Computational Homogenization of Mesoscale Gradient Viscoplasticity

Available online at www.sciencedirect.
com
ScienceDirect
Comput. Methods Appl. Mech. Engrg. 317 (2017) 927–951

www.elsevier.com/locate/cma
Computational homogenization of mesoscale gradient

viscoplasticity
Kenneth Runesson ∗ , Magnus Ekh, Fredrik Larsson
Department of Applied Mechanics, Chalmers University of Technology, Sweden
Received 23 March 2016; received in revised form 25 November 2016; accepted 28 November 2016
Available online 6 January 2017
Abstract
Variationally consistent selective homogenization is applied to a class of gradient-enhanced dissipative materials which
is adopted for the mesoscale modeling. The adopted standard first order homogenization assumption results in the classical
equilibrium equation for a local continuum on the macroscale, while the internal variables “live” on the mesoscale only. Among
the issues considered in the paper, we note (i) the variationally consistent setting of homogenization of gradient theory on the
mesoscale, (ii) the variational basis of the SVE-problem (SVE = Statistical Volume Element) in the time-incremental setting. The
SVE-problem is formulated for the classical boundary conditions (Dirichlet and Neumann) pertinent to the standard momentum
balance as well as the “micro-momentum” balance. The weak format of the SVE-problem constitutes the stationarity condition of
an incremental SVE-potential, which represents an extension of the situation for a single-phase continuum model. The macroscale
stress in a given time-increment is derivable from an incremental “macroscale pseudo-elastic strain energy”. Moreover, the saddle-
point properties of the SVE-potential are shown to form the basis for establishing upper and lower bounds on the pseudo-
elastic strain energy. Bingham viscoplasticity with gradient-enhanced hardening is chosen as the prototype model problem for
the numerical evaluation. The computed stress–strain response relations confirm the theoretical predictions of the influence of
different combinations of boundary conditions on the SVE.
⃝c 2017 Elsevier B.V. All rights reserved.
Keywords: Selective homogenization; Gradient viscoplasticity; Bounds
1. Introduction
A wealth of literature is available on socalled higher order continuum models, that are enriched by the subscale
length scale. It then becomes possible to predict the macroscopically manifested size effect for smooth displacement
fields as well as non-smooth ones that accommodate an emerging localization zone of finite width. Different classes
of higher order models (nonlocal, gradient, micromorphic) can be identified depending on the basic kinematic
assumptions; however, it is sometimes difficult to draw the borderline between these model classes, and an exhaustive
∗ Corresponding author.
E-mail addresses: [email protected], [email protected] (K. Runesson).
http://dx.doi.org/10.1016/j.cma.2016.11.032
0045-7825/⃝ c 2017 Elsevier B.V. All rights reserved.
928 K. Runesson et al. / Comput. Methods Appl. Mech. Engrg. 317 (2017) 927–951
review is outside the scope of this paper. An import class is gradient-enhanced (visco)plasticity models, with early
contribution by e.g. Zbib and Aifantis [1], Fleck et al. [2]. Thermodynamically based model formulations were
proposed by Maugin [3], Valanis [4], Svedberg and Runesson [5]. Admittedly, a significant drawback of higher
order continuum models is that the resulting problem becomes quite complex, both from modeling (higher order
boundary conditions) and numerical (higher regularity conditions) points of view. Numerical issues are discussed
by, among others, Svedberg and Runesson [6], Liebe and Steinmann [7], Liebe et al. [8]. Recent comprehensive
accounts on gradient-extended standard dissipative solids, comprising theoretical and algorithmic fundamentals as
well as various applications, were given by Miehe [9], Welschinger [10], Miehe et al. [11], Miehe [12], Mauthe and
Miehe [13].
A vast amount of work has been devoted to the development of gradient-enhanced crystal plasticity models and
the associated computational procedure, see e.g. Cermelli and Gurtin [14], Acharya et al. [15], Evers et al. [16],
Yefimov and van der Giessen [17], Levkovitch and Svendsen [18], Ohno and Okumura [19], Borg [20], Kuroda
and Tvergaard [21], Bargmann and Reddy [22] Bargmann et al. [23], Bargmann et al. [24], Poh et al. [25], Miehe
et al. [26]. For a polycrystal, it is possible to impose further restrictions along the grain boundaries by formally
introducing “intergrain relations”, e.g. Ekh et al. [27], Gurtin [28], Ekh et al. [29], Gottschalk et al. [30], van Beers
et al. [31], that represent the resistance of the grain boundary to transmit dislocation flux.
A class of higher order continuum models, which are frequently used to represent the macroscale response due
to subscale nonlocality, emerge after some sort of homogenization. As to the underlying kinematics leading to such
a higher order macroscale model, a well-established approach is to introduce higher order “strain-like” quantities as
part of a Taylor series expansion of the displacement field. (In the simplest case of linear expansion, the classical
first order homogenization models are obtained.) Typical proponents for this type of reasoning are the 2nd order
homogenization model by Kouznetsova et al. [32] and the micropolar model by Larsson and Diebels [33]. Via this
type of procedure, a standard continuum description on the subscale level translates to a generalized continuum on the
macroscale.
A quite different approach to the modeling of (in particular) polycrystals is to incorporate a “higher order” model
on the mesoscale, viz. gradient effects within and between the grains. The variety of gradient crystal (visco)plasticity
models are then at our disposal. When applied to polycrystals, it is desirable that (first order) homogenization will lead
to a standard continuum format on the macroscale, involving the macroscale stress as the only relevant homogenized
“flux” variable, whereas possible microstresses are confined to the mesoscale and are automatically suppressed on the
macroscale as part of the homogenization.
In this contribution we apply a fairly general setting of variationally consistent selective homogenization of the
chosen gradient-enhanced continuum models on the mesoscale. The notion “selective” indicates that not all the fields
in the subscale multifield problem are homogenized. The chosen setting is that of the “Variational Multiscale Method”,
originally proposed by Hughes et al. [34] in a quite different context, and it is based on a Variationally Consistent
Macrohomogeneity Condition, VCMC (that generalizes the well-known Hill–Mandel condition to a large class of
time-dependent and multi-field problems, e.g. Larsson et al. [35]. Via the VCMC, the appropriate homogenized
(macroscale) problem is derived. Since the main purpose of this paper is to provide a proof of concept rather than
computing quantitatively useful results, a gradient-enhanced continuum model is adopted (for the sake of simplicity
and transparency). Indeed, crystal (visco)plasticity with gradient effects is more physically motivated; however, it is
significantly heavier from a computational viewpoint and does not offer any conceptual advantages in the present
context.
The paper is outlined as follows: A (sub)class of gradient-enhanced dissipative materials is considered in Sec-
tion 2 in the time-continuous and the time-discrete formats. In particular, we establish the semi-dual format in terms
of a “micro-stress energy”, which is obtained after partial Legendre transformation of the free energy. The result-
ing generalized saddle-point problem, based on an incremental SVE-potential, represents the workhorse for eval-
uation of bounds on the macroscale stress (stiffness) in a given time-increment. Issues related to computational
homogenization are discussed in Section 3, whereby the “canonical format” of the SVE-problem for a single re-
alization is established via the assumption of weakly enforced micro(anti)periodicity of displacement fluctuations
and microtractions. Finally in this Section, energetic bounds are established for a single SVE. The corresponding
FE-approximation and its algorithmic setting are discussed in Section 4. The adopted gradient-enhanced viscoplas-
ticity model, chosen as a prototype model, is presented in Section 5. In Section 6, computational results illustrate the
K. Runesson et al. / Comput. Methods Appl. Mech. Engrg. 317 (2017) 927–951 929
various theoretical findings and predictions. Finally, concluding remarks and an outlook to future work are given in
Section 7.
2. A (sub)class of gradient-enhanced dissipative materials
2.1. Model framework—Thermodynamic balance equations
2.1.1. Preliminaries
As a preliminary we establish the pertinent balance equations of “physical forces” (classical equilibrium) and
“constitutive forces” (commonly denoted micro-equilibrium in the literature) for a (sub)class of models within the
larger class of “gradient-enhanced standard materials”, c.f. Miehe [9], Nguyen [36]. We then consider the time-
dependent deformation in the meso-heterogeneous material, such as a polycrystal with different phases, of a body
occupying the given domain Ω . We restrict to small strain kinematics; hence, the meso-scale deformation is defined by
def
the strain field ϵ(x, t) with ϵ[u] = (u ⊗ ∇)sym , where u(x, t) is the displacement field. Moreover, the meso-scale field
k(x, t) is a column matrix containing internal variables that represent the dissipative microscale mechanisms.1 The
internal variables are tensors of (normally) even order (0th, 2nd, 4th, etc.). Moreover, we introduce the “gradient
def
variable” g(x, t) with g[k] = k ⊗ ∇.
As to the used notation, we distinguish between field variables in the balance equations and constitutive (or
thermodynamic) variables: The latter variables are either independent when used as arguments in the volume-
specific thermodynamic (and other) potentials or they are dependent variables (state functions). A dot · denotes
scalar multiplication of vectors, tensors, whereas a star ⋆ denotes scalar matrix multiplication of two column matrices
(replacing the transpose symbol). Finally, ⋆˙ denotes the combined scalar multiplication.
2.1.2. Canonical constitutive format

