Shiqiang Yan &Q.W.Ma &Jinghua Wang

Abstract In the Lagrangian meshless (particle) methods, such as the smoothed particle hydrodynamics (SPH), moving particle semi-implicit (MPS) method and meshless local Petrov-Galerkin method based on Rankine source solution(MLPG_R), the Laplacian discretisation is often required in order to solve the governing equations and/or estimate physical quantities (such as the viscous stresses). In some meshless applications, the Laplacians are also needed as stabilisation operators to enhance the pressure calculation. The particles in the Lagrangian methods move following the material velocity, yielding a disordered (random) particle distribution even though they may be distributed uniformly in the initial state. Different schemes have been developed for a direct estimation of second derivatives using finite difference, kernel integrations and weighted/moving least square method. Some of the schemes suffer from a poor convergent rate. Some have a better convergent rate but require inversions of high order matrices, yielding high computational costs. This paper presents a quadric semi-analytical finite-difference interpolation (QSFDI) scheme,which can achieve the same degree of the convergent rate as the best schemes available to date but requires the inversion of significant lower-order matrices, i.e. 3 × 3 for 3D cases, compared with 6 × 6 or 10 × 10 in the schemes with the best convergent rate. Systematic patch tests have been carried out for either estimating the Laplacian of given functions or solving Poisson’s equations. The convergence, accuracy and robustness of the present schemes are compared with the existing schemes. It will show that the present scheme requires considerably less computational time to achieve the same accuracy as the best schemes available in literatures, particularly for estimating the Laplacian of given functions.

Keywords Laplaciandiscretisation .Lagrangianmeshlessmethods .QSFDI .Random/disorderedparticledistribution .Poisson’s equation .Patch tests

1 Introduction

In the Lagrangian meshless/particle methods,for example,the smoothed particle hydrodynamics (SPH, e.g. Monaghan 1994; Shao and Lo 2003; Shao et al. 2006; Khayyer et al.2008;Lind et al.2012),moving particle semi-implicit method(MPS, e.g. Koshizuka and Oka 1996; Gotoh and Khayyer 2016; Khayyer and Gotoh 2010),the meshless local Petrov-Galerkin method (MLPG, e.g., Ma 2005a, b; Zhou and Ma 2010), interpolating element-free Galerkin method(Abbaszadeh and Dehghan 2019; Dehghan and Abbaszadeh 2018,2019),the computational domain is represented by particles and the governing equations with associated boundary conditions are discretised to form a linear algebraic equation system,which leads to the approximation of physical quantities at particle locations. These methods become more popular,attributing to their superiority over the conventional mesh-based methods, such as the finite element method and finite volume method, in dealing with various engineering problems with large deformations,such as the breaking wave impact on offshore structures, for which the mesh-based methods may suffer from significant mesh distortions and/or numerical diffusions.

One key technique for the Lagrangian meshless methods is to numerically formulate and discretise the Laplacian operator,which is required mainly for three purposes.The first one is to discretise the governing equations, involving secondorder partial differential terms or Laplacian operators, e.g.the partial differential equation for the heat conduction and thermal diffusion (Chen et al. 1999; Schwaiger 2008), the pressure Poisson’s equation employed by the projectionbased/fractional step methods for solving the Navier-Stokes(NS) equations to deal with fluid-structure interaction problems(Shao and Lo 2003;Ma et al.2016)and the Helmholtz equation widely used in wave and diffusion problems. The second one is to numerically estimate specific physical quantities,which are expressed by the Laplacian of others,e.g.,the viscous stress in the NS equation (Zheng et al. 2018). The third one is to discretise the Laplacian utilised as stabilisation operators to enhance the pressure calculation in various applications(Khayyer and Gotoh 2010;Khayyer and Gotoh 2012;Ikari et al. 2015). In meshless methods, the Laplacian of a function at a specific location is numerically formulated in terms of discretised function values at surrounding particles within the support domain using finite difference,kernel integration,moving least squares(MLS)or weighted least squares(WLS)algorithms.Therefore,the accuracy,convergence and computational efficiency of the schemes are significantly influenced by the particle distribution. In Eulerian meshless methods for steady problems (e.g., Lind and Stansby 2016),the particles are fixed and may be placed uniformly and regularly. For such problems, high-order finite difference schemes based on a uniform particle distribution can be directly applied to formulate the Laplacian operator.However,in the Lagrangian meshless methods for unsteady problems,the particles move at the material velocity and,consequently,their distribution may be highly disordered even though they may be placed regularly and uniformly in the initial state.Such disorderliness/randomness of the particle distribution considerably downgrades the schemes based on a uniform particle distribution.

