Loading [MathJax]/jax/output/SVG/fonts/TeX/fontdata.js
Research article Special Issues

An artificial neural network approach for a class of time-fractional diffusion and diffusion-wave equations

  • In this paper, the artificial neural network method is applied to solve the time-fractional diffusion and diffusion-wave equations. This method combines Taylor series and neural network method, and uses the terms of different power terms of Taylor series as neurons. An error function is given to update the weights of the proposed neural network. In addition, in order to balance the contributions of different error terms in the error function, we propose an adaptive weight adjustment method. In the end, four numerical examples are given to demonstrate the effectiveness of proposed method in solving the time-fractional diffusion and diffusion-wave equations.

    Citation: Yinlin Ye, Hongtao Fan, Yajing Li, Ao Huang, Weiheng He. An artificial neural network approach for a class of time-fractional diffusion and diffusion-wave equations[J]. Networks and Heterogeneous Media, 2023, 18(3): 1083-1104. doi: 10.3934/nhm.2023047

    Related Papers:

    [1] Andrea Corli, Lorenzo di Ruvo, Luisa Malaguti, Massimiliano D. Rosini . Traveling waves for degenerate diffusive equations on networks. Networks and Heterogeneous Media, 2017, 12(3): 339-370. doi: 10.3934/nhm.2017015
    [2] Feiyang Peng, Yanbin Tang . Inverse problem of determining diffusion matrix between different structures for time fractional diffusion equation. Networks and Heterogeneous Media, 2024, 19(1): 291-304. doi: 10.3934/nhm.2024013
    [3] Kexin Li, Hu Chen, Shusen Xie . Error estimate of L1-ADI scheme for two-dimensional multi-term time fractional diffusion equation. Networks and Heterogeneous Media, 2023, 18(4): 1454-1470. doi: 10.3934/nhm.2023064
    [4] Farman Ali Shah, Kamran, Dania Santina, Nabil Mlaiki, Salma Aljawi . Application of a hybrid pseudospectral method to a new two-dimensional multi-term mixed sub-diffusion and wave-diffusion equation of fractional order. Networks and Heterogeneous Media, 2024, 19(1): 44-85. doi: 10.3934/nhm.2024003
    [5] Yaxin Hou, Cao Wen, Yang Liu, Hong Li . A two-grid ADI finite element approximation for a nonlinear distributed-order fractional sub-diffusion equation. Networks and Heterogeneous Media, 2023, 18(2): 855-876. doi: 10.3934/nhm.2023037
    [6] Narcisa Apreutesei, Vitaly Volpert . Reaction-diffusion waves with nonlinear boundary conditions. Networks and Heterogeneous Media, 2013, 8(1): 23-35. doi: 10.3934/nhm.2013.8.23
    [7] Robert Carlson . Myopic models of population dynamics on infinite networks. Networks and Heterogeneous Media, 2014, 9(3): 477-499. doi: 10.3934/nhm.2014.9.477
    [8] Caihong Gu, Yanbin Tang . Global solution to the Cauchy problem of fractional drift diffusion system with power-law nonlinearity. Networks and Heterogeneous Media, 2023, 18(1): 109-139. doi: 10.3934/nhm.2023005
    [9] Elisabeth Logak, Isabelle Passat . An epidemic model with nonlocal diffusion on networks. Networks and Heterogeneous Media, 2016, 11(4): 693-719. doi: 10.3934/nhm.2016014
    [10] Stefan Berres, Ricardo Ruiz-Baier, Hartmut Schwandt, Elmer M. Tory . An adaptive finite-volume method for a model of two-phase pedestrian flow. Networks and Heterogeneous Media, 2011, 6(3): 401-423. doi: 10.3934/nhm.2011.6.401
  • In this paper, the artificial neural network method is applied to solve the time-fractional diffusion and diffusion-wave equations. This method combines Taylor series and neural network method, and uses the terms of different power terms of Taylor series as neurons. An error function is given to update the weights of the proposed neural network. In addition, in order to balance the contributions of different error terms in the error function, we propose an adaptive weight adjustment method. In the end, four numerical examples are given to demonstrate the effectiveness of proposed method in solving the time-fractional diffusion and diffusion-wave equations.



    Fractional calculus is an important branch of mathematics. As its theory has been successfully applied to various fields in recent years, people have gradually found that fractional calculus can describe some "abnormal" phenomena in natural science and engineering applications. For example, mechanical behavior of viscoelastic materials [1], plasma motion under high temperature and high pressure [2], abnormal diffusion of extremely cold atoms or cell protoplasm [3], etc. Time fractional partial differential equation is an important model. When its fractional order α(0,1), it is called time-fractional diffusion equation; when α(1,2), it is called time-fractional diffusion-wave equation. The time-fractional diffusion equation can describe the abnormal slow diffusion phenomenon in porous media with fractal structure [4], while the time-fractional diffusion-wave equation can describe the propagation of mechanical scattered waves in viscoelastic media with power-law creep characteristics [5].

    It is well known that the analytical solution of time-fractional diffusion and diffusion-wave equations is notoriously difficult to obtain. In order to promote the development and progress of science, a great quantity of researchers and scholars have studied the numerical approximation method of time-fractional diffusion and diffusion-wave equations. In recent years, Gu and his collaborators have proposed many effective methods to solve the time-fractional diffusion equations, such as quadratic spline collocation method [6], Lagrange interpolation method [7], parallel-in-time iterative algorithm [8] and the graded L1 scheme method [9]. For the time-fractional diffusion-wave equations, many classical numerical methods have been proposed, including finite difference method [10,11], element-free Galerkin method [12], wavelets method [13,14], meshless local collocation method [15], etc.