The mesoscale material properties of a dissipative material are characterized by two volume-specific potentials: the
free energy density ψ(ϵ, k, g) and the dissipation density φ(k̇).2 Furthermore, for some brevity of notation in what
follows we assume the additive split ψ(ϵ, k, g) = ψ loc (ϵ, k) + ψ gra (g), whereby the gradient effect is decoupled
from the elastic properties. The adopted framework is sufficiently general to contain (visco)plasticity with gradient-
enhanced hardening.
From the balance of energetic changes due to storage and dissipation of energy in an arbitrary part of a body
under quasi-static loading, it is possible to derive the balance equations corresponding to equilibrium of classical
stresses and of “constitutive stresses” after using the proper localization arguments, Miehe [9], Nguyen [36]. The
fields u(x, t), k(x, t) then satisfy the strong formulation3 :
−σ (ϵ[u], k) · ∇ = 0 in Ω × R+ , (1a)
−κ(ϵ[u], k) + κ (k̇) − ξ (g[k]) · ∇ = 0
di
in Ω × R . +
(1b)
In the context of standard local (non-gradient) models, (1b) is known as the Biot equations, e.g. Nguyen [36]. We
introduced the following state functions (“reversible stresses” in the nomenclature of Nguyen [36], or “energetic
microstresses” in the nomenclature of Gurtin [37] in the context of gradient crystal plasticity):
def ∂ψ ∂ψ loc def ∂ψ ∂ψ loc def ∂ψ ∂ψ gra

σ = = , κ =− =− , ξ = = (2)
∂ϵ ∂ϵ ∂k ∂k ∂g ∂g
and the “dissipative stress”
def ∂φ
κ di = . (3)
∂ k̇
1 In the simple model exploited subsequently for the numerical calculation, we take k = [ϵ p , k]T . However, for simplicity of notation we treat
the internal variables as scalar fields, and we set dim(k) = M × 1.
2 This is a restriction of the more general framework in Miehe [9] who adopted the extended parametrization φ(ϵ̇, k̇, ġ; ϵ, k, g).
3 The case with zero body force is considered for the sake of simplicity.
The canonical format is usually non-operational in practice. Rather than constructing φ(k̇) directly, one may
propose an expression for the dual dissipation potential φ ∗ (κ di ) and use the appropriate Legendre transformation
to obtain φ(k̇):
 
φ(k̇) = sup κ̂ di ⋆ k̇ − φ ∗ (κ̂ di ) . (4)
κ̂ di
The stationarity condition gives the evolution rule for k expressed as

∂φ ∗
k̇ = . (5)
∂κ di
As a result, the fields u(x, t), k(x, t), κ di (x, t) satisfy the strong formulation as follows:
−σ (ϵ[u], k) · ∇ = 0 in Ω × R+ , (6a)
−κ(ϵ[u], k) + κ di − ξ (g[k]) · ∇ = 0 in Ω × R+ , (6b)
∂φ ∗
k̇ − di (κ di ) = 0 in Ω × R+ . (6c)
∂κ
Adopting the standard BE-rule to integrate the system (6) in the time interval (tn , tn+1 = tn + ∆t), we obtain the
solution for the time-updated fields u(x), k(x), κ di (x)4 from the time discrete equations:
−σ (ϵ[u], k) · ∇ = 0 in Ω , (7a)
−κ(ϵ[u], k) + κ − ξ (g[k]) · ∇ = 0
di
in Ω , (7b)
∂φ ∗
k − ∆t (κ di ) = nk in Ω . (7c)
∂κ di
Based on this “canonical” strong format of the time-discrete problem, different weak formats can be established. The
straightforward approach leads to the primal format, whereby the fields u, k, κ di are kept as the independent global
fields in the variational setting, c.f. Miehe [9]. However, in this paper we adopt the semi-dual format (as defined
below).
Finally in this Subsection, in order to obtain an operational version of the dual format that comprises
(visco)plasticity, we define the elastic domain E as
E = {κ di | F(κ di ) ≤ 0} (8)
where we introduced the convex “threshold function” F(κ di ) (or quasistatic yield surface in terms of the nomenclature
of viscoplasticity). Typically, we express φ ∗ (κ di ) = t1∗ η(F(κ di )), where t∗ is a relaxation (time) parameter, and the
scalar “overstress” function η(x) is required to possess the properties η′ (x) ≥ 0 for x ≥ 0 and η′ (x) = 0 for x ≤ 0.5
Hence, the evolution rule for k in (5) is expressed as
∂ F di def 1
k̇ = λ(κ di ) (κ ) with λ(κ di ) = η′ (F(κ di )). (9)
∂κ di t∗
2.2. Time-discrete variational formulation—The semi-dual format
It turns out to be convenient (for a number of purposes, as discussed below) to employ a semi-dual, rather than the
straightforward primal, weak format of the micro-balance equation (7b). This is accomplished by exploiting a (partial)
Legendre transformation w.r.t. the gradient variable g as follows:
 
def
ψ ∗,gra (ξ ) = sup ξ ⋆˙ ĝ − ψ gra (ĝ) (10)
ĝ
4 u(x) def
= n+1u(x) etc., i.e. superindex n+1 is dropped for brevity.
5 Notation η′ (x) def
= dη(x)/ dx.
and we define the semi-dual free energy density as

