t) (3)
𝜕t
, 𝛼(x, t) = 1,
where Γ is the Gamma function.
In this study, we focus on the Hermite RBF-DQ method for the variable-order time frac-
tional advection–diffusion equation on complex geometries with Neumann boundary conditions.
First, in this section, we give the time discretization approximation for Equation (1). Using
the time discretization method proposed in [49] for Caputo derivative with the notation
bj (x, tk+1 ) = (j + 1)1−𝛼(x,tk+1) − j1−𝛼(x,tk+1) , we have

1 ∑ k tj+1
1 𝜕u(x, ξ)
c α(x,t)
0 Dt u1 (x, tk+1 ) = 𝑑ξ,
Γ (1 − α (x, tk+1 )) j=0 ∫tj (tk+1 − ξ)α(x,tk+1 ) 𝜕𝜉
( ) ( ) t
1 ∑k
u x, tj+1 − u x, tj j+1
1
= 𝑑ξ + Rk+1 ,
Γ (1 − α (x, tk+1 )) j=0 Δt ∫tj (tk+1 − ξ)α(x,tk+1 )

Δt−α(x,tk+1 ) ∑
k
( ( ) ( ))
= bk−j (x, tk+1 ) u x, tj+1 − u x, tj + Rk+1
Γ (2 − α (x, tk+1 )) j=0

Δt−α(x,tk+1 ) ∑
k
( ( ) ( ))
= bj (x, tk+1 ) u x, tk−j+1 − u x, tk−j + Rk+1 (4)
Γ (2 − α (x, tk+1 )) j=0

where tk = kΔt for k = 0, 1, … , M and M = T/Δt. The truncation error Rk+1 is subjected to [47, 49, 54]

|Rk+1 | ≤ CΔt2−𝛼(x,tk+1 ) . (5)

Substituting Equation (4) into Equation (1), we can obtain
u (x, tk+1 ) − 𝜇 (x, tk+1 ) [𝜅 (xtk+1 ) Δu(x, tk+1 ) − v(x, tk+1 ) ⋅ ∇u(x, tk+1 )]
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 5

∑
k
( ( ) ( ))
= u (x, tk ) − bj (x, tk+1 ) u x, tk−j+1 − u x, tk−j + 𝜇 (x, tk+1 ) f (x, tk+1 ) + Rk+1 , (6)
j=1

where
𝜇 (x, tk+1 ) = Δt𝛼(x,tk+1 ) Γ (2 − 𝛼 (x, tk+1 )) , (7)
and
|Rk+1 | ≤ CΔt2 , (8)
where C and C are constant.
Using the notations uk = uk (x) as the numerical approximation to u(x, tk ), 𝜇k = 𝜇(x, tk ), 𝜅 k =
𝜅 (x, tk ) , vk = v (x, tk ) = (v1 (xtk ) , v2 (xtk )) , bkj = bj (x, tk ) and f k = f (x, tk ), then Equation (1) can be
discretized as follows
[ ]
uk+1 − 𝜇k+1 𝜅 k+1 Δuk+1 − vk+1 ⋅ ∇uk+1
∑
k
( k−j+1 )
= uk − bk+1
j u − uk−j + 𝜇k+1 f k+1 . (9)
j=1

Equation (9) can be rearranged as

[ ]
uk+1 − 𝜇k+1 𝜅 k+1 Δuk+1 − vk+1 ⋅ ∇uk+1
( ) k∑
k−1
( k+1 ) k−j
k u +𝜇 f .
k+1
= 1 − b1 u bj − bk+1
j+1 u + bk+1 0 k+1 k+1
(10)
j=1

3 SPATIAL DISCRETIZATION METHOD

In order to obtain a final numerical scheme of Equation (10), the spatial discretization of variable-order
time fractional advection–diffusion equation should be given. In this section, we first review
RBF-based differential quadrature method for the spatial discretization, which has been reported in
our recent paper [46]. Then the details of the new Hermite RBF-based differential quadrature method
will be proposed to improve the accuracy of approximating solutions near the Neumann boundary.

3.1 Review of RBF-based differential quadrature method

The idea of differential quadrature method came from the numerical integral, that any integral over
a closed domain can be approximated by a linear weighted sum of all the functional values in the
integral domain. In differential quadrature method, similar to other numerical integral, the derivative
value u(m) (x) at the center xi are approximated by a linear weighted sum of the function values at a set
of nodes in a closed domain as
∑
N
( ) ∑
N
( )
u(m)
x (xi ) ≈ w(m)
ij u xj , u(m)
y (xi ) ≈ w(m)
ij u xj , (11)
j=1 j=1

for i = 1, · · ·, N, where N is the total number of the points inside the computational domain, and
w(m) (m)
ij , wij are the weighting coefficients for derivatives of order m with respect to x and y, respectively.
The weighting coefficients of w(m) (m)
ij , and wij are determined by the base functions as the test func-
tion in Equation (11), which will greatly enhance the computational efficiency [21, 55]. Although,
there are many different kind of RBFs available, due to the excellent performance for the scattered data
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
6 LIU ET AL .

interpolation, the multiquadrics (MQ) basis function is used here as the test function [21, 22]. In the
RBF-DQ method, the function in the region of Ω can be locally approximated by MQ RBFs as

∑
N √( )2 ( )2
h(x, y) = 𝜆j x − xj + y − yj + c2j + 𝜆N+1 , (12)
j=1

with shape parameter cj . To make the problem be well-posed, the equation

∑
N
𝜆j = 0, (13)
j=1

is enforced. Substituting Equation (13) into Equation (12) gives

∑
N
h(x, y) = 𝜆j gj (x, y) + 𝜆N+1 , (14)
j=1,j≠i

where
√( )2 ( )2 √
gj (x, y) = x − xj + y − yj + c2j − (x − xi )2 + (y − yi )2 + c2i . (15)

The number of unknowns in Equation (14) is N. As the setting in the paper [21], 𝜆N+1 can be
replaced by 𝜆i , and Equation (14) can be rewritten as

