Accuracy of Integral and Integral2 functions

4 visualizaciones (últimos 30 días)
Syed Abdul Rafay
Syed Abdul Rafay el 9 de Feb. de 2024
Comentada: Syed Abdul Rafay el 13 de Feb. de 2024
I am using integral functions but I am told the integral functions do have some inaccuracy when used. That's why my answers are not same with the paper I am trying to validate. let me share my code:
format long g
clear all
clc
syms lambda
syms x
syms i
rhoEpoxy = 1200;
Em = 3e+09;
R=5;
E_L=Em;
E_r=Em;
rho_L=rhoEpoxy;
rho_r=rhoEpoxy;
nu=3;
A_K=zeros(R+1,R+1);
A_M=zeros(R+1,R+1);
G=zeros(R+1,R+1);
H1_K=zeros(R+1,R+1);
H1_M=zeros(R+1,R+1);
H2_M=zeros(R+1,R+1);
for ri=1:R+1
r = ri-1;
for mi=1:R+1
m = mi-1;
fun1 = @(ksei1) (ksei1.^(r+m)).*((1-(exp(nu*ksei1)-1)./(exp(nu)-1))+E_r/E_L*(exp(nu*ksei1)-1)./(exp(nu)-1));
G(mi,ri)=integral(fun1,0,1);
fun_h1=@(ksei2,s2) (-2*nu/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^r).*(ksei2.^m)+...
(nu^2/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^r).*(ksei2.^(m+1))-...
(nu^2/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^(r+1)).*(ksei2.^m);
fun_h1_lambda=@(ksei3,s3) (-1/6*((ksei3-s3).^3).*((1-(exp(nu*s3)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s3)-1)./(exp(nu)-1))).*(s3.^r).*(ksei3.^m);
fun_h2_lambda=@(ksei4,s4) (1/6*(ksei4.^3).*((1-(exp(nu*s4)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s4)-1)./(exp(nu)-1))-1/2*(ksei4.^2).*s4.*((1-(exp(nu*s4)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s4)-1)./(exp(nu)-1))).*(s4.^r).*(ksei4.^m);
H1_K(mi,ri)=integral2(fun_h1,0,1,0,@(ksei2) ksei2);
H1_M(mi,ri)=integral2(fun_h1_lambda,0,1,0,@(ksei3) ksei3);
H2_M(mi,ri)=integral2(fun_h2_lambda,0,1,0,1);
A_K(mi,ri)=G(mi,ri)+H1_K(mi,ri);
A_M(mi,ri)=H1_M(mi,ri)+H2_M(mi,ri);
end
end
sort(sqrt(eig(A_K,-A_M)))
ans = 6×1
1.0e+00 * 3.51601526867379 22.034800684001 61.7194163744804 128.398091475146 226.847068659209 1094.88143674641
this is my code why I run this code i get following results;
3.516
22.035
61.719
128.4
but the correct results are;
2.4256
18.6031
55.1791
109.5748
  2 comentarios
Torsten
Torsten el 9 de Feb. de 2024
I don't believe that such a big difference between the results can be due to inaccuracies of integral2. There must be some substantial difference between the two computations.
Syed Abdul Rafay
Syed Abdul Rafay el 10 de Feb. de 2024
But the same code give very close results to other paper I am validating the results with

Iniciar sesión para comentar.

Respuestas (1)

Walter Roberson
Walter Roberson el 10 de Feb. de 2024
Editada: Walter Roberson el 10 de Feb. de 2024
Given that code, the eigenvalues come out the 3.516 and so on. It isn't a matter of low accuracy: I switched to high accuracy here.
format long g
clear all
clc
syms lambda
syms x
syms i
Q = @(v) sym(v);
rhoEpoxy = Q(1200);
Em = Q(3)*Q(10)^9;
R = 5;
E_L=Em;
E_r=Em;
rho_L=rhoEpoxy;
rho_r=rhoEpoxy;
nu = Q(3);
A_K=zeros(R+1,R+1,'sym');
A_M=zeros(R+1,R+1,'sym');
G=zeros(R+1,R+1,'sym');
H1_K=zeros(R+1,R+1,'sym');
H1_M=zeros(R+1,R+1,'sym');
H2_M=zeros(R+1,R+1,'sym');
syms ksei1 ksei2 ksei3 ksei4 s2 s3 s4
for ri=1:R+1
r = ri-1;
for mi=1:R+1
m = mi-1;
fun1 = @(ksei1) (ksei1.^(r+m)).*((1-(exp(nu*ksei1)-1)./(exp(nu)-1))+E_r/E_L*(exp(nu*ksei1)-1)./(exp(nu)-1));
G(mi,ri) = int(fun1(ksei1),ksei1,0,1);
fun_h1=@(ksei2,s2) (-2*nu/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^r).*(ksei2.^m)+...
(nu^2/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^r).*(ksei2.^(m+1))-...
(nu^2/(exp(nu)-1)*(E_r/E_L-1)*exp(nu*s2)).*(s2.^(r+1)).*(ksei2.^m);
fun_h1_lambda=@(ksei3,s3) (-1/6*((ksei3-s3).^3).*((1-(exp(nu*s3)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s3)-1)./(exp(nu)-1))).*(s3.^r).*(ksei3.^m);
fun_h2_lambda=@(ksei4,s4) (1/6*(ksei4.^3).*((1-(exp(nu*s4)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s4)-1)./(exp(nu)-1))-1/2*(ksei4.^2).*s4.*((1-(exp(nu*s4)-1)./(exp(nu)-1))+...
rho_r/rho_L*(exp(nu*s4)-1)./(exp(nu)-1))).*(s4.^r).*(ksei4.^m);
%H1_K(mi,ri) = integral2(fun_h1,0,1,0,@(ksei2) ksei2);
H1_K(mi,ri) = int(int(fun_h1(ksei2,s2), s2, 0, ksei2), ksei2, 0, 1);
%H1_M(mi,ri)=integral2(fun_h1_lambda,0,1,0,@(ksei3) ksei3);
H1_M(mi,ri) = int(int(fun_h1_lambda(ksei3,s3), s3, 0, ksei3), ksei3, 0, 1);
%H2_M(mi,ri)=integral2(fun_h2_lambda,0,1,0,1);
H2_M(mi,ri) = int(int(fun_h2_lambda(ksei4,s4), ksei4, 0, 1), s4, 0, 1);
A_K(mi,ri)=G(mi,ri)+H1_K(mi,ri);
A_M(mi,ri)=H1_M(mi,ri)+H2_M(mi,ri);
end
end
dA_K = double(A_K);
dA_M = double(A_M);
sort(sqrt(eig(dA_K,-dA_M)))
ans = 6×1
1.0e+00 * 3.51601526878612 22.0347977911742 61.7162927762873 128.389334759042 223.551343192217 1006.01205007505
  3 comentarios
Torsten
Torsten el 12 de Feb. de 2024
@Walter Roberson wanted to show you that the results are not different with higher accuracy, as was your supposition.
Syed Abdul Rafay
Syed Abdul Rafay el 13 de Feb. de 2024
I tried and now my first 3 modes are correct but my 4 th and 5 th mode are not correct. If i increase R number of mode shapes increases as well but only first 3 modes are correct rest have too much discrepency. Can you look why. Like 4 th mode is 121 and my 4th mode is 128. The 5th mode is 247 and mine is 563. This only start from 4 th mode first 3 are always correct.

Iniciar sesión para comentar.

Etiquetas

Productos


Versión

R2022b

Community Treasure Hunt

Find the treasures in MATLAB Central and discover how the community can help you!

Start Hunting!

Translated by