For a random/disordered particle distribution,efforts have been devoted to the reviews of various schemes,e.g.on dealing with the viscous term in SPH(Zheng et al.2018)and on solving pressure Poisson’s equation (PPE) involved in projection-based meshless methods (Ma et al. 2016). It has been observed that some schemes (classified as type 1 by Ma et al. 2016),e.g. Cummins and Rudman (1999); Lo and Shao (2002); Lee et al. (2008); Xu et al. (2009); Hu and Adams(2007);Gotoh et al.(2014);and Khayyer and Gotoh(2010,2012),converge at a rate less than first order for estimating the Laplacian of a given function,although they do not need inversions of any matrices and thus have relatively low computational costs.Their performances may be improved by reducing the randomness of the particle distribution, e.g.,using the particle shifting scheme proposed by Lind et al.(2012),or by introducing error correction and compensating terms, e.g. Oger et al. (2007) and Ikari et al. (2015), which requires the inversion of matrices. Alternatively, high-order Laplacian discretisation schemes have also been developed for a random/distorted particle distribution. These include the CSPM proposed by Chen et al.(1999),a scheme proposed by Fatehi and Manzari(2011),which was developed from the Brookshaw’s scheme (Brookshaw 1985) by introducing an error compensation, the LSMPS developed by Tamai and Koshizuka (2014) using the MLS algorithm, and LP-MPS proposed by Tamai et al. (2017) using the WLS algorithm.These schemes are classified as type 3 schemes by Ma et al.(2016).Patch tests by Schwaiger(2008),Zheng et al.(2014)and Tamai et al.(2017)have shown that the CSPM,LP-MPS and quadric LSMPS have a higher convergent rate,compared with the type 1 schemes.However,formulating these schemes requires a significant computational cost on inversing matrices at all particle locations and at every time step of the transient simulations.For each particle in 3D problems,the CSPM and LP-MPS need to inverse two matrices with sizes of 6×6 and 3×3,respectively;the quadric LSMPS needs to inverse a matrix with a size of 10 × 10. To overcome this problem,Schwaiger(2008)proposed a CSPH scheme,which is based on the CSPM but reduces the sizes of inversed matrices to two 3×3 for 3D problems,through ignoring the cross-derivative terms of the 2nd derivatives(thus downgrading the accuracy).

In this paper, a quadric semi-analytical finite-difference interpolation scheme (referred to as QSFDI) for numerically formulating the Laplacian operator is developed based on the principle of the linear SFDI(Ma 2008).Its consistency,accuracy and convergence property are similar to the quadric LSMPS and LP-MPS, for randomly distributed particles,whereas the sizes of the matrices to be inversed are considerably reduced to 3×3 for 3D problems,compared with 6×6 in the quadric LSMPS and LP-MPS. The performance of the present scheme will be assessed by systematic patch tests considering both directly estimating the Laplacian of specific functions and solving Poisson’s equation.

2 Mathematical Formulation

Ma (2008) and Ma et al. (2016) developed the linear SFDI,which was based on Taylor’s expansion with a leading truncation term of 2nd derivatives, for numerical interpolations and gradient estimations. The patch test by Ma (2008)suggested that the linear SFDI is superior over the linear MLS when the particles are randomly distributed.The principle of the SFDI is extended here to derive the interpolation,gradient estimation and Laplacian discretisation schemes with a quadric accuracy.

For each particle j at xj, which locates inside the support domain of Point xI,a function p can be expressed as Taylor’s expansion,

Compared with the leading truncation error, it is equally important to look at the number and sizes of the matrices to be inversed in order to formulate the Laplacian operator.For the QSFDI,to approximate the gradient,the 2nd derivatives and the Laplacian of p(x),three matrices,i.e.M2c,I,M2s,Iand M1q,I,need to be inversed.For 3D problems,they all have sizes of 3×3,whereas for 2D problems,the sizes of M2s,Iand M1q,Iare 2 × 2 and M2c,Iis a scalar. The sizes of matrices to be inversed for formulating the Laplacian in the QSFDI are slightly higher than the Schwaiger’s scheme (e.g. CSPH2Γ)for 3D problems but considerably lower than the LP-MPS,the quadric LSMPS and the CSPM,for which the number and the sizes of matrices to be inversed are summarised in Table 1.As mentioned above, an improved scheme for CSPM, i.e.,ICSPM is also introduced.The number and sizes of the matrices to be inversed in the ICSPM are the same as the original CSPM, although the accuracy of the former is one order higher than the latter for a random particle distribution.