∑
N
h(x, y) = 𝜆j gj (x, y) + 𝜆i , (16)
j=1,j≠i

where gi (x, y) = 1 and gj (x, y), j = 1, · · ·, N but j ≠ i given by Equation (15), formed a base vector to
cover the function space of h(x, y).
In RBF-DQ method, the most important contribution is that the weighting coefficients of w(m) ij , and
(m)
wij are determined by all the base functions as the test function in Equation (11). Based on this idea,
we can obtain
∑
N
w(m)
ik = 0, (17a)
k=1

𝜕 m gj (xi , yi ) ∑ (m)
N
= wik gj (xk , yk ) , j = 1, 2, · · · , N, but j ≠ i. (17b)
𝜕xm k=1

The N equations of (17) formed a linear system with N unknowns for the given i, and the weighting
coefficients w(m)
ik can be solved by a numerical method. In a similar manner, the weighting coefficients
w(m)
ij of the y-derivatives can also be computed by Equation (17) with x substituted by y.
Substitution of Equation (11) into time discretized Equation (10) at point xi = (xi , yi ), with uk+1
i as
the approximation solution of u(xi , tk+1 ), yields
[ N (
]
∑ ) N (
∑ )
uk+1
i − 𝜇ik+1 𝜅ik+1 w(2)
ij + w(2)
ij u k+1
j − w(1) k+1
v
ij 1,i + w(1) k+1
v
ij 2,i uk+1
j
j=1 j=1

( ) k ∑ ( k+1 k−1
) k−j
= 1 − bk+1 bi,j − bk+1 i,k ui + 𝜇i fi
+ bk+1 .
0 k+1 k+1
i,1 ui + i,j+1 ui (18)
j=1
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 7

In Equation (18), the following notations are used

𝜇ik+1 = 𝜇 (xi , tk+1 ) , 1,i = v1 (xi , tk+1 ) ,

vk+1 2,i = v2 (xi , tk+1 ) ,
vk+1

𝜅ik+1 = 𝜅 (xi , tk+1 ) , i,j = bj (xi , tk+1 ) ,

bk+1 fik+1 = f (xi , tk+1 ) .
It should be noted that when the number of nodes, N, is large, the coefficient matrix of (17) may
be ill-conditioned. This limits its application. Hence, we use the local RBF-DQ method to solve the
variable-order time fractional advection–diffusion equation.
The key of local RBF-based differential quadrature method is that the m-order derivatives of u(x)
at xi are approximated by the function values at a set of nodes in the neighborhood of xi with N i
supporting nodes (including xi ). That is
∑
Ni
( ) ∑
Ni
( )
u(m)
x (xi ) ≈ w(m)
ij u xj , u(m)
y (xi ) ≈ w(m)
ij u xj , i = 1, 2, · · · , N. (19)
j=1 j=1

The corresponding coefficients w(m) (m)

ij and wij can be determined by Equation (17) with N i local
support nodes in the neighbor of xi .
Substitution of Equation (19) into time discretized Equation (10) at point xi , by the local RBF-based
differential quadrature method, yields
[ Ni (
]
∑ ) Ni (
∑ )
(2) (2)
uk+1
i − 𝜇ik+1 𝜅i
k+1 k+1
wij + wij uj − w(1) k+1 (1) k+1
ij v1,i + wij v2,i uk+1
j
j=1 j=1

( ) k ∑ ( k+1 k−1
) k−j
= 1 − bk+1 bi,j − bk+1 i,k ui + 𝜇i fi
+ bk+1 ,
0 k+1 k+1
i,1 ui + i,j+1 ui (20)
j=1

for i = 1, 2, · · ·, N.
Although the discretization of Equation (1) is simple, the RBF-DQ approximation of the function
contains a shape parameter c that could be node-dependent and must be determined by the user. In this
study, we use the method of normalization of supporting region suggested in the paper of [21], which
can also be referred to [46].

3.2 Hermite RBF-DQ method

In order to improve the approximated accuracy near the boundary with Neumann boundary condi-
tion, for the unknown computational nodes with supporting points on the Neumann boundary, we
suggest to use the Hermite method and include the information on the boundaries to obtain the spatial
discretization of Equation (10).
For convenience of illustration, a sample of mesh points distribution in H-RBF-DQ method on
hybrid boundary conditions is shown in Figure 1, where the Neumann boundary conditions are
enforced on the upper boundary and the Dirichlet boundary conditions are satisfied on the lower
boundary. According to the position of the unknown points, we divide them into four categories. Type
A denotes the point on the Neumann boundary conditions. Type B denotes the internal point with
supporting points on the Dirichlet boundary conditions. Type C denotes the point with all internal
supporting points. And Type D denotes the internal point with supporting points on the Neumann
boundary conditions (see Figure 1). For the relationship structure of points, we use the algorithm in
paper [56].
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
8 LIU ET AL .

FIGURE 1 Mesh points distribution in Hermite radial basis function-based differential quadrature (H-RBF-DQ) method

The approximation of a function u(x) (see [16, 43]) can be written in a Hermite interpolation form
by a linear combination of RBFs at all the N i supporting nodes (including the Nib Neumann boundary
points) within the local support domain and the normal derivatives along the normal direction
n = (nx , ny ) at the Neumann boundary points as
b
Ni
∑Ni
∑ 𝜕𝜙b
u(x) ≈ h(x, y) = 𝜆j 𝜙j + 𝛾l l + 𝜆Ni +1 , (21)
j=1 l=1
𝜕nl
( )
where 𝜆j (j = 1, · · · , Ni ) , 𝛾l l = 1, · · · , Nib are interpolation coefficients, and 𝜙j , 𝜙bl are radial basis
function, respectively.
In this study, similar to the function gj (x, y) given in the previous section, the MQ basis function
is used as the radial basis function, that is
√( √
)2 ( )2 ( )2 ( )2
𝜙j = x − xj + y − yj + c2j , 𝜙bl = x − xlb + y − ybl + c2l , (22)
( )
where xbl = xlb , ybl is the coordinate for the lth normal derivative boundary point. So we can get
( ) ( )
𝜕𝜙bl 𝜕𝜙bl 𝜕𝜙bl x − xlb nx + y − ybl ny
= nx + ny = √ . (23)
𝜕nl 𝜕x 𝜕y ( )
b 2
( )
b 2 2
x − xl + y − yl + cl