    Moreover, artificial neural network method is also proposed to solve fractional differential equations. With the rapid development of machine learning technology, the demand for solving fractional differential equations by using artificial neural network method has increased by leaps and bounds, and some excellent results have been achieved. In 2017, Qu [16] applied cosine radial basis function neural network to solve fractional differential equations with initial value or boundary value problems for the first time. In 2018, Rostami et al. [17] adopted an iterative algorithm combining power series and neural network method to approximately seek the solution of high-order linear and ordinary differential equations. Rizaner et al. [18] proposed a numerical approch for approximately solving the first-order initial value problem using unsupervised radial basis function network, and verified the effectiveness with some numerical examples. In 2020, Hadian-Rasanan et al. [20] solved different types of Lane-Emden equation by constructing a single-layer orthogonal network with a fractional Legendre function as the activation function of the hidden layer, and used the Levenberg-Marquardt algorithm to train them. The above literature is to solve fractional ordinary differential equations by making use of artificial neural network method. However, there are few literatures on solving time-fractional partial equations by using artificial neural networks. Jafarian et al. regarded the artificial neural network method with a generalized sigmoid function as the cost function as the iterative scheme for solving linear fractional ordinary differential equations in [19]. Recently, Qu et al. [21] proposed a neural network method based on cosine basis and sine basis to solve a class of time-fractional diffusion and diffusion-wave equation with zero boundary conditions. Deep neural network methods for solving forward and inverse problems of time-fractional diffusion equations with conformable derivatives were introduced by Ye et al. [22]. In this paper, a new numerical method is proposed for a class of time-fractional diffusion and diffusion-wave equations with Dirichlet boundary conditions by combining Taylor series method with artificial neural network.

    In this paper, We consider the time-fractional diffusion equation with the following form:

    {cDα0,tu(x,t)+c2uxx(x,t)=f(x,t),u(0,t)=λ(t),u(L,t)=μ(t),u(x,0)=φ(x),0<α<1,0<xP,0<tQ, (1.1)

    and the time-fractional diffusion-wave equation with the form as follows:

    {cDα0,tu(x,t)+c2uxx(x,t)=f(x,t),u(0,t)=λ(t),u(L,t)=μ(t),u(x,0)=φ(x),ut(x,0)=ψ(x)=0,1<α<2,0<xP,0<tQ. (1.2)

    For the time-fractional diffusion equation (1.1) and the time-fractional diffusion-wave equation (1.2), we establish a single-layer neural network model with Taylor series of different power terms as neurons, and give the error iteration formula and neural network weight update formula. It is worth mentioning that we have adaptively adjusted the weights of different error terms, which will be shown in Section 4.

    The main contributions of this paper are as below:

    ● The single layer functional link artificial neural network based on Taylor series is established for the first time to solve the time-fractional diffusion and diffusion-wave equation with more general form.

    ● The error function composed of initial error term, boundary error term and internal error term is constructed to update the unknown parameters in the neural network. In addition, the adaptive weight adjustment method is proposed to balance the different error terms.

    ● Compared with the traditional iterative numerical method, the proposed neural network method can calculate the solution at any position in the domain without repeating the whole iterative process.

    The rest of this paper is arranged as follows. Section 2 briefly introduces the related definition and important properties of fractional calculus. In Section 3, we present the main idea of our neural network methods for solving the time-fractional diffusion and diffusion-wave equations (1.1), (1.2). In Section 4, we propose an adaptive weight adjustment method to adjust the weights of different error terms in the iteration. After that, four numerical examples are given to illustrate the effectiveness of the method in Section 5. Finally, we summarize the main work of our paper in Section 6.

    This section gives some basic definition and related properties about fractional calculus; for more details see [23].

    Definition 2.1. The Riemann-Liouville fractional integral of order α>0 for a function f(x)L1((0,Q]) is defined as

    RLIα0+(f(x))=1Γ(α)x0(xt)α1f(t)dt,

    where Γ() denotes the Gamma function.

    Definition 2.2. The Caputo fractional derivative of order α>0 for a function f(x)AC((0,Q]) with nN+ is defined as

    cDα0,xf(x)={RLInα0+(ddx)n(f(x)),n1<α<n,(ddx)nf(x),α=n.

    Since the derivative of a constant is 0, the following useful property holds in the Caputo sense:

    cDα0,x(xk)={0,kZ+,k<α,Γ(k+1)Γ(k+1α)xkα,kZ+,kα,x>0.

    Definition 2.3. The Mittag-Leffler function is defined by

    Eα,β(z)=k=0zkΓ(αk+β),

    where z,βC, α>0, and C is the usual set of complex numbers. Particularly, E1,1(z)=exp(z), that is, the common exponential function.

    In order to facilitate the calculation of the Mittag-Leffler function later, the definition of the extended sine and cos functions and related properties are given [21,23]. The calculation code of the Mittag-Leffler function can be found in [24].

    Definition 2.4. The extended sine function Sinα,β(x) and cos function Cosα,β(x) are defined as

    Sinα,β(x)=k=1(1)k+1x2k1Γ(α(2k1)+β),Cosα,β(x)=k=0(1)kx2kΓ(α(2k)+β),

    where x,βC, α>0. According to the Euler formulas, we have

    Sinα,β(x)=Eα,β(ix)Eα,β(ix)2i,Cosα,β(x)=Eα,β(ix)+Eα,β(ix)2.

    In particular, we have Sin1,1(x)=sin(x), Cos1,1(x)=cos(x).

    Lemma 2.1. For the Caputo fractional derivative of some special functions, we have three useful results as follows:

    cDα0,t(exp(t))=cDα0,t(E1,1(t))=t1αE1,2α(t),
    cDα0,t(sin(t))=cDα0,t(Sin1,1(t))=tαSin1,1α(t),

    and

    cDα0,t(cos(t))=cDα0,t(Cos1,1(t))=tαCos1,1α(t).

    With regard to the time-fractional diffusion equation (1.1) and the time-fractional diffusion-wave equation (1.2), the proposed method is described in detail in this section. In order to solve the above two problems, we built a single layer neural network with a series of rapidly convergent components of Taylor series as neurons, as shown in Figure 1.

    Figure 1.  Neural network architecture for solving the problem (1.1) and (1.2).

    In simple terms, we construct the k-th test solution in the following form:

    uk(x,t)=Ni=0ij=0θki(i+1)2+j+1Tij, (3.1)

    with the first few Taylor components:

    T00(u)=1,T10(u)=t,T01(u)=x,

    and the higher order components of Taylor series may be expressed as the modalities as below:

    Tij=xjtij.

    Here, NN+, (N+1)(N+2)2 and θki(i+1)2+j+1 represent the highest power of the series, the number of neurons and the unknown weights of the neural network in the k-th iteration, respectively.