It is well known that the CPU time spent on the inversion of a matrix with size of M×M with M being 2 to 10 is proportional to M3.The CPU time spent on matrix inversions for 2D problems may be indicated by～2×23for CSPH2Γ and QSFDI,～23+ 33for CSPM, LP-MPS and ICSPM, ～ 63for quadric LSMPS.The CPU time by the QSFDI on matrix inversions is approximately 45.7%of those by the LP-MPS or ICSPM,and 7.4% of that by the quadric LSMPS for 2D problems. The corresponding figures for 3D problems are 40% and 8.1%,respectively. Nevertheless, different schemes require different CPU time on formulating the matrices;the overall robustness of these schemes will be investigated in the following patch tests.

3 Patch Test

To quantify the accuracy,convergence and robustness of the QSFDI and ICSPM, patch tests are performed using various cases.As indicated above,some of the main purposes for the Laplacian discretisation are (1) to find physical quantities,which may be expressed as the Laplacian of others, and (2)to discretise the Poisson’s equations. For the former, the Laplacians of various specified functions f(x,y), which are frequently used in literatures are directly estimated. For the latter,the Poisson’s equation are considered:

Table 1 Summary of the approaches for Laplacian discretisation

where r=djIis used as the kernel function of ICSPM,CSPM and CSPH2Γ, and the weighting function of other schemes.By using Eq.(9),the radius of the support domain is 3h.The effect of the kernel/weighting functions will not be investigated and discussed in the future.

3.1 Estimating Laplacian of Specified Functions

Following Schwaiger(2008),the first group of the test functions used is

The computational domain is a unit square with 2 ≤x ≤3 and 2 ≤y ≤3. The particles are initially generated using a uniform spacing,i.e.Δx=Δy=s0.To reflect the randomness of the particle distribution,a random shift with Δx′=Kδs0and Δy′=Kδs0is applied to all particles,where K is a scale factor and δ is a random number between ?1 and 1. In the patch tests, K is in the range of 0.1 to 0.8. The Laplacian of the function p at any particle i,??2pi?,are estimated by different schemes summarised in Table 1.The relative error is estimated in the same way as in Schwaiger (2008) for the sake of comparison,given by

where ?2piis the analytical value of the Laplacian of the function p at Particle i; N is the total number of particles at which the Laplacian are estimated.It shall be pointed out that all existing schemes suffer from a downgraded accuracy when being applied to the particles near the boundaries due to the fact that the support domain is not full (Chen et al. 1999;Schwaiger 2008)or the neighbouring particles are mainly distributed in a quadrant or a half of the support domain of the particles. For simplicity, as well as being consistent with Schwaiger(2008),only the inner particles within a region of 2.25

Figure 1 compares the relative errors for estimating the Laplacian of p(x,y)=x6y6using different schemes with a randomness specified by K = 0.4. As can be seen, the LPMPS, quadric LSMPS and the present QSFDI result in a linear rate of reduction of the relative error as s0decreases, while the CSPM and CSPH2Γ do not seem to be convergent as s0decreases for a constant ratio of h/s0,conforming to the patch tests in Schwaiger (2008).Compared with CSPM, the relative error of the ICSPM not only is lower but also reduces linearly as s0decreases.It is also observed from Figure 1 that the present QSFDI leads to almost identical results as the LP-MPS, which are more accurate than other schemes, for all values of h/s0and s0applied in the patch test. As analysed in Section 2,the QSFDI demands less computational efforts on matrix inversions than the LP-MPS; the overall CPU time spent by the former is expected to be shorter than the latter.This is confirmed by Figure 2, which shows that the average CPU time spent by the QSFDI is approximately 20% less than that by the LP-MPS for achieving results with the same accuracy. In addition, Figure 3 displays the relative errors of the Laplacian discretisation against CPU times by different schemes in the cases with K = 0.4 and different values of h/s0. For convenience, the CPU times are scaled by the reference duration, TRef, which is the CPU time spent by the CSPH2Γ with s0= 0.1 and h =0.75s0. Both Figures 2 and 3 confirm that the QSFDI requires considerably shorter CPU time than all other schemes for achieving a specific accuracy of estimating the Laplacian of function p(x,y)=x6y6.