With the constrains of (13) and the notation of (15), Equation (21) can be reformulated as
b
Ni
∑
Ni
∑ 𝜕𝜙b
u(x) ≈ h(x, y) = 𝜆j gj (x, y) + 𝜆i + 𝛾l l , (24)
j=1,j≠i l=1
𝜕nl

and
Ni b ( )
𝜕u(x) ∑Ni
𝜕gj (x, y) ∑ 𝜕 𝜕𝜙bl
≈ 𝜆j + 𝛾l , (25)
𝜕n j=1,j≠i
𝜕n l=1
𝜕n 𝜕nl

𝜕𝜙bl
where gi (x, y) = 1, gj (x, y), j = 1, · · ·, N i but j ≠ i and 𝜕nl
, l = 1, · · · , Nib are a base vector for the
function space of h(x, y).
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 9

Let
[ ]T
𝜕ub1 𝜕ubN b
̂
u = u 1 · · · u Ni ··· i
, (26)
𝜕n1 𝜕nNib
1×(Ni +Nib )

and the coefficients vector

[ ]T
a = 𝜆1 · · · 𝜆Ni 𝛾1 · · · 𝛾Nib
̂ , (27)
1×(Ni +Nib )

can be obtained by the interpolations at N i points and the normal derivatives at Nib points on the Neu-
mann boundary in Equations (24) and (25). The process can also be expressed by matrix formulation
as follows
̂
u = Ψ̂ a. (28)
Thus, the unknown coefficients vector
a = Ψ−1̂
̂ u. (29)
Finally, the approximation form of function can be obtained like

a = 𝜙Ψ−1̂
u(x) ≈ 𝜙̂ u = 𝜓̂
u, (30)
where [ ]
𝜕𝜙b 𝜕𝜙N b b

𝜙 = g1 (x, y), · · · , gi−1 (x, y), 1, gi+1 (x, y), · · · , gN (x, y), 1 , · · · , i
, (31)
𝜕n1 𝜕nNib

and the matrix of shape functions 𝜓 can be expressed as follows

[ ]
𝜓 = 𝜓1 , 𝜓2 , · · · , 𝜓Ni , 𝜓1H , 𝜓2H , · · · , 𝜓NHb . (32)
i

Hence, the function u(x) can be approximately expressed as

b
Ni
∑
Ni
∑ 𝜕ubj
u(x) ≈ 𝜓k uk + 𝜓jH . (33)
k=1 j=1
𝜕nj

From the interpolation formulation (33), we can directly get

Nb
𝜕 m u(x) ∑ 𝜕 m 𝜓k ∑
Ni i
𝜕 m 𝜓jH 𝜕ubj
≈ u k + . (34)
𝜕xm k=1
𝜕xm j=1
𝜕xm 𝜕nj
𝜕 u(x) m
Then the mth order derivative value u(m)
x (x) =
𝜕xm
at the center xi can be approximated by a linear
weighted sum of the function values at N i points and the normal derivatives at N b points on Neumann
boundary as
b
Ni
∑
Ni
∑ 𝜕ubj
u(m)
x (xi ) ≈ w(m)
ik uk + s(m) . (35)
k=1 j=1
ij
𝜕nj

In order to get the coefficients in Equations (35), by using the idea of differential quadrature in
Section 3, we take all the base functions in (31) as the test functions and solve the following system
∑
Ni
w(m)
ik = 0, (36a)
k=1
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
10 LIU ET AL .

Nib ( b b)
𝜕 m gj (xi , yi ) ∑ ∑
Ni
(m) (m) 𝜕gj xl , yl
= wik gj (xk , yk ) + sil , j = 1, 2, · · · , Ni , but j ≠ i, (36b)
𝜕xm k=1 l=1
𝜕nl

( ) Nb ( )
𝜕m 𝜕𝜙bp ∑
Ni
𝜕𝜙bp (xk , yk ) ∑ i
𝜕 𝜕𝜙bp ( )
(xi , yi ) = w(m) + s(m) xlb , ybl , p = 1, 2, · · · , Nib . (36c)
𝜕xm 𝜕np k=1
ik
𝜕np l=1
il
𝜕n l 𝜕np

In a similar manner, the weighting coefficients w(m) (m)

ik , and sij of the mth order y-derivatives can
also be computed by Equation (36) with x substituted by y.
With the notation of (35) for the spatial derivative, the numerical scheme at the node D (see
Figure 1) with supporting points on the Neumann boundary will become
[ Ni (
]
∑ ) Ni (
∑ )
uk+1
i − 𝜇ik+1 𝜅ik+1 w(2) (2)
ij + wij
k+1
uj − w(1) k+1 (1) k+1
ij v1,i + wij v2,i uk+1
j
j=1 j=1

( ) k ∑ k−1
( k+1 ) k−j
= 1 − bk+1 bi,j − bk+1 i,k ui + 𝜇i fi
+ bk+1 + Fik+1 ,
0 k+1 k+1
i,1 ui + i,j+1 ui (37)
j=1

where the source term Fik+1 denotes the impact on Neumann boundary conditions, defined by

⎧ Nb ⎫
i [ ( ) ( )] 𝜕ub ⎪
⎪∑ (2) (2) (1) k+1 (1) k+1 j
Fik+1 𝜇ik+1 ⎨ 𝜅i
k+1
.
𝜕nj ⎬
= sij + sij − sij v1,i + sij v2,i (38)
⎪ j=1 ⎪
⎩ ⎭

The boundary treatment for Dirichlet boundary condition in collocation method is trivial. However,
the collocation scheme has difficulties in dealing with Neumann boundary conditions. For the present
( )
Hermite-type RBF-DQ method (37), the Neumann boundary condition at point xib , ybi (see the point
A in Figure 1) can be formulated as

( ) Ni ( ) Nib ( ) 𝜕ub
𝜕u xib , ybi ∑ (1) (1)
∑ j
≈ wik nx + wik ny uk + sij(1) nx + s(1)
ij ny . (39)
𝜕n k=1 j=1
𝜕nj