    Since the weights of the neural network are unknown, in the first iteration, the weights are randomly preseted by a Gaussian distribution. In the subsequent iterations, we renew the weights through the error function. For the initial boundary value problems like Eqs (1.1) and (1.2), the error function is composed of the following three terms:

    Ek=12Ekin2F+12Ekini22+12Ekbou2F, (3.2)

    where F is Frobenius matrix norm and 2 represents vector 2-norm. Let xm=(m1)PNm1, tn=(n1)QNn1, m=1,2,,Nm, n=1,2,,Nn. The matrix Ekin is

    (ek2,2ek2,3ek2,Nnek3,2ek3,3ek3,NnekNm1,2ekNm1,3ekNm1,Nn), (3.3)

    with

    ekm,n=cD0,tαuk(xm,tn)+a2ukxx(xm,tn)f(xm,tn), (3.4)

    for 2mNm1, 2nNn. The vector Ekini is

    (ek1,1ek2,1ekNm,1), (3.5)

    with

    ekm,1=uk(xm,0)φ(xm). (3.6)

    The matrix Ekbou is

    (ek1,1ek1,2ek1,NnekNm,1ekNm,2eNm,Nkn), (3.7)

    with

    ek1,n=uk(0,tn)λ(tn), (3.8)

    and

    ekNm,n=uk(P,tn)μ(tn). (3.9)

    As for the internal points, by substituting the k-th test solution (3.1) into Eq (3.4) and making use of Definition 2.2, we further obtain that

    ekm,n=cD0,tαNi=0ij=0θki(i+1)2+j+1xjmtijn+c2j(j1)Ni=0ij=0θki(i+1)2+j+1xj2mtijnf(xm,tn)=iΓ(ij+1)Γ(ij+1α)Ni=0i1j=0θki(i+1)2+j+1xjmtijαn+c2j(j1)Ni=0ij=0θki(i+1)2+j+1xj2mtijnf(xm,tn). (3.10)

    Concerning the initial points, we gain the initial error by substituting the k-th test solution (3.1) into Eq (3.6)

    ekm,1=Ni=0θk(i+1)(i+2)2ximφ(xm). (3.11)

    With regard to the boundary points, substituting the k-th test solution (3.1) into Eqs (3.8) and (3.9) we acquire the boundary errors

    ek1,n=Ni=0θki(i+1)2+1tinλ(tn), (3.12)

    and

    ekNm,n=Ni=0ij=0θki(i+1)2+j+1Pjtijnμ(tn). (3.13)

    Here, we use the gradient descent method to update the k-th unknown weights θki(i+1)2+j+1. For more details, please refer to [25].

    The weight update formula is as follow:

    θk+1i(i+1)2+j+1=θki(i+1)2+j+1+Δθki(i+1)2+j+1,k=0,1,2,,K, (3.14)

    with

    Δθki(i+1)2+j+1=ηEkθki(i+1)2+j+1=ηNm,Nnm=1,n=1Eekekm,nekm,nθki(i+1)2+j+1=ηNm1,Nnm=2,n=212{Nm1,Nnm=2,n=2(ekm,n)2}ekm,nekm,nθki(i+1)2+j+1ηNmm=112{Nmm=1(ekm,1)2}ekm,1ekm,1θki(i+1)2+j+1ηNnn=112{Nnn=1(ek1,n)2}ek1,nek1,nθki(i+1)2+j+1ηNnn=112{Nnn=1(ekNm,n)2}ekNm,nekNm,nθki(i+1)2+j+1=ηNm1,Nnm=2,n=2(ekm,n)2ekm,nθki(i+1)2+j+1ηNmm=1(ekm,1)2ekm,1θki(i+1)2+j+1ηNnn=1(ek1,n)2ek1,nθki(i+1)2+j+1ηNnn=1(ekNm,n)2ekNm,nθki(i+1)2+j+1ηNm1,Nnm=2,n=2(ekm,n)×{iΓ(ij+1)Γ(ij+1α)Ni=0i1j=0xjmtijαn+a2j(j1)Ni=0ij=0xj2mtijn}ηNmm=1(ekm,1)Ni=0ximηNnn=1(ek1,n)Ni=0tinηNnn=1(ekNm,n)Ni=0ij=0Pjtijn.

    Here, θ0i(i+1)2+j+1 are some numbers randomly generated by Gaussian distribution, K is an integer representing the maximum number of iterations, and η is the learning rate of neural network.

    For the time-fractional wave equation (1.2), we just need to rewrite Eqs (3.10) and (3.14) into

    ekm,n=cD0,tαNi=0ij=0θki(i+1)2+j+1xjmtijn+a2j(j1)Ni=0ij=0θki(i+1)2+j+1xj2mtijnf(xm,tn)=i(i1)Γ(ij+1)Γ(ij+1α)Ni=0i2j=0θki(i+1)2+j+1xjmtijαn+a2j(j1)Ni=0ij=0θki(i+1)2+j+1xj2mtijnf(xm,tn), (3.15)

    and

    θk+1i(i+1)2+j+1=θki(i+1)2+j+1+Δθki(i+1)2+j+1,k=0,1,2,,K, (3.16)

    with

    Δθki(i+1)2+j+1=ηEkθki(i+1)2+j+1=ηNm1,Nnm=2,n=2(ekm,n){i(i1)Γ(ij+1)Γ(ij+1α)Ni=0i2j=0xjmtijαn+a2j(j1)Ni=0ij=0xj2mtijn}ηNmm=1(ekm,1)Ni=0ximηNnn=1(ek1,n)Ni=0tinηNnn=1(ekNm,n)Ni=0ij=0Pjtijn,

    respectively.

    It is well known that a system of partial differential equations can have infinite solutions without specifying proper bounds or initial conditions [26]. So is the time-fractional diffusion and diffusion-wave equation. Therefore, we believe that the gradient of internal error (ˉθEin), the gradient of initial error (ˉθEini) and the gradient of boundary error (ˉθEbou) should match each other in magnitude, otherwise the network will eventually tend to learn any solution satisfying the equation due to the dominant effect of the gradient of internal error, and then, it is easy to fall back to the wrong prediction. In this section, let us talk over how to balance the interaction between internal error term (Ein), initial error term (Eini) and boundary error term (Ebou). For simplicity, ˉθ denotes θi(i+1)2+j+1 in this section.