Figure 1 Relative error of Laplacian discretisation vs mean particle spacing s0 for estimating Laplacian of p(x, y) = x6y6 (K = 0.4; error estimation domain 2.25 < x < 2.75, 2.25 < y < 2.75). a h = 0.75s0.b h=0.9s0.c h=1.2s0

Figure 2 Ratio of the CPU time spent by the LP-MPS and that by the QSFDI on estimating Laplacian of p(x,y)=x6y6(K=0.4)

Similar behaviours to Figures 1 and 2 have been observed in the cases with other values of K. Some results are shown in Figures 5 and 6 for demonstration.To save the space, not all the results with different ratios of h/s0are presented. In Figure 5, h/s0= 1.2 is used by the CSPM, CSPH2Γ and ICSPM; h/s0= 0.75 is adopted by others. In such a way, the CSPM, CSPH2Γ,ICSPM and the LSMPS are expected to have the best robustness and the accuracy compared to other values of h/s0, whereas the results of the QSFDI and the LP-MPS adopting h/s0= 0.75 are worse than the corresponding results with a greater h/s0, as demonstrated by Figure 4.Again, it is clearly seen that the QSFDI and the LPMPS lead to the highest accuracy and convergence properties compared to others. The comparison of the CPU time spent by the LP-MPS and the QSFDI in Figure 6 again confirms the superiority of the former in terms of saving CPU time for the cases with a different randomness.

Considering the fact that most of analytical formulations used in the engineering problems may be represented by hyperbolic/exponential and/or trigonometric functions (e.g.the velocity potential associated with a linear propagation wave and the temperature in 2D heat conduction problems),the exponential function applied by Tamai et al.(2017)

Figure 3 Relative error of Laplacian discretisation vs CPU time for estimating Laplacian of p(x,y)=x6y6 (K = 0.4; TRef is the CPU time spent by CSPH2Γ with s0 = 0.1 and h = 0.75s0; error estimation domain 2.25 < x < 2.75, 2.25 < y < 2.75). a h = 0.75s0. b h = 0.9s0.c h=1.2s0

are also considered.In these tests,the computational domain is taken as a unit square with 0 ≤x ≤1;0 ≤y ≤1.The particle generation is the same as that used in the first test case,i.e.a random distribution with different values of K. The relative error in the tests is defined by

Figure 4 Relative error of Laplacian discretisation vs CPU time for estimating Laplacian of p(x,y)=x6y6 in the cases with different ratios h/s0(K=0.4;TRef is the CPU time spent by CSPH2Γ with s0=0.1 and h= 0.75 s0;error estimation domain 2.25

where N is the number of particles for error estimations to be consistent with the references. Similar to the cases shown in Figures 1, 2, and 3, the relative error at inner particles away from boundaries, i.e. 0.25 ≤x ≤0.75; 0.25 ≤y ≤0.75, is assessed. In addition, to reflect the overall performances of different schemes at both the inner particles and boundary particles, the relative error estimated by considering all particles are also assessed.

Figure 5 Relative error for estimating Laplacian of p(x,y)=x6y6 for K = 0.6 and K = 0.8 (h = 1.2s0 for CSPM, CSPH2Γ and ICSPM,h = 0.75s0 for other schemes; the slopes of the dotted line is 1; error estimation domain 2.25

Figure 6 Ratio of the CPU time spent by the LP-MPS and that by the QSFDI on estimating Laplacian of p(x,y)=x6y6(h=0.75s0)for the cases with different randomness of particle distribution

Figure 7 displays the relative errors for estimating Laplacians of Eqs. (12) and (13) in the cases with K= 0.4. For clarity, the corresponding results by ICSPM are not shown. From Figure 7a, c, in which only inner particles are taken into account when estimating the relative error, it is observed that all schemes converge at a rate between linear and quadric for relatively coarse particle resolutions, i.e. s0> 0.01; thereafter, the relative errors of the CSPM and CSPH2Γ seem not to be reduced, whereas the QSFDI, the LP-MPS and quadric LS-MPS converge at a rate slightly higher than a linear rate. As expected, for a specific particle resolution, the relative errors of all the schemes considering all particles (Figure 7b, d) including these on boundaries are relatively higher than the corresponding results considering the inner particles only. The CSPM and CSPH2Γ converge at a rate much lower than a linear rate (the mean slopes of the curves for CSPM and CSPH2Γ are about 1/0.5). The QSFDI, the LP-MPS and quadric LSMPS converge at a rate with the mean slopes of approximately 1/1.5.