   
ϕ(ϵ, k, ξ ) = inf ψ(ϵ, k, ĝ) − ξ ⋆˙ ĝ = ψ loc (ϵ, k) − sup ξ ⋆˙ ĝ − ψ gra (ĝ) = ψ loc (ϵ, k) − ψ ∗,gra (ξ ). (11)
ĝ ĝ
The stationarity condition of (10) gives the constitutive equation for g expressed as
∂ψ ∗,gra ∂ϕ
g= =− . (12)
∂ξ ∂ξ
We note that, since (10) represents only a partial Legendre transformation, the semi-dual free energy ϕ is the
appropriate potential for σ and κ as well:
∂ϕ ∂ϕ
σ = , κ=− . (13)
∂ϵ ∂k
Boundary conditions6 will be imposed on “displacements” and “tractions” as follows:

def
u = ūpre on ∂Ωu , t[σ ] = σ · n = t̄pre on ∂Ωt , ∂Ω = ∂Ωu ∪ ∂Ωt , (14)
def
k = k̄ pre on ∂Ωk , p[ξ ] = ξ · n = p̄ pre on ∂Ωp , ∂Ω = ∂Ωp ∪ ∂Ωk . (15)
Remark. The boundary condition on ∂Ωp is essential, whereas the condition on ∂Ωk is natural. The roles of these
two conditions are thereby switched as compared to the corresponding primal variational setting.
Next, we introduce the appropriate solution and test spaces

U = {u ∈ H1 (Ω ), u = ūpre on ∂Ωu } (16a)
def
X = {ξ ∈ [H1div (Ω )] M , p[ξ ] = ξ · n = p̄ pre on ∂Ωp } (16b)
U0 = {u ∈ H1 (Ω ), u = 0 on ∂Ωu } (16c)
def
X0 = {ξ ∈ [H1div (Ω )] M , p[ξ ] = ξ · n = 0 on ∂Ωp } (16d)
K = [L2 (Ω )] M
(16e)
Here, L2 (Ω ) denotes the conventional (Hilbert) space of Lebesgue square integrable functions. Moreover, H1 (Ω ) and
H1div (Ω ) denote the spaces for vector fields with gradients and divergence, respectively, in L2 (Ω ).
We are now in the position to pose the semi-dual variational format corresponding to the system (7) and (12): Find
the fields u ∈ U, ξ ∈ X, k ∈ K, and κ di ∈ K that solve

σ (ϵ[u], k) : δϵ dV = l (u) (δu) ∀δu ∈ U0 , (17a)
  Ω

−g(ξ )˙⋆δξ − k ⋆ χ [δξ ] dV = l (ξ ) (δξ ) ∀δξ ∈ X0 , (17b)
  Ω

−κ(ϵ[u], k) + κ di − χ[ξ ] ⋆ δk dV = 0 ∀δk ∈ K, (17c)
Ω
∂φ ∗
   
k − ∆t di (κ di ) ⋆ δκ di dV = n
k ⋆ δκ di dV ∀δκ di ∈ K (17d)
Ω ∂κ Ω
6 Subscript “pre” denotes “prescribed value”.

def
where we introduced the “micro-force” χ[ξ ] = ξ · ∇, and where we introduced the data in terms of the linear
functionals

(u)
l (δu) = t̄pre · δu dS, (18a)
∂ Ωt

l (ξ ) (δξ ) = − k̄ pre ⋆ p[δξ ] dS. (18b)
∂ Ωk
Note that g(ξ ) in (17b) is a dependent state-variable, similarly to σ (ϵ, k) and κ(ϵ, k).
The derivation of (17) deserves a comment: Eq. (17a) is the standard weak format of equilibrium, whereas (17c)
and (17d) are the weak forms of (7b) and (7c), respectively, without using integration by parts. Eq. (17b) is, however,
somewhat nonstandard; it is obtained upon stating the identity g(ξ ) = k ⊗ ∇ in weak form as
  
g(ξ ) − k ⊗ ∇ ⋆˙ δξ dV = 0 (19)
Ω
for suitable test functions δξ and integrating by parts in the last term in order to shift differentiation from k to δξ . We
emphasize that vanishing of the integrands in (17c) and (17d) represents the corresponding strong forms, which are
utilized in the numerical algorithm in Section 4.4.
Next, we introduce the volume-specific incremental potential π as
π(ϵ, k, ξ , χ, κ di ) = ϕ(ϵ, k, ξ ) − nϕ + κ di ⋆ [k − nk] − χ ⋆ k − ∆t φ ∗ (κ di ). (20)
Moreover, we introduce the global incremental potential Π = Π int + Π ext , where the internal and external parts Π int
and Π ext are defined as

Π (u, ξ , k, κ ) =
int di
π(ϵ[u], k, ξ , χ[ξ ], κ di ) dV (21)
Ω
Π ext (u, ξ ) = −l (u) (u − nu) − l (ξ ) (ξ ) (22)
and we note that the system (17) represents the stationarity conditions of the problem
min max min max Π (û, ξ̂ , k̂, κ̂ di ). (23)

û∈U ξ̂ ∈X k̂∈K κ̂ di ∈K
That the extremal properties in (23) are, indeed, the relevant ones are shown in Appendix A.
Special case: For the local (non-gradient) model, ξ = 0 and (17c) infers the identity κ di = κ(ϵ, k); hence, the
system (17) reduces to that of solving for u ∈ U and k ∈ K from

σ (ϵ[u], k) : ϵ[δu] dV = l (u) (δu) ∀δu ∈ U0 (24a)
Ω
∂φ ∗
k − ∆t (κ(ϵ[u], k)) = nk in Ω (24b)
∂κ
whereby it is noted that (24b) is identical to the strong format in (7c) and represents a local problem pointwise in Ω
that can be solved for k{ϵ} for any given ϵ. This value represents the stationary point of π(ϵ, k{ϵ}, 0, 0, κ(ϵ, k{ϵ})) in
the direction of k for any given value of ϵ. We may thus compute the stationary value
w{ϵ} = π(ϵ, k{ϵ}, 0, 0, κ(ϵ, k{ϵ}))
= ψ loc (ϵ, k{ϵ}) − nψ loc + κ(ϵ, k{ϵ}) ⋆ [k{ϵ} − nk] − ∆tφ ∗ (κ(ϵ, k{ϵ})) (25)
and it can be shown, e.g. Miehe [9], that this “volume-specific incremental pseudo-elastic strain energy” serves as the
potential for σ {ϵ} in the given time interval, i.e.
∂ψ loc (ϵ, k{ϵ})

dw{ϵ} = |k : dϵ = σ {ϵ} : dϵ. (26)
∂ϵ
It should be noted that no such “local” interpretation of any (volume-specific) incremental pseudo-elastic strain energy
w{ϵ} is possible when gradient effects are present. This drawback was “overcome” by Hurtado [38] in the context of
the primal variational formulation by choosing simply g = ng as a “parameter” in ψ(ϵ, k, g).
3. Variationally consistent homogenization
3.1. Preliminaries
The classical approach (which is adopted in this paper) is to introduce “model-based homogenization”, whereby a
local field y is replaced by the spatially homogenized (volume-averaged) field ⟨y⟩ :

def 1
y(x̄) → ⟨y⟩ (x̄) = y dV, x̄ ∈ Ω , (27)
|Ω (x̄)| Ω
representing a smoothing approximation on a Statistical Volume Element (SVE). In practice, the SVE’s are finite-
sized and occupy the subscale region Ω with boundary Γ . Subsequently, we shall assume the SVE to be a cube
(square in 2D) with side length L . The SVE is centered at the macroscale position x̄, i.e. ⟨x − x̄⟩ = 0 for any given
x̄ ∈ Ω .7
Inside each SVE, a typical field y = y M + y µ is split into the smooth part, y M , and the subscale fluctuation, y µ . The
scales are linked by expressing y M (x̄, x), for x̄ ∈ Ω and x ∈ Ω (x̄), in terms of the macroscale solution ȳ(x̄), x̄ ∈ Ω ,
def
in an explicit fashion and defining the coupling of the fluctuation y µ to ȳ, i.e. defining the prolongation8 y µ = y µ { ȳ}
for given ȳ. This prolongation allows for computing the homogenized quantities.
In this paper, standard (model-based) first order homogenization will be assumed for the displacements, i.e. the
macroscale field uM is assumed to vary linearly within the SVE. Hence, we have the explicit representation
uM (x̄, x) = ū(x̄) + ϵ̄(x̄) · [x − x̄] for x ∈ Ω (28)

