Practical error bounds for the forces

This is a simple example showing how to compute error estimates for the forces on a ${\rm TiO}_2$ molecule, from which we can either compute asymptotically valid error bounds or increase the precision on the computation of the forces.

The strategy we follow is described with more details in [CDKL2021] and we will use in comments the density matrices framework. We will also needs operators and functions from src/scf/newton.jl.

using DFTK
using Printf
using LinearAlgebra
using ForwardDiff
using LinearMaps
using IterativeSolvers


We setup manually the ${\rm TiO}_2$ configuration from Materials Project.

Ti = ElementPsp(:Ti; psp=load_psp("hgh/lda/ti-q4.hgh"))
O  = ElementPsp(:O; psp=load_psp("hgh/lda/o-q6.hgh"))
atoms     = [Ti, Ti, O, O, O, O]
positions = [[0.5,     0.5,     0.5],  # Ti
             [0.0,     0.0,     0.0],  # Ti
             [0.19542, 0.80458, 0.5],  # O
             [0.80458, 0.19542, 0.5],  # O
             [0.30458, 0.30458, 0.0],  # O
             [0.69542, 0.69542, 0.0]]  # O
lattice   = [[8.79341  0.0      0.0];
             [0.0      8.79341  0.0];
             [0.0      0.0      5.61098]];

We apply a small displacement to one of the $\rm Ti$ atoms to get nonzero forces.

positions[1] .+= [-0.022, 0.028, 0.035]
3-element Vector{Float64}:

We build a model with one $k$-point only, not too high Ecut_ref and small tolerance to limit computational time. These parameters can be increased for more precise results.

model = model_LDA(lattice, atoms, positions)
kgrid = [1, 1, 1]
Ecut_ref = 35
basis_ref = PlaneWaveBasis(model; Ecut=Ecut_ref, kgrid)
tol = 1e-5;


We compute the reference solution $P_*$ from which we will compute the references forces.

scfres_ref = self_consistent_field(basis_ref; tol, callback=identity)
ψ_ref = DFTK.select_occupied_orbitals(basis_ref, scfres_ref.ψ, scfres_ref.occupation).ψ;

We compute a variational approximation of the reference solution with smaller Ecut. ψr, ρr and Er are the quantities computed with Ecut and then extended to the reference grid.

Choice of convergence parameters

Be careful to choose Ecut not too close to Ecut_ref. Note also that the current choice Ecut_ref = 35 is such that the reference solution is not converged and Ecut = 15 is such that the asymptotic regime (crucial to validate the approach) is barely established.

Ecut = 15
basis = PlaneWaveBasis(model; Ecut, kgrid)
scfres = self_consistent_field(basis; tol, callback=identity)
ψr = DFTK.transfer_blochwave(scfres.ψ, basis, basis_ref)
ρr = compute_density(basis_ref, ψr, scfres.occupation)
Er, hamr = energy_hamiltonian(basis_ref, ψr, scfres.occupation; ρ=ρr);

We then compute several quantities that we need to evaluate the error bounds.

  • Compute the residual $R(P)$, and remove the virtual orbitals, as required in src/scf/newton.jl.
res = DFTK.compute_projected_gradient(basis_ref, ψr, scfres.occupation)
res, occ = DFTK.select_occupied_orbitals(basis_ref, res, scfres.occupation)
ψr = DFTK.select_occupied_orbitals(basis_ref, ψr, scfres.occupation).ψ;
  • Compute the error $P-P_*$ on the associated orbitals $ϕ-ψ$ after aligning them: this is done by solving $\min |ϕ - ψU|$ for $U$ unitary matrix of size $N×N$ ($N$ being the number of electrons) whose solution is $U = S(S^*S)^{-1/2}$ where $S$ is the overlap matrix $ψ^*ϕ$.
function compute_error(basis, ϕ, ψ)
    map(zip(ϕ, ψ)) do (ϕk, ψk)
        S = ψk'ϕk
        U = S*(S'S)^(-1/2)
        ϕk - ψk*U
err = compute_error(basis_ref, ψr, ψ_ref);
  • Compute ${\boldsymbol M}^{-1}R(P)$ with ${\boldsymbol M}^{-1}$ defined in [CDKL2021]:
P = [PreconditionerTPA(basis_ref, kpt) for kpt in basis_ref.kpoints]
map(zip(P, ψr)) do (Pk, ψk)
    DFTK.precondprep!(Pk, ψk)
function apply_M(φk, Pk, δφnk, n)
    DFTK.proj_tangent_kpt!(δφnk, φk)
    δφnk = sqrt.(Pk.mean_kin[n] .+ Pk.kin) .* δφnk
    DFTK.proj_tangent_kpt!(δφnk, φk)
    δφnk = sqrt.(Pk.mean_kin[n] .+ Pk.kin) .* δφnk
    DFTK.proj_tangent_kpt!(δφnk, φk)
function apply_inv_M(φk, Pk, δφnk, n)
    DFTK.proj_tangent_kpt!(δφnk, φk)
    op(x) = apply_M(φk, Pk, x, n)
    function f_ldiv!(x, y)
        x .= DFTK.proj_tangent_kpt(y, φk)
        x ./= (Pk.mean_kin[n] .+ Pk.kin)
        DFTK.proj_tangent_kpt!(x, φk)
    J = LinearMap{eltype(φk)}(op, size(δφnk, 1))
    δφnk = cg(J, δφnk; Pl=DFTK.FunctionPreconditioner(f_ldiv!),
              verbose=false, reltol=0, abstol=1e-15)
    DFTK.proj_tangent_kpt!(δφnk, φk)
function apply_metric(φ, P, δφ, A::Function)
    map(enumerate(δφ)) do (ik, δφk)
        Aδφk = similar(δφk)
        φk = φ[ik]
        for n = 1:size(δφk,2)
            Aδφk[:,n] = A(φk, P[ik], δφk[:,n], n)
Mres = apply_metric(ψr, P, res, apply_inv_M);