The same conclusions are achieved in the cases with other randomness, e.g.Figure 8, which displays the relative errors considering all particles for estimating Laplacian of Eqs.(12)and (13), respectively, obtained by using K = 0.8.Examination of the corresponding robustness is also carried out for these cases.The CPU time for different accuracy corresponding to the results of Figure 8 is illustrated in Figure 9.Once again,the superiority of the QSFDI over other schemes in terms of the accuracy and computational costs is clearly observed as in Figures 3 and 6.

3.2 Finding solutions of Poisson’s equation

Another purpose of the Laplacian discretisation is to discretise the governing equation,e.g.the Poisson’s equation,to find its numerical solution. Relevant patch tests presented in thissection will use a computational domain of a unit square(0 ≤x ≤1,0 ≤y ≤1),the same as that for Figures 7,8,and 9.In this test,the Poisson’s equation defined by Eq.(8),of which(R.H.S)are specified by the Laplacians of a given function.The Dirichlet condition on all boundaries of the domain is applied, i.e. p at all boundary particles are specified to be consistent with(R.H.S).Under this condition,the analytical solution of the Poisson’s equation (Eq. (8)) in the computational domain is the function p(x,y).For example,if Eq.(13)is applied as the function p(x,y),(R.H.S)= ?100 sin(6x+8y)and the values on the boundary at x=1.0 is sin(6+8y).The Poisson’s equation is discretised at all internal particles by using different schemes to form linear algebraic equations,which is solved by the GMRESS solver,resulting in the solutions of p(x,y) at all internal particles.For the results shown below,the control residual adopted by the GMRESS solver is 10?4,which is sufficiently small(comparison with the corresponding results obtained using a control residual of 10?8shows that the difference is smaller than 0.1%).The relative error of the numerical solution against the analytical solution is evaluated by

Figure 7 Relative error for estimating Laplacians of Eq.(12)and(13)(K=0.4;h=1.2s0 for CSPM and CSPH2Γ,h=0.9s0 for other schemes;internal particles located at 2.25

Figure 10 compares the relative errors of numerical solutions to the Poisson’s equations,whose(R.H.S)are given by the Laplacian of Eqs. (12) and (13), respectively, with the moderate randomness of particle distribution (K = 0.4). It is observed that the convergent rates of the LP-MPS, quadric LS-MPS, ICSPM and the QSFDI are quadric for all particle spacing.In contrast,as s0decreases,the relative errors of the CSPM and CSPH2Γ reduces at a quadric rate for relatively coarse resolutions (when the error is large); however, it reduces to a linear or lower rate for finer particle resolutions(when the error becomes acceptably small).The comparison of the relative errors for a specific particle spacing indicates that the QSFDI and the LP-MPS result in the most accurate solutions. The corresponding comparisons of the CPU time are illustrated in Figure 11. Unlike the direct Laplacian discretisation presented in Section 3.1, the CPU time spent on achieving the solutions to the Poisson’s equation is also influenced by the effectiveness of the linear algebraic solver and its pre-conditioner, i.e. the initial value. By using the solver briefed above, the total CPU time spent by the QSFDI is slightly shorter than the LP-MPS but significantly shorter than all other schemes.

Figure 8 Relative error for estimating Laplacians of Eqs.(12)and(13)at all particles (K = 0.8; h = 1.2s0 for CSPM and CSPH2Γ, h = 0.9s0 for other schemes

Figure 9 CPU times for estimating Laplacians of Eq.(12)and(13)(K=0.8;h=1.2s0 for CSPM and CSPH2Γ,h=0.9s0 for other schemes;TRef is the CPU time spent by CSPH2Γ with s0=0.1)

Figure 10 Relative error of solution to Poisson’s equation based on Eq. (12) and Eq. (13) (K = 0.4, h = 1.2s0 for CSPM, CSPH2Γ and ICSPM,h=0.9s0 for other schemes)