def
where ϵ̄ = (ū ⊗ ∇)sym is the macroscale strain that is associated with the (global) macroscale displacement field
ū ∈ Ū that is “sufficiently smooth”.
On the other hand, no homogenization at all will be imposed on the internal variables. That the internal variables
“live only on the mesoscale” is the very essence of the adopted strategy to obtain a local continuum formulation on
the macroscale. Summarizing, we conclude that SVE-problem will provide solutions uµ {ϵ̄}, k{ϵ̄}, κ di {ϵ̄} from which
σ {ϵ̄}, ξ {ϵ̄} can be computed.
Variationally consistent homogenization is obtained from (17a) upon testing with the smooth function δuM (x̄, x) =
δ ū(x̄) + δ ϵ̄(x̄) · [x − x̄] within each SVE. As a result, we may obtain the macroscale problem in standard fashion: Find
ū ∈ Ū s.t.
 
σ̄ {ϵ̄} : ϵ[δ ū] dV = t̄p · δ ū dS ∀δ ū ∈ Ū0 (29)
Ω ∂ Ωt
def
where the macroscale stress is the volume average σ̄ = ⟨σ (ϵ[u], k)⟩ .
Before turning to the formulation of the SVE-problem, the solution of which admits the calculation of σ̄ for given
ϵ̄, we note that this problem must be defined so as to satisfy the Hill–Mandel condition on macrohomogeneity (rather,
a generalized condition as compared to the standard stress problem). However, for brevity in this paper we omit the
explicit discussion of the underlying variational structure. We merely conclude that (i) it is possible to establish a
generic formulation of the macrohomogeneity condition that is based on “variationally consistent homogenization”,
as given by Larsson & al. [35], (ii) the choice of SVE-formulation below does indeed satisfy the macrohomogeneity
condition.
7 Henceforth, the argument x̄ is suppressed unless there is a risk of confusion.

8 Curly brackets {(•)} indicate implicit functional dependence on (•).
3.2. SVE-problem—Weak enforcement of micro-(anti)periodicity
We consider the SVE-problem corresponding to the problem in (17) that is based on the semi-dual variational
format:

1
⟨σ (ϵ[u], k) : ϵ[δu]⟩ − t · δu dS = 0, (30a)
|Ω | Γ

1
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ + k ⋆ p[δξ ] dS = 0, (30b)
|Ω | Γ
⟨−κ(ϵ[u], k) ⋆ δk⟩ + ⟨κ di ⋆ δk⟩ − ⟨χ[ξ ] ⋆ δk⟩ = 0, (30c)
∂φ ∗
 
⟨k ⋆ δκ di ⟩ − ∆t di (κ di ) ⋆ δκ di = ⟨ nk ⋆ δκ di ⟩ , (30d)
∂κ
where the solution and test spaces are as yet undefined.
In order to ensure a uniquely solvable SVE-problem, we now adopt the following additional model assumptions:
def
• We assume that uµ is periodic in a weak sense, while the tractions t[σ ] = σ · n are strongly anti-periodic and
subject to modeling on Γ , cf. Larsson et al. [39]:
 
δλt · uµ (x+ ) − uµ (x− ) dS = δλt · [[uµ ]] dS = 0 ∀δλt ∈ Lt, ,
 
(31)
Γ+ Γ+
def
where we introduced the “difference operator” [[• ]] = •(x+ ) − •(x− ), whereby x+ ∈ Γ+ is a “mirror point”
whereas x− ∈ Γ \ Γ+ is the corresponding “image point”, cf. Fig. 1(a).
def
• We make the “dual” assumption that micro-tractions p[ξ ] = ξ · n are weakly anti-periodic, while k is strongly
periodic and subject to modeling on Γ :
   
δλk ⋆ p(x+ ) + p(x− ) dS = δλk ⋆ [[ξ ]] ·n dS = 0 ∀δλk ∈ Lk, . (32)
Γ+ Γ+
Clearly, δλt in (31) represents a field of virtual boundary tractions, whereas δλk in (32) represents a field of virtual
internal variables. Now, upon introducing the forms (for the sake of brevity)
 
(u) def 1 (k) def 1
d (λt , u) = λt · [[u ]] dS, d (λk , ξ ) = λ ⋆ [[[ξ ]] ·n] dS (33)
|Ω | Γ+ |Ω | Γ+ k
we may establish a “canonical” variational format of the SVE-problem for the case of macroscale strain control: For
given value of ϵ̄, find the fields u ∈ U , λt ∈ Lt, , ξ ∈ X , λk ∈ Lk, , k ∈ K , and κ di ∈ K that solve the system
(u)
⟨σ (ϵ[u], k) : ϵ[δu]⟩ − d (λt , δu) = 0 ∀δu ∈ U , (34a)
(u) (u)
−d (δλt , u) = −d (δλt , uM (ϵ̄)) ∀δλt ∈ Lt, , (34b)
(k)
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ + d (λk , δξ ) =0 ∀δξ ∈ X , (34c)
(k)
d (δλk , ξ ) =0 ∀δλk ∈ Lk, , (34d)
⟨−κ(ϵ[u], k) ⋆ δk⟩ + ⟨κ di ⋆ δk⟩ − ⟨χ[ξ ] ⋆ δk⟩ =0 ∀δk ∈ K , (34e)
∂φ ∗
 
⟨k ⋆ δκ di ⟩ − ∆t di (κ di ) ⋆ δκ di = ⟨ nk ⋆ δκ di ⟩ ∀δκ di ∈ K (34f)
∂κ
where the solution and test spaces are given as
  
U = u ∈ H1 (Ω ), u dS = 0 (35a)
Γ
X = [H1div (Ω )] M (35b)
+ −
Fig. 1. (a) SVE in 2D with “image” (Γ ) and “mirror” (Γ ) boundaries. (b) Possible FE-discretizations of Ω (representations of u, ξ , k, κ di )
+
and of Γ (representations of Lagrange multiplier fields λt , λk ).
Source: Figure from [39].
K = [L2 (Ω )] M (35c)

Lt, = [L2 (Γ+ )]3 , Lk, = [L2 (Γ+ )] M . (35d)
The Lagrange multiplier fields λt and λk are depicted schematically in Fig. 1(b) in a FEM context.
Remark. Although a quite general setting of the SVE-problem based on (weak) microperiodicity has been
formulated, this general problem is never solved in practice in this paper while exploiting FE-approximation of the
Lagrangian multiplier fields. Rather, to meet the purpose in the paper, it suffices to consider combinations of the
two special cases defined as Dirichlet and Neumann conditions. Moreover, the common choice of strongly periodic
boundary conditions (which do not require any Lagrange multipliers in the operational setting) cannot be bound by
such combinations of boundary conditions and is, therefore, omitted.
3.3. Incremental SVE-potential—Variational problem
The volume-specific incremental π is the one already defined in (20); hence, the “internal” SVE-potential Πint ,
which is pertinent to a specific time-increment (whereby the arguments of Πint represent values at the endpoint), is
given as
Πint (u, ξ , k, κ di ) = ⟨π(ϵ[u], k, ξ , χ[ξ ], κ di )⟩ (36)
whereas Πext = 0. However, due to the jump terms in (34), the total incremental potential Π becomes
(u) (k)
Π (u, λt , ξ , λk , k, κ di ) = Πint (u, ξ , k, κ di ) − d (λt , u − uM (ϵ̄)) + d (λk , ξ ). (37)
With the abbreviation z = (u, λt , ξ , λk , k, κ di ), the system (34) represents the stationarity conditions corresponding
to the saddle-point problem [Appendix A]
 