Although, the derivatives in the Neumann condition can also be discretized by the local RBF-DQ
method of (19) as
( ) Ni ( )
𝜕u xib , ybi ∑
≈ w(1) n x + w(1)
n y uk , (40)
𝜕n k=1
ik ik

the accuracy is much lower than the present H-RBF-DQ method, which will be shown in the next
section.
It is worth pointing out that, in the present H-RBF-DQ method, the normal derivatives at the
Neumann boundary nodes are added as additional DOFs, which are enforced by Equation (35). For
an unknown collocation node at xi , if its local support nodes do not include the Neumann boundary
node, the conventional differential quadrature formulas (19) are used to approximate the derivatives
in Equation (10), and Equation (20) is used to discretize the variable-order time fractional
advection–diffusion equation. If its support neighboring nodes include the Neumann boundary nodes,
the formulation (35) is used. Hence, by the present H-RBF-DQ method, if the computational node with
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 11

supporting points on the Neumann boundary, the scheme of (37) is used, otherwise the scheme (20)
is used. The throughout procedure can be referred to the Algorithm 1. Although the above description
is based on the variable-order time fractional advection–diffusion equation, the same treatment is still
suitable for other Neumann boundary-value problems.
We must point out that the present Hermite RBF point interpolation method is different from ones
in [16, 43]. The method in [16, 43] firstly required to obtain the matrix of shape functions 𝜓 in (30);
then the differential of components of 𝜓 must be calculated to obtain Equation (34). Furthermore, for
time related problems, their method demands that the calculation of the matrix of shape functions 𝜓
with their differential of components should be solved at every time step. In our method, however,
the coefficients in (35) is only related to the support points set; and for time related problems, the
system of (36) is just solved only once, therefore, the coefficients in (35) can be stored before the time
evolution.

Algorithm 1. H-RBF-DQ METHOD

1: input the known terms in Equations (1) and (2)

2: generate the mesh pointes
3: set the time step Δt
4: set the label to distinguish the boundary nodes and inner nodes
5: ascertain the supporting points for every unknown node
6: calculate the weights in Equations (19) and (35)
7: while time t < = tend do
8: for every unknown point xi in Ω do
9: if inner node xi without any supporting point on the Neumann boundary then
10: form the Equation (20)
11: end if
12: if inner node xi with supporting points on the Neumann boundary then
13: form the Equation (37) with source (38)
14: end if
15: if node xi on the Neumann boundary then
16: form the boundary equation by (39)
17: end if
18: end for
19: assemble the solver matrix for every time step
20: solve the system for every time step
21: t = t + Δt
22: end while

4 NUMERICAL EXAMPLES

In this section, numerical experiments are carried out to demonstrate the effectiveness of the present
new H-RBF-DQ method for variable-order time fractional advection–diffusion equation with Neu-
mann boundary conditions. Three different test problems are chosen to show the capability and
accuracy of the currently proposed method. All the tests are carried on a Macbook Pro laptop (2.2
GHZ Intel Core i7, 16GB 1600 MHz DDR3) with Matlab R2015b.
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
12 LIU ET AL .

To show the accuracy of the proposed method, the L2 , L∞ errors and root-mean-square (RMS) of
errors are measured using the following definitions:
√
√N
√∑ ∑
N
√ | |2
|u (xi ) − u (xi )| ∕ |u (xi )| ,
2 h 2
L =
i=1 i=1
√ ( )
√
| | √1 ∑ N
L = max |u (xi ) − u (xi )| , RMS =
∞ h √ |u (xi ) − uh (xi )| ,
2
1≤i≤N | | N i=1 | |

where uh (xi ) is the numerical solution of u(xi ).

Example 1 Consider the following variable-order time fractional advection–diffusion

equation

𝜕 𝛼(x,t) u(x, t) 𝜕 2 u(x, t) 𝜕 2 u(x, t) 𝜕u(x, t) 𝜕u(x, t) 2t2−𝛼(x,t)

= + − − + + 2x + 2y − 4, (41)
𝜕t 𝛼(x,t) 𝜕x 2 𝜕y2 𝜕x 𝜕y Γ(3 − 𝛼(x, t))
with the exact solution u(x, t) = x2 + y2 + t2 . Equation (41) was given in paper [49] to
evaluate the MLS method to deal with Dirichlet boundary condition. In the present study,
the Neumann boundary condition is considered by the present Hermite RBF-DQ method.
The unit circle domain is considered to test the accuracy. The two constant orders (𝛼 = 0.5,
0.8) and a variable order

𝛼(x, t) = 0.8 − 0.1 cos(xt) sin(x) − 0.1 cos(yt) sin(y) (42)

are considered. The spatial convergence is evaluated by three different grids. For example,
the point distribution with N = 214 is shown in Figure 2a. A contour plot of numerical
solution is shown in Figure 2b. The comparison between the new H-RBF-DQ method
(39) and the RBF-DQ method (40) (see the paper [46]) is presented in Table 1. The results
show that the errors can be greatly reduced by the use of the present method. This demon-
strates that the new developed H-RBF-DQ method is more accurate than the RBF-DQ
method. In addition, the matrix condition numbers and time costs are also provided in
Table 1. Although the condition numbers for the H-RBF-DQ method are slightly bigger
than the RBF-DQ method, they are still in a reasonable region. In Table 2, the numerical
comparisons at final time t = 2 are also given. It is very interesting that the L∞ and RMS
error accumulation are slowly increased. However, for the present special example, with
the increase of time, the solution will increase, which will lead to the decrease of relative
L2 error. The grid convergence for variable-order (42) is provided in Table 3. The condi-
tion numbers and the time costs are also reported in Table 3. The time convergence for
variable-order (42) is provided in Table 4.

Similar to the RBF-DQ method, the approximated accuracy for the H-RBF-DQ method is also
affected by the shape parameter c. The impact of the shape parameter c is drawn in Figure 3, which
shows the choice c2 = 5 is suitable for the present example. For the choosing of c2 , one can refer to
papers [21, 46, 57–59].