    Recall the error function Eq (3.2) and weight update formula Eq (3.14) proposed in Section 3, we get that

    ˉθk+1=ˉθkηˉθkEk=ˉθkη[12ˉθkEkin2F+12ˉθkEkini22+12ˉθkEkbou2F]. (4.1)

    Usually, the internal error term dominates the overall. To balance the interaction between different terms in this error, a simple method is to weight the initial error term and the boundary error term. Thus, Eq (3.2) and Eq (3.14) can be rewritten as

    Ek=12Ekin2F+12ωk1Ekini22+12ωk2Ekbou2F, (4.2)

    and

    ˉθk+1=ˉθkηˉθkEk=ˉθkη[12ˉθkEkin2F+12ωk1ˉθkEkini22+12ωk2ˉθkEkbou2F]=ηNm1,Nnm=2,n=2(ekm,n){iΓ(ij+1)Γ(ij+1α)Ni=0i1j=0xjmtijαn+c2j(j1)Ni=0ij=0xj2mtijn}ηωk1Nmm=1(ekm,1)Ni=0ximηωk2Nnn=1(ek1,n)Ni=0tinηωk2Nnn=1(ekNm,n)Ni=0ij=0Pjtijn, (4.3)

    respectively, ωk1,ωk2 are the weighted coefficients. This method is similar to the weighted optimization method, which improves the optimization effect by penalizing the initial error term and boundary error term.

    For the time-fractional diffusion-wave equation (1.2), Eq (3.16) can be rewritten as

    ˉθk+1=ˉθkηˉθkEk=ˉθkη[12ˉθkEkin2F+12ωk1ˉθkEkini22+12ωk2ˉθkEkbou2F]=ηNm1,Nnm=2,n=2(ekm,n){i(i1)Γ(ij+1)Γ(ij+1α)Ni=0i2j=0xjmtijαn+c2j(j1)Ni=0ij=0xj2mtijn}ηωk1Nmm=1(ekm,1)Ni=0ximηωk2Nnn=1(ek1,n)Ni=0tinηωk2Nnn=1(ekNm,n)Ni=0ij=0Pjtijn. (4.4)

    Next, we will discuss how to choose the weighted coefficients ωk1,ωk2. For different models, the optimal constant can be very different, which means that we cannot find a fixed empirical formula that can be transferred between different time-fractional diffusion and diffusion-wave equations. Also, the error function always consists of different parts that are used to constrain the equation. It is impractical to manually assign different weights to different terms of the error function. The traditional parameter adjustment method can neither guarantee the efficiency nor ensure the optimization to the global optimal value. Therefore, inspired by the learning rate annealing algorithm in [27], we propose an adaptive weight coefficients adjustment method.

    At the (k+1)th iterations, we calculate the estimated values of ωk1 and ωk2 by the formula as below:

    ^ω1k+1=¯|ˉθEin|¯|ˉθEini|,^ω2k+1=¯|ˉθEin|¯|ˉθEbou|, (4.5)

    where ¯|ˉθEin|, ¯|ˉθEini| and ¯|ˉθEbou| denote the means of |ˉθEin|, |ˉθEini| and |ˉθEbou|.

    Then, the weighted coefficients of (k+1) iterations are updated by the weighted moving average method:

    ωk+11=(1ξ)ωk1+ξ^ω1k+1,ωk+12=(1ξ)ωk2+ξ^ω2k+1 (4.6)

    with ξ is given as 0.1.

    Finally, the algorithm of this paper is summarized as Algorithm 1.

    Algorithm 1 Neural network method based on adaptive weight adjustment

    1: Initialization: Generate grid points (xm,tn)(m=1,2,,Nm, n=1,2,,Nn) on the rectangular area [0,P]×[0,Q]. The weights θ0i(i+1)2+j+1 (i=0,1,,N, j=0,1,,i) are randomly generated by Gaussian distribution.
    2: while the error function Ek<ϵ (ϵ is a small parameter, which is given as 107) or the iteration time k<K do
    3:  Compute the error function Ek, k>K (K is maximum number of iterations) and weight variation Δθki(i+1)2+j+1 (i=0,1,,N, j=0,1,,i).
    4:  Adjust the weights θki(i+1)2+j+1 according to the Eq (4.3).
    5:  Calculate the weighted coefficients ωk1,ωk2 by the Eqs (4.5) and (4.6).
    6: end while
    7: Save network weights θki(i+1)2+j+1, calculate neural network solution uk(x,t) and test other points on the solution area [0,P]×[0,Q].

    In this section, we give four examples to illustrate the applicability of our proposed neural network method in solving time-fractional diffusion equation (1.1) and time-fractional diffusion-wave equation (1.2). All experiments were performed on a computer with the following configurations: Intel(R) Core(TM) i5-8265U CPU @ 1.60 GHz 1.80 GHz using MATLAB R2019a.

    Example 1. We consider Eq (1.1) with the conditions: λ(x)=μ(x)=0, φ(x)=sin(πxP) and f(x,t)=t1αsin(πxP)E1,2α(t)(cπP)2exp(t)sin(πxP). The exact solution of the equation may be calculated by means of Definition 2.4 and Lemma 2.1.

    u(x,t)=sin(πxP)exp(t).

    At first, the parameters of the model are set to the following: P=1, Q=1, c=1 and 0<α<1. Then, the parameters of the proposed method we use are as follows: N=5, m=5, n=5, learning rate μ=4×104, and accuracy threshold ϵ=1×107. When α=0.9, Figure 2 represent the comparisons between the exact solution and the neural network solution, and the number of iterations is 2000, 5000, 10000, 20000, respectively. We observe that the approximate solution gets better and better as the number of iterations increases. Then, we check the approximate solution on the grid of Nm×Nn=21×21 in Figure 3a. The variation of the error value with the number of iterations is shown in Figure 4. Also, in Figure 3, a comparison of the exact and the approximate solution is given under α=0.3,0.5 and 0.7. In addition, we list the numerical solutions and absolute errors of some specific test points for α=0.3,0.5,0.7 and 0.9 in Table 1. We can see that we get similar results for different fractional order α. Eventually, the weights of the neural network are listed in Table 2.