z = arg min max max min min max Π (ẑ) . (38)
û∈U λ̂t ∈Lt, ξ̂ ∈X λ̂k ∈Lk, k̂∈K κ̂ di ∈K
It should be noticed that the solution z{ϵ̄} depends implicitly on the “control” variable ϵ̄. Upon using (34b) and (34d),
we conclude that the stationarity solution satisfies the condition
(u) (k)
d (λt {ϵ̄}, u{ϵ̄} − uM (ϵ̄)) = 0, d (λk {ϵ̄}, ξ {ϵ̄}) = 0. (39)
The stationary point thus takes the value

def
w̄ {ϵ̄} = Π (ϵ̄; z{ϵ̄}) = Πint (z̃{ϵ̄}) (40)
where we introduced the abbreviated notation z̃ = (u, ξ , k, κ di )
for the “reduced” solution excluding the Lagrangian
multipliers that are pertinent to the RVE-boundary.
We may now use (40), the definition of Πint (z̃{ϵ̄}) in (36) and the stationarity conditions (34) in order to show
[Appendix B] that
dw̄ {ϵ̄} = ⟨σ {ϵ̄}⟩ : dϵ̄ = σ̄ : dϵ̄. (41)
We thus conclude that w̄ {ϵ̄} is the incremental “pseudo-elastic strain energy” that serves as a potential for σ̄ for the
selected SVE (of finite size), i.e.
dw̄ {ϵ̄}
σ̄ {ϵ̄} = . (42)
dϵ̄
3.4. Dirichlet and Neumann problems for a single SVE—Variational properties and bounds
3.4.1. Dirichlet boundary condition

It is possible to identify the problem with Dirichlet boundary conditions on the displacement field from the
canonical format upon restricting the spaces U and X , i.e. introducing UD 9 D
⊆ U and X ⊆ X defined as
= {u ∈ U : ∃ϵ̂ ∈ Rsym s. t. u = ϵ̂ · [x − x̄] on Γ }

UD 3×3
(43)
 
def 1
X = ξ ∈ X : ∃χ̂ ∈ R s. t. p[ξ ] = ξ · n = χ̂ ⊗ [x − x̄] · n on Γ
D M
(44)
3
while LD D D
t, = Lt, , Lk, = Lk, and K = K are left unrestricted. By the choice (43) and (44) the constraint
equations (34b) and (34d) will be satisfied only for the choice ϵ̂ = ϵ̄ and χ̂ = 0.
Remark. The choice (43) and (44) does not represent the “operational” format of the SVE-problem that is suitable
for algorithmic implementation; such a format is described subsequently in Section 4.1.
3.4.2. Neumann boundary condition

The problem with Neumann boundary conditions is obtained from the canonical format upon restricting the spaces
Lt, and Lk, , i.e. introducing LN N
t, ⊆ Lt, and Lk, ⊆ Lk, defined as
t, = {λt ∈ Lt, : ∃σ̂ ∈ Rsym s.t. λt = σ̂ · n on Γ }

LN 3×3
(45)
k, = {λk ∈ Lk, : ∃k̂ ∈ R s.t. λk = k̂ on Γ }

LN M
(46)
while UN N N
= U , X = X and K = K are left unrestricted. Clearly, the choice in (45) and (46) restricts the
tractions to be piecewise constant on each of the three positive boundary faces of the SVE-cube, whereas the internal
variables take a constant value on all (the positive) boundary faces.
Remark. The Neumann condition represents the weakest possible way of enforcing the micro-periodicity condition
(31). Here it is considered as a model assumption; however, it is also possible to view this choice as the crudest
possible FE-approximation of the traction field, cf. Larsson et al. [39].
3.4.3. Mixed Dirichlet–Neumann boundary conditions

In the previous subsections, we considered the “pure” Dirichlet and Neumann problems in the sense that the same
type of boundary conditions was applied to both balance equations of interest. However, to complete the picture
9 For a cubic/square SVE, we obtain x̄ = ⟨x⟩ 1  D

= |Γ | Γ x dS, which guarantees U ⊆ U .
with regard to the appropriate choice of prolongation conditions and to pave the way for estimating bounds on the
“incremental macroscale strain energy”, we shall consider the more general case of arbitrarily “mixed” conditions
in the sense that the Dirichlet or Neumann boundary condition is applied to either of the two balance equations. For
brevity, we introduce the notation (uN-ξ D), as an example, for “Neumann condition on the displacement and Dirichlet
condition on the micro-stress”.
We are now in the position to establish a compact format of both the “pure” Dirichlet and Neumann problems (as
discussed above) and the “mixed” Dirichlet/Neumann problems upon introducing the generic space
def
= U × Lt, × X × Lk, × K × K ,
ZXY X X Y Y
X,Y = D or N. (47)
Hence, the RVE-problem (uX-ξ Y) for given (prescribed) macroscopic ϵ̄ may be formulated in condensed form for
any combination of boundary conditions as follows: Find z XY ∈ ZXY
such that
(Π )′ (z XY ; δz) = 0 ,
∀δz ∈ ZXY X,Y = D or N. (48)
The stationary point represents a saddle-point of Π in the following sense:
 
z XY = arg min max max min min max Π (ẑ) . (49)
û∈UX
λ̂t ∈Lt, ξ̂ ∈X λ̂k ∈Lk, k̂∈K κ̂ ∈K
X Y Y di
3.4.4. Upper and lower bounds

It is of interest to assess the choice of boundary conditions on the SVE in terms of expected upper and lower bounds
on the macroscale stiffness properties. It will be shown below that the mixed boundary conditions (uN-ξ D) represent
a lower bound, whereas (uD-ξ N) represent an upper bound. Indeed, these conclusions based on theoretical arguments
are verified (at least in part) by the computational results shown in Section 6.
We recall from (40) that, due to the exact fulfillment of the constraint equations (corresponding to the Lagrangian
multipliers λt and λk at the saddle-point), we obtain
w̄
XY
{ϵ̄} = Π (z XY {ϵ̄}) = Πint (z̃ XY {ϵ̄}) (50)
where Πint is the reduced functional defined in (36). We shall then make explicit use of the properties
⊆ U ,
UD ⊆ X , t, ⊆ Lt, ,
N
XD N
LN D
LN D
k, ⊆ Lk, (51)
in order to construct the following inequalities for any given value ϵ̄:
min max max min min max
w̄
ND
= Π (ẑ)
û ∈ UN λ̂t ∈ LN
t, ξ̂ ∈ XD λ̂k ∈ LD
k, k̂ κ̂ di
≤ Π (ẑ) = w̄
NN
t, ξ̂ ∈ XN λ̂k ∈ LN
k, k̂ κ̂ di
≤ Π (ẑ) = w̄
DN
,
û ∈ UD λ̂t ∈ LD
k, k̂ κ̂ di
and
w̄
ND
= Π (ẑ)
k, k̂ κ̂ di
≤ Π (ẑ) = w̄
DD
k, k̂ κ̂ di
≤ Π (ẑ) = w̄
DN
.
k, k̂ κ̂ di
We may now conveniently summarize the bounds:
w̄
ND
≤ w̄
NN
≤ w̄
DN
, w̄
ND
≤ w̄
DD
≤ w̄
DN
. (52)
NN and w̄ DD ; only that they are both bounded from
In particular, we note that it is not possible (in general) to order w̄
below by w̄ and from above by w̄ .
ND DN
Remark. The bounds in (52) hold for the “apparent” response that is obtained for any given SVE (given realization of
a random microstructure), with finite size L < ∞. They hold for the continuous solution and the FE-approximation.
However, values for a single finite-sized SVE do not represent bounds on the true “effective” energy w̄{ϵ̄} = w̄ {ϵ̄}
for L = ∞.
4. Operational variational formats
4.1. SVE-problem with Dirichlet boundary conditions (DD-problem)
The operational format with Dirichlet boundary conditions is conveniently obtained upon introduction of the
spaces
UD,0
= {u ∈ U : u = 0 on Γ }, (53)
XD,0
= {ξ ∈ X : p[ξ ] = ξ · n = 0 on Γ }. (54)
We first conclude that
(u)
d (λt , δu) = 0 ∀δu ∈ UD,0
,
(k) (55)
d (λk , δξ ) = 0 ∀δξ ∈ XD,0
.
Moreover, the constraints in (34b) and (34d) are satisfied identically a priori for u = uM + uµ with uµ ∈ U D,0
and
D,0
for ξ ∈ X , respectively. As a result, the operational format of the DD-problem reads: For given value of ϵ̄, find the
fields uµ ∈ UD,0 D,0
, ξ ∈ X , k ∈ K , and κ ∈ K that solve the system
di
⟨σ (ϵ̄ + ϵ[uµ ], k) : ϵ[δu]⟩ = 0 D,0