Example 2 We will consider a variable-order fractional equations, say

c 𝛼(x,t)
( ) 𝜕u(x, t) 𝜕u(x, t)
0 Dt u(x, t) = 3 + t2 Δu(x, t) − − + f (x, t), (43)
𝜕x 𝜕y
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 13

(a) (b)

FIGURE 2 Point distribution of the Circle domain with 214 points and the contours of the numerical solution of Example 1

TABLE 1 The errors, condition numbers (Cond.), and the time costs (Sec.) obtained by local H-RBF-DQ and RBF-DQ
methods for Example 1 with Δt = 1/200 and N = 873 at time t = 1

H-RBF-DQ method (39) RBF-DQ method (40)

𝜶 L2 L∞ RMS Cond. Sec. L2 L∞ RMS Cond. Sec.
0.2 1.1694e-3 1.8845e-3 1.8420e-3 15420 116.62 1.5685e-3 3.4946e-3 2.4708e-3 3870.7 77.23
0.5 1.1053e-3 1.7853e-3 1.7411e-3 1275.3 112.72 1.7253e-3 3.7746e-3 2.7177e-3 332.04 73.74
0.8 2.6599e-4 4.6496e-4 4.1900e-4 255.83 115.93 2.4109e-3 4.9065e-3 3.7977e-3 186.44 77.36
𝛼(x, t) 9.1107e-5 2.8501e-4 1.4351e-4 418.56 117.63 2.6559e-3 5.4846e-3 4.1837e-3 187.62 77.70

Abbreviation: RMS, root-mean-square.

TABLE 2 The errors, condition numbers (Cond.), and the time costs (Sec.) obtained by local H-RBF-DQ and RBF-DQ
methods for Example 1 with Δt = 1/200 and N = 873 at time t = 2

H-RBF-DQ method (39) RBF-DQ method (40)

𝜶 L2 L∞ RMS Cond. Sec. L2 L∞ RMS Cond. Sec.
0.2 4.6448e-4 2.1592e-3 2.1158 e-3 15420 327.77 6.2522 e-4 3.8901 e-3 2.8481 e-3 3870.7 239.27
0.5 5.4052e-4 2.5081e-3 2.5081 e-3 1275.3 309.26 8.5173 e-4 4.9739 e-3 3.8799 e-3 332.04 226.17
0.8 1.5963e-4 7.7516e-4 7.2718 e-4 255.83 324.82 1.4713 e-3 7.8504 e-3 6.7024 e-3 186.44 239.28
𝛼(x, t) 2.8047e-5 2.2271e-4 1.2776 e-4 279.63 332.38 1.6118 e-3 8.5692 e-3 7.3424 e-3 186.36 246.11

Abbreviation: RMS, root-mean-square.

TABLE 3 Grid convergence, condition numbers and the time costs with Δt = 1/200 for Example 1 at time t = 1

N L2 L∞ RMS Cond. Sec.

56 2.4731e-3 3.7333e-3 3.4964e-3 26.698 12.09
105 2.1829e-3 3.2831e-3 3.1148e-3 58.505 20.48
214 6.6816e-4 1.2703e-3 1.0549e-3 38.317 36.12
421 3.6598e-4 7.2805e-4 5.7275e-4 77.264 62.56
873 9.1107e-5 2.8501e-4 1.4351e-4 418.56 117.6

Abbreviation: RMS, root-mean-square.

 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
14 LIU ET AL .

TABLE 4 Time discretization convergence, condition numbers and the time costs with fixed N = 873 for Example 1 at time
t=1

𝚫t L2 L∞ RMS Cond. Sec.

1/25 1.1449e-2 1.9981e-2 1.8035e-2 1860.1 12.84
1/50 4.5590e-3 8.1718e-3 7.1815e-3 1049.3 20.73
1/100 1.4366e-3 2.7570e-3 2.2630e-3 644.31 46.80
1/200 9.1107e-5 2.8501e-4 1.4351e-4 418.65 117.6

Abbreviation: RMS, root-mean-square.

FIGURE 3 Errors versus shape parameter c2 for the Hermite radial basis function-based differential quadrature (H-RBF-DQ)
method

FIGURE 4 Point distribution of the irregular domain with 2125 points

 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 15

(a)

(b)

FIGURE 5 The numerical solution and the error of Example 2

with Gauss pulse exact solution,

−(x−0.05)2 −(y−0.05)2
u(x, t) = t2 e 𝛽 ,
and the variable order 𝛼(x, t) in (42), where f (x, t) can be obtained by the exact solution.
In order to show the versatility of the method, an irregular computational domain with Neumann
boundary condition is considered again. The computational domain is formed by the polar coordinate
method [60], and is defined by
n+1
𝜌= (n + 1 − cos n𝜃), 𝜃 ∈ [0, 2𝜋].
n2
As an example, the geometry for n = 6 with 𝛽 = 0.1 is considered. The distribution of grid points
with total number 2125 points is shown in Figure 4. The numerical solution and the error are shown
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
16 LIU ET AL .

FIGURE 6 Point distribution of the irregular domain with 1364 points enforced by a hybrid Dirichlet and Neumann
boundary condition (Neumann boundary condition labeled by red “⋆”)