Different values of K and h are also used in this investigation.Some results are illustrated in Figures 12 and 13 for K=0.6 and 0.8 respectively, where h = 1.2s0are used for all schemes.For clarity,the corresponding results with the quadric LP-MPS and the ICSPM are not shown.As can be seen,with severer randomness of particle distribution,the relative errors of the CSPM and CSPH2Γ reduce at a rate less than the linear rate as s0decreases, quite different from what has been seen in Figure 10.In contrast,the accuracy and convergent properties of the QSFDI and LP-MPS seem to be insignificantly affected by increasing the randomness (Figures 12a, c and 13a, c).Figures 12b and d and 13b and d reveal that the CPU time by the present QSFDI is generally shorter than all other schemes for achieving satisfactory results,e.g.relative error smaller than 1%.Following the comparison of the robustness of the QSFDI and the LP-MPS in the previous section on estimating Laplacians, the average ratios of the CPU time spent by the LP-MPS and that by the QSFDI are displayed in Figure 14,where h=1.2s0,for finding the solutions to the Poisson’s equation based on Eq.(13).It clearly shows that the CPU time spent by the QSFDI is approximately 5%–10%shorter than the LPMPS, although averagely 20% less CPU time on discretising the Poisson’s equation than that by the QSFDI is recorded in the patch tests in Section 3.1.

Figure 11 CPU times for solving Poisson’s equation based on Eqs.(12)and(13)(K=0.4,h=1.2s0 for CSPM,CSPH2Γ and ICSPM,h=0.9s0 for other schemes)

Figure 12 Relative error of solution to Poisson’s equation based on Eq.(12)and the corresponding CPU time in the cases with different particle randomness(h=1.2s0;TRef is the CPU time spent by CSPH2Γ with s0=0.1 and K=0.6)

4 Conclusions

This paper develops a new scheme called QSFDI, which adopts the same principle of SFDI, to discretise the Laplacian operator for Lagrangian meshless (particle)methods, in which the particles move during the numerical simulation and exhibit a disordered/random distribution.The accuracy and consistency of the QSFDI are similar to the LSMPS and LP-MPS but higher than the CSPM and CSPH for randomly distributed particles.However,the matrices required to be inversed by the QSFDI have smaller sizes than the LSMPS and LP-MPS.For example,for 3D problems,the size of the matrices to be inversed in the QSFDI is 3×3,while it is 6×6 in LP-MPS.

Systematic patch tests considering both directly estimating the Laplacian of specific functions and solving Poisson’s equations are carried out. In these tests, different functions including polynomials, hyperbolic and trigonometric functions,which may represent typical spatial variations of physical quantities in engineering such as the water waves and the thermodynamics, are considered. The particles used in the patch tests are randomly distributed. It is observed that theQSFDI has the same convergent rate as the LP-MPS and quadric LSMPS,which is higher than that of the CSPM and CSPH in all the cases studied. It is also observed that the QSFDI requires considerably less computational time than all other schemes(such as the LP-MPS)to achieve the same accuracy.

Figure 13 Relative error of solution to Poisson’s equation based on Eq.(13)and the corresponding CPU time in the cases with increased particle randomness(h=1.2s0;TRef is the CPU time spent by CSPH2Γ with s0=0.1 and K=0.6)

Figure 14 Average ratio of the CPU time spent by the LP-MPS against that by the QSFDI in the cases with different randomness for finding solutions to Poisson’s equation(h=1.2s0)

It is worth noting that the QSFDI method presented in the paper does not only give a new formula for the Laplacian discretisation but also provides the new schemes for the numerical interpolation and gradient estimations. This means that one may extend the QSFDI to deal with the first and 2ndderivatives in differential equations,e.g.the NS equation and advection-diffusion equations, with a linear consistency and quadric accuracy.

It shall be also noted that the implementation of the QSFDI in a Lagrangian meshless method to solve engineering problems,such as wave-structure interaction in maritime engineering, is under study and results will be discussed in other publications.

Appendix 1:Derivation of QSFDI

Appendix 2:Error Analysis of CSPM and improvement

Open AccessThis article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing,adaptation,distribution and reproduction in any medium or format,as long as you give appropriate credit to the original author(s)and the source,provide a link to the Creative Commons licence,and indicate if changes were made.The images or other third party material in this article are included in the article's Creative Commons licence,unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use,you will need to obtain permission directly from the copyright holder. To view a copy of this licence,visit http://creativecommons.org/licenses/by/4.0/.