    Figure 2.  The exact solution and the neural network solution for Example 1 at α=0.9. Red: exact solution, blue: neural network solution.
    Figure 3.  The exact solution and the neural network solution on the grid of Nm×Nn=21×21 for Example 1 at α=0.9 with k=20000. Red: exact solution, blue: neural network solution.
    Figure 4.  Error decreasing trend with regard to iterations for fractional order α=0.9 and maximum training time K=20000.
    Table 1.  The numerical solutions and absolute errors for Example 1 at α=0.3,0.5,0.7 and 0.9 when k=20000.
    0.3 0.5 0.7 0.9
    solution error solution error solution error solution error
    (0,0) 0.0029 0.0029 0.0046 0.0046 0.0031 0.0031 0.0030 0.0030
    (0.2,0.2) 0.6996 0.0184 0.7029 0.0150 0.7034 0.0145 0.7003 0.0176
    (0.4,0.4) 1.4004 0.0184 1.4026 0.0162 1.4084 0.0104 1.4042 0.0147
    (0.6,0.6) 1.7059 0.0270 1.7076 0.0254 1.7087 0.0243 1.7054 0.0275
    (0.8,0.8) 1.2730 0.0351 1.2781 0.0301 1.2695 0.0386 1.2694 0.0387
    (1,1) 0.0025 0.0025 0.0014 0.0014 0.0024 0.0024 0.0032 0.0032

     | Show Table
    DownLoad: CSV
    Table 2.  The weights of the proposed neural network method for Example 1 at α=0.3,0.5,0.7 and 0.9 when k=20000.
    Fractional order θki(i+1)2+j+1,k=20000
    0.3 0.0029 -0.0344 2.9583 -0.0048 3.4271 0.3197 0.0755
    2.1478 -1.7237 -3.7582 0.1448 0.4967 -1.4288 -4.5874
    -2.9112 -0.1865 0.1164 -0.4064 -1.0538 3.0193 3.3892
    0.5 0.0046 0.0168 2.9111 -0.2540 3.5760 0.5170 0.3759
    2.0230 -2.0247 -4.1744 0.1251 0.5603 -1.1722 -4.4873
    -2.4350 -0.2712 0.0068 -0.2784 -1.2850 3.0901 3.1768
    0.7 0.0031 -0.0509 2.9372 0.2153 3.5563 0.4696 -0.2334
    1.6944 -1.8766 -4.1846 -0.0421 0.8784 -0.7943 -4.6371
    -2.4028 0.1078 0.0842 -0.8617 -1.0688 3.0298 3.1788
    0.9 0.0030 -0.0381 2.9011 0.1165 3.5459 0.6432 -0.0392
    1.8245 -2.1044 -4.4339 -0.1421 0.8739 -0.9409 -4.2895
    -2.2707 0.0989 -0.0227 -0.6420 -1.1731 2.9353 3.1575

     | Show Table
    DownLoad: CSV

    Example 2. In this example, we consider Eq (1.1) with the conditions as follows:

    {λ(t)=0,μ(t)=P2t3,φ(x)=0,f(x,t)=Γ(4)Γ(4α)t3αx2+2c2t3,

    in this situation, the exact solution is

    u(x,t)=x2t3.

    Let P=1, Q=1, c=1 and 0<α<1. In our proposed neural network method, the parameters are as follows: N=7, Nm=5, Nn=5. We set the learning rate η and accuracy threshold ϵ to 1×104 and 1×107, respectively. Through 20000 iterations, we obtain the neural network solution of Example 2 for different fractional orders α=0.3,0.5,0.7 and 0.9. Figure 5 give a comparison of neural network solutions and exact solutions for different fractional orders. In Table 3, we list the numerical solutions and absolute errors of several test points under different α. Next, we subdivided the computational grid into Nm×Nn=21×21 and tested our neural network solution on Figure 6. In Figure 7, we plot the variation of the error for different fractional order. Finally, the weights of neurons after iteration are listed in Table 4.

    Figure 5.  The exact solution and the approximate solution for Example 2 with k=20000. Red: exact solution, blue: neural network solution.
    Table 3.  The numerical solutions and absolute errors for Example 2 at α=0.3,0.5,0.7 and 0.9 when k=20000.
    0.3 0.5 0.7 0.9
    solution error solution error solution error solution error
    (0,0) -0.0016 0.0016 -0.0025 0.0025 -0.0023 0.0023 -0.0001 0.0001
    (0.2,0.2) 0.0061 0.0058 0.0035 0.0032 -0.0064 0.0067 -0.0076 0.0080
    (0.4,0.4) 0.0188 0.0085 0.0088 0.0014 0.0014 0.0088 0.0001 0.0101
    (0.6,0.6) 0.0816 0.0039 0.0656 0.0122 0.0740 0.0037 0.0705 0.0073
    (0.8,0.8) 0.3148 0.0128 0.3071 0.0206 0.3260 0.0017 0.3144 0.0133
    (1,1) 1.0011 0.0011 1.0055 0.0055 1.0020 0.0020 1.0026 0.0026

