Physics informed Neural Network

37 次查看（过去 30 天）

Sankarganesh P 2023-7-11

1
链接

此问题的直接链接

https://ww2.mathworks.cn/matlabcentral/answers/1994258-physics-informed-neural-network

评论： Sankarganesh P 2023-7-27

Dear Matlab Community Members,

i would like to apply my equations in physics-informed neural networks, so i reffered solving Burger's equation L-BFGS method. For example i would like to apply 1D equation

d^u/dt=(d^2g/dx^2)-d^4u/dx^4; where g=2*u*(1-u)*(1-2*u) with 
u(x=0,t)=0, u(x=1,t)=0 and 
u(t=0,x)=sin(4*pi*x/L)

Model Loss Function

%%%%%%%%%%%%%%%%%%%%%%%%%%%%
U = model(parameters,X,T);
G=2.*U.*(1-U).*(1-2.*U);
gradientsG = dlgradient(sum(G,"all"),{X,T},EnableHigherDerivatives=true);
Gx = gradientsG{1};
Gxx = dlgradient(sum(Gx,"all"),X,EnableHigherDerivatives=true);
gradientsU = dlgradient(sum(U,"all"),{X,T},EnableHigherDerivatives=true);
Ux = gradientsU{1};
Ut = gradientsU{2};
Uxx = dlgradient(sum(Ux,"all"),X,EnableHigherDerivatives=true);
Uxxx = dlgradient(sum(Uxx,"all"),X,EnableHigherDerivatives=true);
Uxxxx = dlgradient(sum(Uxxx,"all"),X,EnableHigherDerivatives=true);
f=Ut-Gxx+2.*Uxxxx;
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%

i have made the above changes in the code...but i didnt get any expected results....can u please help me with this...?

4 个评论
显示 2更早的评论隐藏 2更早的评论

Ben 2023-7-26

@Sankarganesh P I notice that your ODE is

but the model loss function uses 2.*Uxxxx. That seems like an issue at first glance.

Otherwise your model loss function looks as expected to me. Do you get good training losses?

Of course a small training loss only means the network is "nearly" solving the ODE at the training points - for complex ODEs/PDEs this may only give weak approximations to the true solution in general.

Sankarganesh P 2023-7-27