∀δu ∈ U , (56a)
D,0
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ = 0 ∀δξ ∈ X , (56b)
⟨−κ(ϵ̄ + ϵ[uµ ], k) ⋆ δk⟩ + ⟨κ di ⋆ δk⟩ − ⟨χ[ξ ] ⋆ δk⟩ = 0 ∀δk ∈ K , (56c)
∂φ ∗
 
⟨k ⋆ δκ di ⟩ − ∆t di (κ di ) ⋆ δκ di = ⟨ nk ⋆ δκ di ⟩ ∀δκ di ∈ K . (56d)
∂κ
When the solution has been found the homogenized stress σ̄ is computed in a “post-processing step”: σ̄ = ⟨σ ⟩ =
def
|Ω | Γ t ⊗ [x − x̄] dS with t = σ · n.
1

4.2. SVE-problem with Neumann boundary conditions (NN-problem)
From the construction of the spaces LN N

t, and Lk, we note the identities
(u)
d (σ̂ · n, u) = σ̂ : ⟨⟨[[u]] ⊗ n⟩⟩ = σ̂ : ⟨ϵ[u]⟩ ∀σ̂ ∈ R3×3
sym , ∀u ∈ U , (57)
(k)
d (k̂, ξ ) = k̂ ⋆ ⟨⟨[[ξ ]] · n⟩⟩ = k̂ ⋆ ⟨χ[ξ ]⟩ ∀k̂ ∈ R N , ∀ξ ∈ X , (58)
where we introduced the notation for surface average on Γ+ :


def 1
⟨⟨y⟩⟩ = y dV. (59)
|Ω | Γ+
Remark. The last identities in (57) and (58), which are obtained using the divergence theorem, are valid only when
UN N
and X are sufficiently smooth.
, σ̂ ∈ Rsym ,
The operational (standard) format of the NN-problem then becomes: For given value of ϵ̄, find u ∈ UN 3×3
ξ ∈ XN
, k̂ ∈ R , k ∈ K , and κ ∈ K that solve the system
N di
⟨σ (ϵ[u], k) : ϵ[δu]⟩ − σ̂ : ⟨⟨[[δu]] ⊗ n⟩⟩ = 0 ,

∀δu ∈ UN (60a)
−δ σ̂ : ⟨⟨[[u]] ⊗ n⟩⟩ = −δ σ̂ : ϵ̄ ∀δ σ̂ ∈ R3×3
sym , (60b)
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ + k̂ ⋆ ⟨⟨[[δξ ]] · n⟩⟩ = 0 ∀δξ ∈ XN , (60c)
δ k̂ ⋆ ⟨⟨[[ξ ]] · n⟩⟩ = 0 ∀δ k̂ ∈ R ,
N
(60d)
⟨−κ(ϵ[u], k) ⋆ δk⟩ + ⟨κ di ⋆ δk⟩ − ⟨χ [ξ ] ⋆ δk⟩ = 0 ∀δk ∈ K , (60e)
∂φ ∗
 
⟨k ⋆ δκ di ⟩ − ∆t di (κ di ) ⋆ δκ di = ⟨ nk ⋆ δκ di ⟩ ∀δκ di ∈ K . (60f)
∂κ
It is of interest to give a “physical” interpretation of σ̂ and k̂.

def
(i) Firstly, testing (60a) with δu = δ ĥ · [x − x̄] gives the condition that σ̂ must take the value σ̂ = ⟨σ ⟩ = σ̄ .
(ii) Secondly, choosing δξ = δ χ̂ ⊗ 13 [x − x̄], where δ χ̂ ∈ R N is constant, and inserting into (60c), we obtain
 
1
k̂ = ⟨k⟩ + g · [x − x̄] . (61)
3
In the special case that the field k = k 0 is homogeneous within Ω , then it follows directly from (61) that k̂ = k 0 ,
which is the expected result.
4.3. SVE-problem with mixed Dirichlet–Neumann boundary conditions (DN- and ND-problems)
Regarding the appropriate choice of boundary conditions on the SVE, the DN- and ND-problems formulations are
of particular interest since the corresponding solutions represent upper and lower bounds, respectively, as discussed
above. The operational format of these problems is, therefore, given subsequently.
The operational format with Dirichlet boundary conditions on the standard equilibrium equation and Neumann
boundary condition on the micro-equilibrium equation reads: For given value of ϵ̄, find the fields uµ ∈ U
D,0
, ξ ∈ XN
,
k̂ ∈ R , k ∈ K , and κ ∈ K that solve the system
N di
⟨σ (ϵ̄ + ϵ[uµ ], k) : ϵ[δu]⟩ = 0 ∀δu ∈ UD,0

, (62a)
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ + k̂ ⋆ ⟨⟨[[δξ ]] · n⟩⟩ = 0 ∀δξ ∈ ,
XN (62b)
δ k̂ ⋆ ⟨⟨[[ξ ]] · n⟩⟩ = 0 ∀δ k̂ ∈ R ,N
(62c)
µ
⟨−κ(ϵ̄ + ϵ[u ], k) ⋆ δk⟩ + ⟨κ ⋆ δk⟩ − ⟨χ[ξ ] ⋆ δk⟩ = 0
di
∀δk ∈ K , (62d)
∂φ ∗
 
⟨k ⋆ δκ di ⟩ − ∆t di (κ di ) ⋆ δκ di = ⟨ nk ⋆ δκ di ⟩ ∀δκ di ∈ K . (62e)
∂κ
When the solution has been found the homogenized stress σ̄ is computed in a “post-processing step”: σ̄ = ⟨σ ⟩ =
def
|Ω | Γ t ⊗ [x − x̄] dS with t = σ · n.
1

Next, the operational format with Neumann boundary conditions on the standard equilibrium equation and Dirichlet
boundary condition on the micro-equilibrium equation reads: For given value of ϵ̄, find the fields u ∈ UN
, σ̂ ∈ Rsym ,
3×3
ξ ∈ XD,0
, k ∈ K , and κ ∈ K that solve the system
di
⟨σ (ϵ[u], k) : ϵ[δu]⟩ − σ̂ : ⟨⟨[[δu]] ⊗ n⟩⟩ = 0 ,