     | Show Table
    DownLoad: CSV
    Figure 6.  The exact solution and the neural network solution on the gird of Nm×Nn=21×21 for Example 2 with k=20000. Red: exact solution, blue: neural network solution.
    Figure 7.  Error decreasing trend with regard to iterations where maximum training time K=20000.
    Table 4.  The weights of the proposed neural network method for Example 2 at α=0.3,0.5,0.7 and 0.9 when k=20000.
    Fractional order θki(i+1)2+j+1,k=20000
    0.3 -0.0016 -0.0002 0.0571 0.1500 -0.0567 -0.2120 -0.2402 0.0174 0.0459 0.1907 0.2427 -0.2256
    0.4118 0.1506 0.1091 -0.4905 0.1521 0.3623 0.3238 -0.5528 -0.3150 0.0692 0.2436 0.3069
    -0.0435 -0.1723 0.0865 0.3350 0.2728 -0.1388 -0.1348 -0.1450 0.2798 -0.1127 0.1988 -0.1630
    0.5 -0.0025 0.0238 0.0540 0.0662 -0.2532 -0.1620 -0.1016 0.1590 0.4392 0.1513 -0.3655 -0.0698
    -0.1684 -0.0452 0.0068 0.3707 0.2070 0.1868 0.5117 -0.4333 0.0094 0.1612 0.1791 0.2668
    0.1049 -0.2033 0.3859 -0.2370 -0.1491 -0.2049 0.2384 -0.3002 -0.0007 0.1888 -0.1895 0.1808
    0.7 -0.0023 -0.0265 0.0230 0.1398 -0.2252 -0.0691 -0.0352 0.0130 -0.0024 0.2792 -0.1484 0.3733
    0.4558 0.2513 -0.4968 -0.1691 0.1646 0.0865 -0.1089 0.0741 0.1482 0.0129 0.1104 -0.1779
    0.0332 0.0814 -0.3102 0.2354 0.2346 -0.3240 0.4362 -0.2617 0.4132 -0.2044 0.1158 -0.1176
    0.9 -0.0001 -0.0464 0.0058 0.0056 -0.0356 0.0137 0.0867 -0.1687 0.2562 -0.1507 -0.0945 0.3658
    0.4054 -0.3842 0.3073 0.2090 -0.2674 -0.0495 0.5397 0.0868 -0.1517 -0.0252 -0.2444 0.1457
    -0.1503 -0.0782 -0.1682 -0.1482 -0.1317 0.4388 -0.0699 0.4540 -0.1167 -0.2030 0.2436 0.1234

     | Show Table
    DownLoad: CSV

    Example 3. Let P=1, Q=1, c=1 and 1<α<2, we consider Eq (1.2) with λ(t)=0, μ(t)=P2t3, φ(x)=0, ψ(x)=0 and f(x,t)=Γ(4)Γ(4α)t3αx2+2c2t3. Through calculation, we can gain the exact solution u(x,t)=x2t3.

    In our proposrd method, we set N=7 and sample gird Nm×Nn=5×5. The learning rate η and accuracy threshold ϵ are set to 1×104 and 1×107, respectively. With 20000 iterations, we acquire a good neural network solution. Figure 8 show a comparison of the neural network solution and the exact solution on sampling grid, their comparison on denser grid Nm×Nn=21×21 and the change of the error value with the number of iterations, respectively. In Table 5, we list the numerical solutions and absolute errors of several test points for α=1.9. Also, Table 6 lists the weights of the proposed neural network in this example. The results illustrate that the proposed neural network method is also very effective for seeking the solution of the time-fractional diffusion-wave equations. (On account of too many graphs, we only list the situation of α=1.9 here.)

    Figure 8.  The exact solution and the neural network solution for Example 3 at α=1.9 with k=20000. (a) Comparison of the neural network solution and the exact solution on sampling grid. (b) comparison on denser grid Nm×Nn=21×21. (c) the change of the error value with the number of iterations. Red: exact solution, blue: neural network solution.
    Table 5.  The numerical solutions and absolute errors for Example 3 at α=1.9 when k=20000.
    Test point Numerical solution Absolute error Test point Numerical solution Absolute error
    (0,0) 0.0015 0.0015 (0.5,0.5) 0.0317 0.0004
    (0.1,0.1) 0.0013 0.0013 (0.6,0.6) 0.0773 0.0004
    (0.2,0.2) 0.0011 0.0008 (0.7,0.7) 0.1651 0.0030
    (0.3,0.3) 0.0028 0.0004 (0.8,0.8) 0.3209 0.0068
    (0.4,0.4) 0.0106 0.0004 (1,1) 1.0015 0.0015

     | Show Table
    DownLoad: CSV
    Table 6.  The weights of the proposed neural network method for Example 3 at α=1.9 with k=20000.
    Fractional order θki(i+1)2+j+1,k=20000
    1.9 0.0015 0.0231 -0.0217 -0.0077 0.01778 -0.0464
    -0.2435 0.4708 -0.4991 0.2667 0.3001 -0.3739
    0.0398 0.5381 -0.0875 0.0560 0.1424 0.2924
    0.3189 -0.3403 -0.1458 -0.2037 -0.0313 0.4367
    0.2250 0.0753 -0.1526 -0.2626 0.07791 -0.0908
    0.2382 -0.4721 0.3075 -0.5056 0.3633 0.2947

     | Show Table
    DownLoad: CSV

    Example 4. In this example, we consider Eq (1.2) with the conditions as follows:

    {λ(t)=μ(t)=0,φ(x)=sin(πxP),f(x,t)=sin(πxP)cos1,1α(t)(cπP)2sin(πxP)cos(t),

    in this situation, the exact solution is

    u(x,t)=sin(πxP)cos(t).

    Let P=1, Q=1, c=1 and 1<α<2, other parameters are listed as follows: N=5, Nm=Nn=5, η=6×105 and ϵ=1×107. Similar to Example 3, we only show the case of α=1.5 by 40000 iterations. In Figure 9, we show the comparison of the neural network solution and the exact solution on sampling grid Nm×Nn=5×5, their comparison on denser grid Nm×Nn=21×21 and the change of error value with the number of iterations, respectively. In addition, we list the numerical solutions and absolute errors of several test points in Table 7. The weights of the neural network in this example are listed in Table 8.