in Figure 5, which demonstrate the present method is effective for Gauss pulse. Furthermore, a com-
plicated boundary condition, a hybrid Dirichlet and Neumann boundary condition is also studied here.
For the sake of simplification, we rethink Equation (43) with hybrid boundary conditions, but the com-
putational domain is formed by a pentagonal star and a circular arc with 𝜋4 which is shown in Figure 6.
It is worth pointing out that such a geometry was first proposed in [46] for testing the RBF-DQ method.
The point coordinates on the boundary of the pentagonal star are represented by the following formula.
[90 + 72(i − 1)]𝜋 [90 + 72(i − 1)]𝜋
x2i−1 = R cos , y2i−1 = R sin , i = 1, 2, 3, 4
180 180
[126 + 72(i − 1)]𝜋 [126 + 72(i − 1)]𝜋
x2i = r cos , y2i = r sin , i = 1, 2, 3
180 180
where the parameters R = 2 and r = 1 are considered. As a hybrid boundary condition, the Neumann
boundary condition is enforced on the right of the shape labeled by red “⋆,” and the Dirichlet boundary
condition is enforced on the rest. Clearly, the computational domain is singular, which increases the
difficulty of using numerical methods. The meshless distribution with 1364 points is shown in Figure 6.
The total cost of computation is 140 s. It can be seen that although the boundaries with the Neumann
boundary conditions include singular corner points, the final numerical result seems still very fine as
shown in Figure 7, where is the error distribution plot. Hence, it can be concluded that the present
method handles the Neumann boundary conditions very well.
Example 3 As the final example, we studied a fractional approximate modeling of
air pollution, in which the process of transport and diffusion of pollutants in the atmo-
sphere with obstacles is investigated. The equation represents a mass balance statement
of pollutant materials when they are transported through the air, and is defined by
c 𝛼(x,t)
0 Dt C(x, t) = 𝜉ΔC(x, t) − v ⋅ ∇C(x, t) + f (x, t), (44)
where C(x, t) is the concentration of pollutant materials, v is the wind velocity and 𝜉 is
the diffusion coefficient.
Although, the authors in [49] considered the modeling of air pollution, the homogeneous Dirichlet
boundary condition is not realistic in the practical applications and the computational domain without
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 17

(a)

(b)

FIGURE 7 The numerical solution and the error of Example 2 enforced by the hybrid Dirichlet and Neumann boundary
conditions

obstacles is very simple. Because the penetrability cannot be satisfied near the computational bound-
aries, the method suggested in [49] is assumed that the air pollution does not touch the boundaries.
In our recent research [46], a solution for this problem was provided from an application with the
Neumann boundary conditions with obstacles by using RBF-DQ method, in which the geometry is
complicated with obstacles. For this case, the wind velocity distribution affected by the obstacles is
not constant, and is obtained by the method of computational fluid dynamics.
Here, we reconsidered the air pollution modeling with constant 𝜉 = 0.2, but with 𝛼(x, t) = 0.55
and 𝛼(x, t) = 0.55 + 0.45 sin(xyt), and a single source of pollutant with strength f = 5 is located
in the domain of 2.5 ≤ x ≤ 2.7 and 2.5 ≤ y ≤ 2.7. In this application, a computational grid with 3145
points and the time step Δt = 0.02 are used. For the simplicity, the boundaries of holes in the
10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
AL .
LIU ET

Pollutant distributions and the wind velocity vectors at different time

(b)

(d)

(f)
(a)

(c)

(e)

FIGURE 8
18
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 19

computational domain is enforced by extrapolation to satisfy wall no penetration boundary condition.

Using the H-RBF-DQ method, the pollutant distributions at different time are shown in Figure 8. It is
found that the present method captures the pollutant feature very well throughout the computational
domain. As expected, the concentration of pollutant materials with the variable-order time model
𝛼(x, t) = 0.55+0.45 sin(xyt) moves more slowly than that with the constant-order model 𝛼(x, t) = 0.55.
It should be noted that the present method guarantees the values of concentration are positive, and the
air pollution is successfully transferred across the boundaries as shown in Figure 8. The results show
that the proposed method can properly capture the distribution of pollutant on the domain with a given
velocity field.

5 CONCLUSION

A local meshless H-RBF-DQ method on Neumann boundary condition is developed in this paper
to simulate the fractional calculus problems with complex geometries. Although the method is pre-
sented based on the variable-order time fractional advection–diffusion equations, it is also suitable
for other PDEs with the Neumann boundary conditions. It is worth noting that the present Hermite
RBF method is very different from ones in [16, 43]. The advantage for the use of the current method
is that the coefficients of the shape functions are only related to the support points set, hence, for
time related problems, the system of equations is just solved only once, therefore, the coefficients can
be stored before the time evolution. Numerical examples show the present method is more accurate
than the RBF-DQ method. The current method is successfully applied to a variable-order fractional
model of air pollution. It is an evident that the numerical method can keep the values of the con-
centration are positive. Furthermore, our numerical examples demonstrate that the current method is
suitable for complicated boundary value problems, including hybrid Dirichlet and Neumann bound-
ary conditions. In the future work, the high order temporal discretization for other PDEs will be
studied.

D A T A A V A I L A B I L I T Y ST A T E M E N T
All the data and the source code used in this paper can be obtained from https://github.com/
jamesliu516/H-RBFDQ_DATA.git.

ORCID
Jianming Liu https://orcid.org/0000-0002-5890-0680
Xin Kai Li https://orcid.org/0000-0002-6281-8639
Xiuling Hu https://orcid.org/0000-0002-1357-9491

REFERENCES

[1] V. V. Uchaikin, Fractional derivatives for physicists and engineers, Higher Education Press, Beijing, 2013.
[2] I. Podlubny, Fractional differential equations, Academic Press, San Diego, 1999.
[3] W. Chen and H. Sun, Fractional differential equations and statistical models for anomalous diffusion, Science
Press, Beijing, 2017.
[4] F. Liu, P. Zhuang, and Q. Liu, Numerical method of fractional partial differential equation and its application,
Science Press, Beijing, 2015.
[5] X. Zhao, Z. Sun, and G. E. Karniadakis, Second-order approximations for variable order fractional derivatives:
Algorithms and applications, J. Comput. Phys. 293 (2015), 184–200.
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
20 LIU ET AL .