∀δu ∈ UN (63a)
−δ σ̂ : ⟨⟨[[u]] ⊗ n⟩⟩ = −δ σ̂ : ϵ̄ ∀δ σ̂ ∈ sym ,
R3×3 (63b)
−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ = 0 ,

∀δξ ∈ XN (63c)
⟨−κ(ϵ[u], k) ⋆ δk⟩ + ⟨κ ⋆ δk⟩ − ⟨χ[ξ ] ⋆ δk⟩ = 0
di
∀δk ∈ K , (63d)
∂φ ∗ di
 
⟨k ⋆ δκ ⟩ − ∆t di (κ ) ⋆ δκ
di di
= ⟨ nk ⋆ δκ di ⟩ ∀δκ di ∈ K . (63e)
∂κ
4.4. FE-approximation
Finally in this Section, we make a few remarks on the adopted FE-approximation. To be specific, we thereby refer
to the DD-problem formulation in (56).10 FE-spaces for the various solution fields are introduced as follows on an
element triangulation (in 2D) of Ω :
• U,h ∈ U is the space of p.w. linear displacements, i.e. ϵ[uh ] is constant within any given element.
def
• X,h ∈ X is the space of p.w. linear micro-stresses. In particular, this means that χ h = ξ h · ∇ is constant within
any given element.
• K,h ∈ K is the space of p.w. constants on each element i = 1, 2, . . . , N E L.
Since δk ∈ K,h , we conclude from (56c) that this equation must hold for each individual element. Hence, we may
solve for κ idi , associated with a particular element, to obtain
κ idi = κ i (ϵ i , k i ) + χ i [ξ ], i = 1, 2, . . . , N E L (64)
where the gradient effect enters only via the nonlocal character of χ i [ξ ]. Upon inserting (64) into (56d), which must
hold for each δκ idi ∈ K,h that represents a constant value of δκ di , we conclude that the local residual must vanish;
hence,
∂φ ∗
k i − ∆t (κ (ϵ i , k i ) + χ i [ξ ]) − nk i = 0, i = 1, 2, . . . , N E L . (65)
∂κ di i
Clearly, (65) reminds strongly about the “local constitutive driver” for the conventional non-gradient theory. Indeed,
this feature is considered a significant advantage of (i) the particular choice of “dual mixed” variational format
combined with (ii) the particular choice of FE-spaces adopted here.
5. Homogenization of polycrystal response
5.1. Preliminaries
One important application of the present homogenization strategy is the prediction of polycrystal response,
whereby the grains will interact via their boundaries. Fig. 2 shows a typical polycrystal with grain boundaries in
2D. Clearly, for a metallic material the obvious candidate for mesoscale modeling is crystal (visco)plasticity, and
the grain boundary interaction is controlled by (among other factors) the crystal orientation misfit. However, in
this paper we highlight the generic features of the homogenization strategy, and we therefore choose to introduce
strongly simplifying assumptions: (i) Standard continuum viscoplasticity is assumed for the grains, (ii) grain boundary
interaction is modeled via a simple linear relation for “dislocation resistance”, as proposed in the literature on metal
polycrystals, cf. Kouznetsova et al. [32], Ekh et al. [27].
5.2. Prototype model for grains: Bingham’s viscoplasticity model with gradient-enhanced linear isotropic hardening
We consider a standard viscoplasticity model of the Bingham type with gradient-enhanced isotropic hardening,
originally suggested by Svedberg and Runesson [6], to represent the mesoscale material response. The model is
defined by two internal variables k = {ϵ p , k}, where ϵ p is the plastic (part of) strain and k is the variable that represents
10 Similar arguments apply to the other SVE-problems.

Fig. 2. SVE with polycrystal mesostructure (used in the numerical example in Section 6.2). Only the simple “microhard” condition, k = 0, is
adopted as interface condition between grains in the numerical examples.
def
isotropic hardening. The elastic strain is then given as ϵ e (ϵ, ϵ p ) = ϵ − ϵ p . The semi-dual free energy is proposed as
1   e 2 2  1 loc 2 1
ϕ(ϵ e (ϵ, ϵ p ), k, ξ ) = 3G e(ϵ ) + K v(ϵ e ) + H k − 2 gra |ξ |2 .

(66)
2
  2   2l H  
ψ loc ψ ∗gra
def
 (66) we used the effective elastic strain e(ϵ ) and the volumetric elastic strain v(ϵ ) via the definitions e(ϵ) =
In e e
def def
3 |ϵ dev | and v(ϵ) = ϵ : I, and the decomposition ϵ = ϵ dev + 3 v(ϵ)I. Moreover, we used the notation g = ∇k.
2 1
The material parameters are the (state-independent but generally spatially inhomogeneous) elastic moduli G, K , the
standard (local) hardening modulus H loc , the “gradient hardening” modulus H gra and the intrinsic length l.
From (66) we obtain the constitutive state equations as follows:
σ = σ dev + σm I with σ dev = 2Gϵ edev , σm = K v(ϵ e ) (67)
def ∂ϕ def ∂ϕ def ∂ϕ 1
σp = − [=σ ], κ =− = −H loc k, g=− = 2 gra ξ . (68)
∂ϵ p ∂k ∂ξ l H
Comparing with the generic equations (64), we obtain the dissipative stresses as
σ di = σ (ϵ, ϵ p ) (69)
κ di
= κ(k) + χ = −H loc
k+χ (70)
def
where χ = ξ · ∇. The loading function expressing isotropic hardening is given as
F(σ di , κ di ) = s(σ di ) − σY − κ di
= s(σ (ϵ, ϵ p )) − σY + H loc k − χ = 3Ge(ϵ e ) − σY + H loc k − χ
def
= f (ϵ e , k, χ ) (71)

def
where s(σ ) = 32 |σ dev | is the effective stress, whereas σY is the quasistatic yield stress (material parameter). We
also choose the simple “overstress” model
1 1
φ∗ = η(F) with η(F) = ⟨F⟩2 (72)
t∗ 2C
where C is a reference “stress”, and we obtain
∂φ ∗ 3σ dev ϵ edev
= λ = λ (73)
∂σ di 2s(σ ) e(ϵ e )
∂φ ∗
= −λ (74)
∂κ di
Table 1
Material parameter values.
E [MPa] ν [–] l [mm] H gra [MPa] H loc [MPa] σY [MPa] t∗ [s]

200 · 103 0.3 10−2 107 104 103 5 · 104
where the “plastic multiplier” state function is given as

