Adjoint method in gridap

New to Julia
,
1

Hi, I am trying to understand this code of gridap tutorial 18 but i couldnt is there anyone who can explain how below equation is written in julia as below?
dg/dpt=−2ℜ[w^†(dA/dpt u)]

function Dgfdpf(pf_vec; β, η, phys_params, fem_params)
    pfh = FEFunction(fem_params.Pf, pf_vec)
    pth = (pf -> Threshold(pf; β, η)) ∘ pfh
    A_mat = MatrixA(pth; phys_params, fem_params)
    b_vec = assemble_vector(v->(∫(v)fem_params.dΓ_s), fem_params.V)
    u_vec = A_mat \ b_vec
    O_mat = MatrixOf(fem_params)

    uh = FEFunction(fem_params.U, u_vec)
    w_vec =  A_mat' \ (O_mat * u_vec)
    wconjh = FEFunction(fem_params.U, conj(w_vec))

    l_temp(dp) = ∫(real(-2 * DAdpf(uh, wconjh, pfh; phys_params, β, η)) * dp)fem_params.dΩ_d
    dgfdpf = assemble_vector(l_temp, fem_params.Pf)
    return dgfdpf
end
2

−2ℜ[w^†(dA/dpt u)] is a linear form in u (here = the FEM basis functions), so it uses the above two lines via assemble_vector.

1 Like
3

why it isnt multiplied by w^T? it is just using uh and wconjh as inputs of dA. considering dAdp as below.

DAdpf(u, v, pfh; phys_params, β, η) = ((p -> Dξdpf(p, phys_params.n_air, phys_params.n_metal, β, η)) ∘ pfh) * (∇(v) ⊙ ∇(u))

as you can see there is ∇(v) ⊙ ∇(u) in dA, so what if there is (ε(v) ⊙ (σfun∘(ε(u))) in dA?
thank you

4

That’s what the wconjh argument does. If A is the FEM matrix of a bilinear form a(u,v), then \vec{u}^T A \vec{v} = a(u,v) = \int \text{something}(u,v) d\Omega, where u(x) = \sum_k u_k p_k(x) is the function formed from the FEM basis functions p_k(x) \in P with coefficients \vec{u}, and similarly for v(x) and \vec{v}.

Now, for the adjoint method, it’s the same sort of thing but with the derivative of A with respect to the parameters. The \vec{w}^\dagger turns into a function \overline{w(x)} in the integrand, which is represented here by wconjh.

2 Likes
5

so as you have mentioned that a(u,v) is the integration of something and in my case is

A_Disp(u,v,pth) =  ((p->Em(p))∘pth) * (ε(v) ⊙ (σfun∘(ε(u))))
    a_Disp(u,v) =∫(A_Disp(u,v,pth))fem_params.dΩ

l_temp would be as tutorial as below (without 2 as i think it is due to problem)

    l_temp(dp) = ∫(-(DAdpf(uh, wconjh, pfh; β, η)) * dp)fem_params.dΩ

knowing that

DAdpf(u, v, pfh; β, η) = ((p->DEdpf(p, β, η)) ∘ pfh) * (ε(v) ⊙ (σfun∘(ε(u))))
6

in this case it is assumed that b is independent from p so its derivation is zero. if i have a traction on boundary how can i assemble vector for derivation. DAdp is defined on domain while DbDp for traction and without body force defined on boundary?

dg/dp=- w^t(dA/dp * u - db/dp)

b(v)=(v ⋅ n ⋅ (σfun∘(ε(v)))

thanks
7

@stevengj is there any documents or papers to help me in this wat? as you say if u^TAv=a(u,v) however it seems that dor dA, the conjugate of w is going through dA function. Am i right? Also I want to know that if i have real numbers, just w goes through dA?
thank you

8

For complex numbers where your weak forms and/or objective functions involve complex conjugatiohn, you apply the adjoint method using CR calculus; there are some tutorials on CR calculus online.

9

@stevengj what if i dont have complex numbers?

10

For real numbers, the you can remove the conjugation (it does nothing anyway).