[6] H. Liu and S. Lü, Gauss-Lobatto-Legendre-Birkhoff pseudospectral approximations for the multi-term time frac-
tional diffusion-wave equation with Neumann boundary conditions, Numer. Methods Partial Differ. Equ. 34 (2018),
no. 6, 2217–2236.
[7] R. M. Ganji, H. Jafari, and S. Nemati, A new approach for solving integro-differential equations of variable order,
J. Comput. Appl. Math. 379 (2020), 112946.
[8] N. H. Tuan, R. M. Ganji, and H. Jafari, A numerical study of fractional rheological models and frac-
tional Newell-Whitehead-Segel equation with non-local and non-singular kernel, Chin. J. Phys. 68 (2020),
308–320.
[9] V. R. Hosseini, F. Yousefi, and W.-N. Zou, The numerical solution of high dimensional variable-order time
fractional diffusion equation via the singular boundary method, J. Adv. Res. 32 (2021), 73–84.
[10] R. M. Ganji, H. Jafari, and D. Baleanu, A new approach for solving multi variable orders differential equations
with Mittag-Leffler kernel, Chaos Soliton Fractals 130 (2020), 109405.
[11] V. R. Hosseini, M. Remazani, and W. Zou, The meshless approach for solving 2D variable-order time-fractional
advection-diffusion equation arising in anomalous transport, Eng. Comput. 38 (2022), 2289–2307. https://doi.
org/10.1007/s00366-021-01379-7.
[12] R. C. Mittal and S. Pandit, Quasilinearized scale-3 Haar wavelets-based algorithm for numerical simulation of
fractional dynamical systems, Eng. Comput. 35 (2018), no. 5, 1907–1931.
[13] R. C. Mittal and S. Pandit, A numerical algorithm to capture spin patterns of fractional Bloch nuclear magnetic
resonance flow models, J. Comput. Nonlinear Dyn. 14 (2019), no. 8, 081001.
[14] S. Pandit and R. C. Mittal, A numerical algorithm based on scale-3 Haar wavelets for fractional advection
dispersion equation, Eng. Comput. 38 (2021), no. 4, 1706–1724.
[15] T. Belytschko, Y. Krongauz, D. Organ, M. Fleming, and P. Krysl, Meshless methods: An overview and recent
developments, Comput. Methods Appl. Mech. Eng. 139 (1996), no. 1, 3–47.
[16] G. R. Liu and Y. T. Gu, An introduction to meshfree methods and their programming, Springer, Netherlands,
2005.
[17] G. Fasshauer, Meshfree approximation methods with MATLAB, World Scientific Publishers, Singapore,
2007.
[18] Y. Cheng, Meshfree methods, Science Press, Beijing, 2015.
[19] E. J. Kansa, Multiquadrics—A scattered data approximation scheme with applications to computational fluid
dynamics I surface approximations and partial derivative estimates, Comput. Math. Appl. 19 (1990), no. 8,
127–145.
[20] E. J. Kansa, Multiquadrics–A scattered data approximation scheme with applications to computational
fluid-dynamics-II solutions to parabolic, hyperbolic and elliptic partial differential equations, Comput. Math.
Appl. 19 (1990), no. 8, 147–161.
[21] C. Shu, H. Ding, and K. S. Yeo, Local radial basis function-based differential quadrature method and its applica-
tion to solve two-dimensional incompressible Navier-Stokes equations, Comput. Methods Appl. Mech. Eng. 192
(2003), no. 7, 941–954.
[22] C. Shu, H. Ding, H. Q. Chen, and T. G. Wang, An upwind local RBF-DQ method for simulation of inviscid
compressible flows, Comput. Methods Appl. Mech. Eng. 194 (2005), no. 18, 2001–2017.
[23] G. B. Wright and B. Fornberg, Scattered node compact finite difference-type formulas generated from radial basis
functions, J. Comput. Phys. 212 (2006), no. 1, 99–123.
[24] C. M. C. Roque, D. Cunha, C. Shu, and A. J. M. Ferreira, A local radial basis functions—Finite dif-
ferences technique for the analysis of composite plates, Eng. Anal. Bound. Elem. 35 (2011), no. 3,
363–374.
[25] M. Dehghan, M. Abbaszadeh, and A. Mohebbi, A meshless technique based on the local radial basis functions
collocation method for solving parabolic-parabolic Patlak-Keller-Segel chemotaxis model, Eng. Anal. Bound.
Elem. 56 (2015), 129–144.
[26] M. Dehghan and V. Mohammadi, A numerical scheme based on radial basis function finite difference (RBF-FD)
technique for solving the high-dimensional nonlinear Schrödinger equations using an explicit time discretization:
Runge–kutta method, Comput. Phys. Commun. 217 (2017), 23–34.
[27] V. Bayona, N. Flyer, B. Fornberg, and G. A. Barnett, On the role of polynomials in RBF-FD approximations: II.
Numerical solution of elliptic PDEs, J. Comput. Phys. 332 (2017), 257–273.
[28] R. Jiwari, Sanjay Kumar, R. C. Mittal, and J. Awrejcewicz, A meshfree approach for analysis and computational
modeling of non-linear Schrödinger equation, Comput. Appl. Math. 39 (2020), no. 2, 95.
[29] S. Kumar, R. Jiwari, R. C. Mittal, and J. Awrejcewicz, Dark and bright soliton solutions and computational
modeling of nonlinear regularized long wave model, Nonlinear Dyn. 104 (2021), no. 1, 661–682.
[30] R. Jiwari and A. Gerisch, A local radial basis function differential quadrature semi-discretisation tech-
nique for the simulation of time-dependent reaction-diffusion problems, Eng. Comput. 38 (2021), no. 6,
2666–2691.
[31] S. Kumar, R. Jiwari, and R. C. Mittal, Radial basis functions based meshfree schemes for the simulation of
non-linear extended Fisher-Kolmogorov model, Wave Motion 109 (2022), 102863.
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
LIU ET AL . 21