    Figure 9.  The exact solution and the neural network solution for Example 4 at α=1.5 with k=40000. (a)Comparison of the neural network solution and the exact solution on sampling grid. (b)comparison on denser grid Nm×Nn=21×21. (c) the change of the error value with the number of iterations. Red: exact solution, blue: neural network solution.
    Table 7.  The numerical solutions and absolute errors for Example 4 at α=1.5 when k=40000.
    Test point Numerical solution Absolute error Test point Numerical solution Absolute error
    (0,0) -0.0028 0.0028 (0.5,0.5) 0.8497 0.0279
    (0.1,0.1) 0.3185 0.0110 (0.6,0.6) 0.7633 0.0216
    (0.2,0.2) 0.5758 0.0002 (0.7,0.7) 0.6026 0.0161
    (0.3,0.3) 0.7558 0.0171 (0.8,0.8) 0.3929 0.0166
    (0.4,0.4) 0.8484 0.0276 (1,1) -0.0011 0.0011

     | Show Table
    DownLoad: CSV
    Table 8.  The weights of the proposed neural network method for Example 4 at α=1.5 with k=40000.
    Fractional order θki(i+1)2+j+1,k=40000
    1.5 -0.0028 0.0963 3.3865 -0.2123 -0.8028 -1.4247 0.0071
    -0.2903 0.3029 -2.5919 0.1321 -0.3393 0.4284 0.6957
    -1.0843 -0.0163 -0.1139 0.2444 0.2087 -0.3435 1.7188

     | Show Table
    DownLoad: CSV

    In this paper, different power terms of Taylor polynomial are used as neurons of the proposed artificial neural network to solve the time-fractional diffusion and diffusion-wave equations. The initial weight of the neural network is given randomly according to the Gaussian distribution, and what is more the weight in each iteration is updated by the error function. In order to balance the contribution of initial error term, boundary error term and internal error term in the error function, we weight different error terms, and give an adaptive weight adjustment algorithm. The trained network is found to be effective by testing other points within the solution region. In addition, the value of the error function decreases as the number of iterations increases. This illustrates that our method is very effective for solving time-fractional diffusion and diffusion-wave equations. In the future work, we will consider to extend our methods for solving other time-fractional partial differential equations, such as the time-fractional mixed diffusion and diffusion-wave equations [28], time-fractional diffusion equations with a time-invariant type variable order [29] and the generalized time space fractional diffusion equations with variable coefficients [30].

    The authors express their sincere thanks to the anonymous reviewer for his/her careful reading of the paper, giving valuable comments and suggestions. It is their contributions that greatly improve the paper. The authors also thank the editors for their kind help.

    This work was supported by the National Natural Science Foundation of China (Nos. 11801452, 11701456); Shaanxi Fundamental Science Research Project for Mathematics and Physics (Grant No. 22JSQ020); Fundamental Research Project of Natural Science in Shaanxi Province General Project (Youth) (Nos.2019JQ-415, 2019JQ-196); Chinese Universities Scientific Funds (Nos. 2452022104, 2452022105, 2452022372); and Innovation and Entrepreneurship Training Program for College Students of Northwest A & F University (202110712186, 202210712048, S202210712377).

    The authors declare there is no conflict of interest.