We can now compute the modified residual $R_{\rm Schur}(P)$ using a Schur complement to approximate the error on low-frequencies[CDKL2021]:

\[\begin{bmatrix} (\boldsymbol Ω + \boldsymbol K)_{11} & (\boldsymbol Ω + \boldsymbol K)_{12} \\ 0 & {\boldsymbol M}_{22} \end{bmatrix} \begin{bmatrix} P_{1} - P_{*1} \\ P_{2}-P_{*2} \end{bmatrix} = \begin{bmatrix} R_{1} \\ R_{2} \end{bmatrix}.\]

  • Compute the projection of the residual onto the high and low frequencies:
resLF = DFTK.transfer_blochwave(res, basis_ref, basis)
resHF = res - DFTK.transfer_blochwave(resLF, basis, basis_ref);
  • Compute ${\boldsymbol M}^{-1}_{22}R_2(P)$:
e2 = apply_metric(ψr, P, resHF, apply_inv_M);
  • Compute the right hand side of the Schur system:
# Rayleigh coefficients needed for `apply_Ω`
Λ = map(enumerate(ψr)) do (ik, ψk)
    Hk = hamr.blocks[ik]
    Hψk = Hk * ψk
ΩpKe2 = DFTK.apply_Ω(e2, ψr, hamr, Λ) .+ DFTK.apply_K(basis_ref, e2, ψr, ρr, occ)
ΩpKe2 = DFTK.transfer_blochwave(ΩpKe2, basis_ref, basis)
rhs = resLF - ΩpKe2;
  • Solve the Schur system to compute $R_{\rm Schur}(P)$: this is the most costly step, but inverting $\boldsymbol{Ω} + \boldsymbol{K}$ on the small space has the same cost than the full SCF cycle on the small grid.
(; ψ) = DFTK.select_occupied_orbitals(basis, scfres.ψ, scfres.occupation)
e1 = DFTK.solve_ΩplusK(basis, ψ, rhs, occ; tol).δψ
e1 = DFTK.transfer_blochwave(e1, basis, basis_ref)
res_schur = e1 + Mres;

Error estimates

We start with different estimations of the forces:

  • Force from the reference solution
f_ref = compute_forces(scfres_ref)
forces   = Dict("F(P_*)" => f_ref)
relerror = Dict("F(P_*)" => 0.0)
compute_relerror(f) = norm(f - f_ref) / norm(f_ref);
  • Force from the variational solution and relative error without any post-processing:
f = compute_forces(scfres)
forces["F(P)"]   = f
relerror["F(P)"] = compute_relerror(f);

We then try to improve $F(P)$ using the first order linearization:

\[F(P) = F(P_*) + {\rm d}F(P)·(P-P_*).\]

To this end, we use the ForwardDiff.jl package to compute ${\rm d}F(P)$ using automatic differentiation.

function df(basis, occupation, ψ, δψ, ρ)
    δρ = DFTK.compute_δρ(basis, ψ, δψ, occupation)
    ForwardDiff.derivative(ε -> compute_forces(basis, ψ.+ε.*δψ, occupation; ρ=ρ+ε.*δρ), 0)
  • Computation of the forces by a linearization argument if we have access to the actual error $P-P_*$. Usually this is of course not the case, but this is the "best" improvement we can hope for with a linearisation, so we are aiming for this precision.
df_err = df(basis_ref, occ, ψr, DFTK.proj_tangent(err, ψr), ρr)
forces["F(P) - df(P)⋅(P-P_*)"]   = f - df_err
relerror["F(P) - df(P)⋅(P-P_*)"] = compute_relerror(f - df_err);
  • Computation of the forces by a linearization argument when replacing the error $P-P_*$ by the modified residual $R_{\rm Schur}(P)$. The latter quantity is computable in practice.
df_schur = df(basis_ref, occ, ψr, res_schur, ρr)
forces["F(P) - df(P)⋅Rschur(P)"]   = f - df_schur
relerror["F(P) - df(P)⋅Rschur(P)"] = compute_relerror(f - df_schur);

Summary of all forces on the first atom (Ti)

for (key, value) in pairs(forces)
    @printf("%30s = [%7.5f, %7.5f, %7.5f]   (rel. error: %7.5f)\n",
            key, (value[1])..., relerror[key])
                        F(P_*) = [1.47896, -1.25376, 0.81012]   (rel. error: 0.00000)
                          F(P) = [1.13550, -1.01531, 0.40016]   (rel. error: 0.20484)
        F(P) - df(P)⋅Rschur(P) = [1.29122, -1.10178, 0.69055]   (rel. error: 0.07832)
          F(P) - df(P)⋅(P-P_*) = [1.50902, -1.28635, 0.86147]   (rel. error: 0.08072)

Notice how close the computable expression $F(P) - {\rm d}F(P)⋅R_{\rm Schur}(P)$ is to the best linearization ansatz $F(P) - {\rm d}F(P)⋅(P-P_*)$.

  • CDKL2021E. Cancès, G. Dusson, G. Kemlin, and A. Levitt Practical error bounds for properties in plane-wave electronic structure calculations Preprint, 2021. arXiv