[32] M. Dehghan and V. Mohammadi, The numerical solution of Cahn-Hilliard (CH) equation in one, two and
three-dimensions via globally radial basis functions (GRBFs) and RBFs-differential quadrature (RBFs-DQ)
methods, Eng. Anal. Bound. Elem. 51 (2015), 74–100.
[33] C. Shu, Differential quadrature and its application in engineering, Springer, London, 2000.
[34] M. Dehghan and A. Nikpour, Numerical solution of the system of second-order boundary value problems using
the local radial basis functions based differential quadrature collocation method, Appl. Math. Model. 37 (2013),
no. 18, 8578–8599.
[35] M. R. Hashemi and F. Hatam, Unsteady seepage analysis using local radial basis function-based differential
quadrature method, Appl. Math. Model. 35 (2011), no. 10, 4934–4950.
[36] Y. L. Chan, L. H. Shen, C. T. Wu, and D. L. Young, A novel upwind-based local radial basis function differential
quadrature method for convection-dominated flows, Comput. Fluids 89 (2014), 157–166.
[37] M. Dehghan and M. Shirzadi, Meshless simulation of stochastic advection-diffusion equations based on radial
basis functions, Eng. Anal. Bound. Elem. 53 (2015), 18–26.
[38] A. Golbabai and A. Nikpour, Computing a numerical solution of two dimensional non-linear Schrödinger equation
on complexly shaped domains by RBF based differential quadrature method, J. Comput. Phys. 322 (2016),
586–602.
[39] A. Golbabai and E. Mohebianfar, A new method for evaluating options based on multiquadric RBF-FD method,
Appl. Math. Comput. 308 (2017), 130–141.
[40] M. Dehghan and M. Abbaszadeh, Solution of multi-dimensional Klein-Gordon-Zakharov and
Schrödinger/Gross-Pitaevskii equations via local radial basis functions-differential quadrature
(RBF-DQ) technique on non-rectangular computational domains, Eng. Anal. Bound. Elem. 92 (2018),
156–170.
[41] M. Dehghan and N. Shafieeabyaneh, Local radial basis function–finite-difference method to simulate some mod-
els in the nonlinear wave phenomena: Regularized long-wave and extended Fisher-Kolmogorov equations, Eng.
Comput. 37 (2021), no. 2, 1159–1179.
[42] T. J. Liszka, C. A. M. Duarte, and W. W. Tworzydlo, hp-meshless cloud method, Comput. Methods Appl. Mech.
Eng. 139 (1996), no. 1, 263–288.
[43] X. Liu, G. R. Liu, K. Tai, and K. Y. Lam, Radial point interpolation collocation method (RPICM) for partial
differential equations, Comput. Math. Appl. 50 (2005), no. 8–9, 1425–1442.
[44] D. Stevens, H. Power, M. Lees, and H. Morvan, The use of PDE centres in the local RBF Hermitian method for
3D convective-diffusion problems, J. Comput. Phys. 228 (2009), no. 12, 4606–4624.
[45] A. Krowiak, Hermite type radial basis function-based differential quadrature method for higher order equations,
Appl. Math. Model. 40 (2016), no. 3, 2421–2430.
[46] J. Liu, X. Li, and X. Hu, A RBF-based differential quadrature method for solving two-dimensional variable-order
time fractional advection-diffusion equation, J. Comput. Phys. 384 (2019), 222–238.
[47] P. Zhuang, Y. T. Gu, F. Liu, I. Turner, and P. K. D. V. Yarlagadda, Time-dependent fractional advection-diffusion
equations by an implicit MLS meshless method, Int. J. Numer. Methods Eng. 88 (2011), no. 13,
1346–1362.
[48] A. Mardani, M. R. Hooshmandasl, M. H. Heydari, and C. Cattani, A meshless method for solving the time
fractional advection-diffusion equation with variable coefficients, Comput. Math. Appl. 75 (2018), no. 1,
122–133.
[49] A. Tayebi, Y. Shekari, and M. H. Heydari, A meshless method for solving two-dimensional variable-order time
fractional advection diffusion equation, J. Comput. Phys. 340 (2017), 655–669.
[50] J. Lin, J. Bai, S. Reutskiy, and J. Lu, A novel RBF-based meshless method for solving time-fractional
transport equations in 2D and 3D arbitrary domains, Eng. Comput. (2022), 1–18. https://doi.org/10.1007/
s00366-022-01601-0.
[51] F. Zhuo-Jia, L.-W. Yang, Q. Xi, and C.-S. Liu, A boundary collocation method for anomalous heat conduction
analysis in functionally graded materials, Comput. Math. Appl. 88 (2021), 91–109.
[52] Z. Tang, F. Zhuojia, H. G. Sun, and X. Liu, An efficient localized collocation solver for anomalous diffusion on
surfaces, Fract. Calc. Appl. Anal. 24 (2021), no. 3, 865–894.
[53] Q. Xi, F. Zhuojia, T. Rabczuk, and D. Yin, A localized collocation scheme with fundamental solutions for long-time
anomalous heat conduction analysis in functionally graded materials, Int. J. Heat Mass Transf. 180 (2021),
121778.
[54] Y. Lin and C. Xu, Finite difference/spectral approximations for the time-fractional diffusion equation, J. Comput.
Phys. 225 (2007), no. 2, 1533–1552.
[55] Y. L. Wu and C. Shu, Development of RBF-DQ method for derivative approximation and its application to simulate
natural convection in concentric annuli, Comput. Mech. 29 (2002), no. 6, 477–485.
[56] P.-O. Persson and S. Gilbert, A simple mesh generator in MATLAB, SIAM Rev. 46 (2004), no. 2,
329–345.
[57] G. E. Fasshauer and J. G. Zhang, On choosing “optimal”shape parameters for rbf approximation, Numer.
Algorithms 45 (2007), no. 1, 345–368.
 10982426, 0, Downloaded from https://onlinelibrary.wiley.com/doi/10.1002/num.22997 by Indian Institute Of Technology, Wiley Online Library on [17/01/2023]. See the Terms and Conditions (https://onlinelibrary.wiley.com/terms-and-conditions) on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
22 LIU ET AL .

[58] C. S. Chen, W. Chen, and Z.-J. Fu, Recent advances in radial basis function collocation methods, Springer Briefs
in Applied Sciences and Technology, Springer, Heidelberg, 2014.
[59] F. Zhuo-Jia, J. Zhang, P.-W. Li, and J.-H. Zheng, A semi-lagrangian meshless framework for numerical solutions
of two-dimensional sloshing phenomenon, Eng. Anal. Bound. Elem. 112 (2020), 58–67.
[60] M. Hajiketabi and S. Abbasbandy, The combination of meshless method based on radial basis functions with a
geometric numerical integration method for solving partial differential equations: Application to the heat equation,
Eng. Anal. Bound. Elem. 87 (2018), 36–46.

How to cite this article: J. Liu, X. K. Li, and X. Hu, A novel local Hermite radial basis
function-based differential quadrature method for solving two-dimensional variable-order
time fractional advection–diffusion equation with Neumann boundary conditions, Numer.
Methods Partial Differ. Eq. (2023), 1–22. https://doi.org/10.1002/num.22997