    [1] Y. A. Rossikhin, M. V. Shitikova, Applications of fractional calculus to dynamic problems of linear and nonlinear hereditary mechanics of solids, Appl. Mech. Rev., 50 (1997), 15–67. https://doi.org/10.1115/1.3101682 doi: 10.1115/1.3101682
    [2] D. del-Castillo-Negrete, B. A. Carreras, V. E. Lynch, Front dynamics in reaction-diffusion systems with Lévy flights: a fractional diffusion approach, Phys. Rev. Lett., 91 (2003), 018302. https://doi.org/10.1103/PhysRevLett.91.018302 doi: 10.1103/PhysRevLett.91.018302
    [3] A. Dechant, E. Lutz, Anomalous spatial diffusion and multifractality in optical lattices, Phys. Rev. Lett., 108 (2012), 230601. https://doi.org/10.1103/PhysRevLett.108.230601 doi: 10.1103/PhysRevLett.108.230601
    [4] M. Giona, H. E. Roman, Fractional diffusion equation for transport phenomena in random media, Phys. A, 185 (1992), 87–97. https://doi.org/10.1016/0378-4371(92)90441-R doi: 10.1016/0378-4371(92)90441-R
    [5] F. Mainardi, Fractional diffusive waves in viscoelastic solids, In: J. I. Wegner, F. R. Norwood, eds., Nonlinear Waves in Solids., Fairfield: ASME/AMR, 1995, 93–97.
    [6] W. H. Luo, T. Z. Huang, G. C. Wu, X. M. Gu, Quadratic spline collocation method for the time fractional subdiffusion equation, Appl. Math. Comput., 276 (2016), 252–265. https://doi.org/10.1016/j.amc.2015.12.020 doi: 10.1016/j.amc.2015.12.020
    [7] W. H. Luo, C. Li, T. Z. Huang, X. M. Gu, G. C. Wu, A high-order accurate numerical scheme for the Caputo derivative with applications to fractional diffusion problems, Numer. Funct. Anal. Optim., 39 (2018), 600–622. https://doi.org/10.1080/01630563.2017.1402346 doi: 10.1080/01630563.2017.1402346
    [8] X. M. Gu, S. L. Wu, A parallel-in-time iterative algorithm for Volterra partial integro-differential problems with weakly singular kernel, J. Comput. Phys., 417 (2020), 109576. https://doi.org/10.1016/j.jcp.2020.109576 doi: 10.1016/j.jcp.2020.109576
    [9] Y. L. Zhao, X. M. Gu, A. Ostermann, A preconditioning technique for an all-at-once system from Volterra subdiffusion equations with graded time steps, J. Sci. Comput., 88 (2021), 11. https://doi.org/10.1007/s10915-021-01527-7 doi: 10.1007/s10915-021-01527-7
    [10] Z. Liu, A. Cheng, X. Li, A novel finite difference discrete scheme for the time fractional diffusion-wave equation, Appl. Numer. Math., 134 (2018), 17–30. https://doi.org/10.1016/j.apnum.2018.07.001 doi: 10.1016/j.apnum.2018.07.001
    [11] R. Du, Y. Yan, Z. Liang, A high-order scheme to approximate the Caputo fractional derivative and its application to solve the fractional diffusion wave equation, J. Comput. Phys., 376 (2019), 1312–1330. https://doi.org/10.1016/j.jcp.2018.10.011 doi: 10.1016/j.jcp.2018.10.011
    [12] X. Li, S. Li, A fast element-free Galerkin method for the fractional diffusion-wave equation, Appl. Math. Lett., 122 (2021), 107529. https://doi.org/10.1016/j.aml.2021.107529 doi: 10.1016/j.aml.2021.107529
    [13] M. H. Heydari, M. R. Hooshmandasl, F. M. M. Ghaini, C. Cattani, Wavelets method for the time fractional diffusion-wave equation, Phys. Lett. A, 379 (2015), 71–76. https://doi.org/10.1016/j.physleta.2014.11.012 doi: 10.1016/j.physleta.2014.11.012
    [14] M. H. Heydari, Z. Avazzadeh, M. F. Haromi, A wavelet approach for solving multi-term variable-order time fractional diffusion-wave equation, Appl. Math. Comput., 341 (2019), 215–228. https://doi.org/10.1016/j.amc.2018.08.034 doi: 10.1016/j.amc.2018.08.034
    [15] A. Kumar, A. Bhardwaj, B. V. R. Kumar, A meshless local collocation method for time fractional diffusion wave equation, Comput. Math. Appl., 78 (2019), 1851–1861. https://doi.org/10.1016/j.camwa.2019.03.027 doi: 10.1016/j.camwa.2019.03.027
    [16] H. Qu, Cosine radial basis function neural networks for solving fractional differential equations, Adv. Appl. Math. Mech., 9 (2017), 667–679. https://doi.org/10.4208/aamm.2015.m909 doi: 10.4208/aamm.2015.m909
    [17] F. Rostami, A. Jafarian, A new artificial neural network structure for solving high-order linear fractional differential equations, Int. J. Comput. Math., 95 (2018), 528–539. https://doi.org/10.1080/00207160.2017.1291932 doi: 10.1080/00207160.2017.1291932
    [18] F. B. Rizaner, A. Rizaner, Approximate solutions of initial value problems for ordinary differential equations using radial basis function networks, Neural Process. Lett., 48 (2018), 1063–1071. https://doi.org/10.1007/s11063-017-9761-9 doi: 10.1007/s11063-017-9761-9
    [19] A. Jafarian, S. M. Nia, A. K. Golmankhaneh, B. Baleanu, On artificial neural networks approach with new cost functions, Appl. Math. Comput., 339 (2018), 546–555. https://doi.org/10.1016/j.amc.2018.07.053 doi: 10.1016/j.amc.2018.07.053
    [20] A. H. Hadian-Rasanan, D. Rahmati, S. Gorgin, K. Parand, A single layer fractional orthogonal neural network for solving various types of Lane-Emden equation, New Astron., 75 (2020), 101307. https://doi.org/10.1016/j.newast.2019.101307 doi: 10.1016/j.newast.2019.101307
    [21] H. Qu, X. Liu, Z. She, Neural network method for fractional-order partial differential equations, Neurocomputing, 414 (2020), 225–237. https://doi.org/10.1016/j.neucom.2020.07.063 doi: 10.1016/j.neucom.2020.07.063
    [22] Y. Ye, H. Fan, Y. Li, X. Liu, H. Zhang, Deep neural network methods for solving forward and inverse problems of time fractional diffusion equations with conformable derivative, Neurocomputing, 509 (2022), 177–192. https://doi.org/10.1016/j.neucom.2022.08.030 doi: 10.1016/j.neucom.2022.08.030
    [23] A. A. Kilbas, H. M. Srivastava, J. J. Trujillo, Theory and Applications of Fractional Differential Equations, New York: Elsevier, 2006.
    [24] R. Garrappa, The Mittag-Leffler function, MATLAB Central File Exchange. Available from: https://www.mathworks.com/matlabcentral/fileexchange/48154-the-mittag-leffler-function
    [25] M. H. Hassoun, Fundamentals of artificial neural networks, Cambridge: MIT Press, 1995.
    [26] L. C. Evans, Partial Differential Equations, 2nd edition, Providence: American Mathematical Society, 2010.
    [27] S. Wang, Y. Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM J. Sci. Comput., 43 (2021), A3055–A3081. https://doi.org/10.1137/20M1318043 doi: 10.1137/20M1318043
    [28] J. Shen, X. M. Gu, Two finite difference methods based on an H2N2 interpolation for two-dimensional time fractional mixed diffusion and diffusion-wave equations, Discrete Contin. Dyn. Syst. Ser. B, 27 (2022), 1179–1207. https://doi.org/10.3934/dcdsb.2021086 doi: 10.3934/dcdsb.2021086
    [29] X. M. Gu, H. W. Sun, Y. L. Zhao, X. Zheng, An implicit difference scheme for time-fractional diffusion equations with a time-invariant type variable order, Appl. Math. Lett., 120 (2021), 107270. https://doi.org/10.1016/j.aml.2021.107270 doi: 10.1016/j.aml.2021.107270
    [30] X. M. Gu, T. Z. Huang, Y. L. Zhao, P. Lyu, B. Carpentieri, A fast implicit difference scheme for solving the generalized time-space fractional diffusion equations with variable coefficients, Numer Methods Partial Differ Equ, 37 (2021), 1136–1162. https://doi.org/10.1002/num.22571 doi: 10.1002/num.22571
  • This article has been cited by:

    1. Zehui Ma, Rahmatjan Imin, A meshless particle method for solving time-fractional wave equations, 2024, 11, 2196-4378, 2157, 10.1007/s40571-024-00771-6
    2. Lin Li, Zhong Chen, Hong Du, Wei Jiang, Biao Zhang, A meshless approach based on fractional interpolation theory and improved neural network bases for solving non-smooth solution of 2D fractional reaction–diffusion equation with distributed order, 2024, 138, 10075704, 108245, 10.1016/j.cnsns.2024.108245
    3. Fangfang Wu, Jia Tong, Shaobo Qu, Yingying Wang, Numerical solution of generalized time-fractional wave equation based on lattice Boltzmann method, 2025, 100, 0031-8949, 015281, 10.1088/1402-4896/ad9e36
  • Reader Comments
  • © 2023 the Author(s), licensee AIMS Press. This is an open access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0)
通讯作者: 陈斌, bchen63@163.com
  • 1. 

    沈阳化工大学材料科学与工程学院 沈阳 110142

  1. 本站搜索
  2. 百度学术搜索
  3. 万方数据库搜索
  4. CNKI搜索

Metrics

Article views(1791) PDF downloads(86) Cited by(3)

Figures and Tables

Figures(9)  /  Tables(8)

Other Articles By Authors

/

DownLoad:  Full-Size Img  PowerPoint
Return
Return

Catalog