1
λ= ⟨ f (e(ϵ e ), k, χ )⟩. (75)
t∗ σY
Next, we derive the time-discrete formulation. The local residual equations (65) for the time-updated values of the
def
two internal variables ϵ p and k have the solution in terms of ϵ e,tr and γ = λ∆t:
ϵ edev ϵ e,tr def µ
ϵ p = nϵ p + γ = ϵ
n p
+ γ dev
with ϵ e,tr = nϵ e + ∆ϵ̄
 + ∆ϵ  (76)
e(ϵ ) e e(ϵ )
e,tr
∆ϵ
k = nk − γ . (77)
The scalar field γ = γ (x) is the solution of the “local-type” constitutive equation
t∗
σY γ − ⟨ f (ϵ e,tr , γ , χ )⟩ = 0 (78)
∆t
where
def
f (ϵ e,tr , γ , χ ) = f tr (ϵ e,tr ) − h p γ − χ with h p = 3G + H loc (79)
and where the “trial value” of the loading function f corresponding to pure elastic incremental response is given as
def
f tr (ϵ e,tr ) = 3Ge(ϵ e,tr ) − σY + H loc nk. (80)
Since this is a viscoplastic model, the distinction between active/inactive region (in a given spatial point) is determined
by the trial value ϵ e,tr for any given value of χ . Clearly, the presence of χ ̸= 0 is what makes the gradient theory
different from standard local viscoplasticity in the context of the constitutive algorithm. Inelastic loading at any spatial
point x ∈ Ω is characterized by F > 0 (or f > 0), whereby (78) can be rewritten as
t∗
def f tr (ϵ e,tr ) − χ
h vp γ + χ = f tr (ϵ e,tr ) with h vp = h p + σY , ⇒ γ = . (81)
∆t h vp
Hence, when the gradient term χ is known in the considered spatial point, it is possible to obtain the solution γ from
(81) in the same fashion as for the local (non-gradient) theory. Moreover, in the special case that l = 0, i.e. local
theory is presumed, then the standard solution is obtained without “smearing” of the plastic zone.
6. Numerical investigation
6.1. 1D bar model
To illustrate the influence of the different combinations of Dirichlet and Neumann boundary conditions on the
solution of the SVE-problem, we first investigate a simple 1D bar consisting of three “grains” with different lengths
(0.2, 1 and 1 µm) according to the Fig. 3. The adopted material data are shown in Table 1.
The homogenized stress–strain responses for the bar model with different combinations of boundary conditions
are shown in Fig. 4. We observe that the microhard condition (k = 0) gives the stiffest response. As to the boundary
conditions on the SVE, we note that the stiffest response is obtained from the Neumann boundary condition on ξ ,
whereas the weakest response is obtained from the Dirichlet condition on ξ . The interface condition between the
grains is chosen as microhard for all three cases. The derived theoretical bounds are confirmed by the numerical
results in Fig. 4.
Fig. 3. Bar model consisting of three grains. Dimensions are given in µm.
Fig. 4. Homogenized stress σ̄ vs strain ϵ̄. Results are shown for different combinations of Dirichlet/Neumann boundary conditions. (All interfaces
are chosen as microhard.)
Fig. 5. Distribution of hardening variable k and micro-stress ξ along the bar for ϵ̄ = 0.05. Results are shown for different combinations of
Dirichlet/Neumann boundary conditions.
The distribution of k along the bar for the three different boundary conditions are shown in Fig. 5 (left) together
with the distribution of ξ (right). It is readily shown here that, for the 1D-problem setting, the solution for the Neumann
condition has the property that the micro-stress ξ (and thereby dk/dx) is equal at x = 0 and x = L. The Dirichlet
condition, on the other hand, gives the result that the hardening variable k is equal at x = 0 and x = L.
6.2. 2D grain model
Next, we investigate two different examples of 2D grain models (plane strain assumption). Both models are
subjected to simple shear loading with a loading time of 10 [s]. Similarly to the 1D example, we assume the microhard
condition (k = 0) for the interface between grains. The chosen set of material parameters are still those given in
Table 1.
The first SVE is composed of 4 grains with the total size 2 × 2 [µm], whereas the second SVE is composed of
25 grains with the total size 5 × 5 [µm]. This means that the average grain-size in the two examples is the same. We
Fig. 6. Homogenized shear stress τ̄ vs shear deformation γ̄ for two different SVE-sizes: 4 grains (left figure) and 25 grains (right figure). Results
are shown for different combinations of Dirichlet/Neumann boundary conditions on the SVE.
can conclude from Fig. 6 that the derived theoretical bounds are confirmed. The stiffest response is obtained from
the Dirichlet condition on u and the Neumann condition on ξ , whereas the weakest response is obtained from the
Neumann condition on u and the Dirichlet condition on ξ . Upon comparing the results for the 4 grain example with
the result for the 25 grain SVE, we can conclude that the significance of the choice of boundary condition decreases
with size of the SVE (as desired). In particular, this is true for the microstress balance equation. [Although not shown
in the paper, results for even larger SVE-sizes confirm that the macroscale stress does, indeed, converge and become
less influenced by the particular choice of boundary conditions.] Finally, Fig. 7 shows the typical variation of the
hardening variable k, clearly demonstrating the influence of the chosen “microhard” inter-grain condition. What is
less visible is the difference in boundary displacement between Dirichlet (uD) and Neumann (uN) conditions.
7. Conclusions and outlook
In this paper we have applied variationally consistent selective homogenization of subscale gradient-enhanced
viscoplasticity, whereby the macroscale model becomes a standard continuum. In other words, the field equations
representing gradient effects live only on the subscale. Since the main purpose of the paper is to provide a proof
of concept, it was considered sufficient to choose a gradient-enhanced continuum model without crystal features to
represent the subscale constitutive response. From an algorithmic viewpoint it is advantageous, as shown in Ekh
et al. [27], to base the formulation of the SVE-problem on the mixed variational principle exploiting the displacement
and the microstress fields. The corresponding variational format was coined “semi-dual” from its construction that
employs a partial Legendre transformation.
A new result is that the macroscale stress in a given time-increment (after time integration) is derivable from
a macroscopic pseudoelastic strain energy pertinent to a given Statistical Volume Element (SVE). In its turn, this
pseudo-elastic energy represents the saddle-point of an incremental SVE-potential, and the extremal properties of this
SVE-potential were shown to form the basis for establishing upper and lower bounds on the incremental pseudo-
elastic energy. Indeed, it was shown that the specific combination of boundary conditions (uD-ξ N) produces the upper
bound, whereas the combination of boundary conditions (uN-ξ D) gives the lower bound. The validity of these bounds
are demonstrated heuristically in the simple numerical examples; however, the results are consistent.
As to future developments, it is desirable to adopt more realistic subscale modeling of the intragrain response
(gradient crystal plasticity) as well as the intergrain interaction (grain interface resistance of dislocation flux and
decohesion). For the numerical results in this paper, we simply assumed the microhard condition (k = 0) in order
to provoke subscale heterogeneity. The modeling of intergrain interaction is a challenging research topic in its own
right, e.g. Gurtin [28], Wulfinghoff et al. [40], Gottschalk et al. [30]. Another issue of generic importance is to obtain
bounds on the macroscale stress vs. strain relation rather than restricting to bounds on the pseudoelastic energy vs.
strain relation for a single time increment (as in this paper). The numerical results seem to support the heuristic
presumption that the incremental characteristics carry over to the entire stress–strain response function for any given
history of the controlling macroscale strain; however, a formal proof is lacking.
Fig. 7. Snapshots of hardening variable k for different boundary conditions on the SVE with 25 grains. Results are shown for different combinations
of Dirichlet/Neumann boundary conditions.
Acknowledgment
Financial support from the Swedish Research Council, grant no. 621-2013-3091, is gratefully acknowledged.
Appendix A. Extremal properties of time-incremental problem in the semi-dual format
The task is to verify the extremal properties
min max min max Π (û, ξ̂ , k̂, κ̂ di ) (A.1)

û ξ̂ k̂ κ̂ di
where

Π (u, ξ , k, κ ) = di
π(ϵ[u], k, ξ , χ[ξ ], κ di ) dV − l (u) (u − nu) − l (ξ ) (ξ ) (A.2)
Ω
with
π(ϵ, k, ξ , χ, κ di ) = ϕ(ϵ, k, ξ ) − nϕ + κ di ⋆ [k − nk] − χ ⋆ k − ∆t φ ∗ (κ di ). (A.3)
As a preliminary, we note the coercivity of the diagonal forms of the Hessian of Π as follows:
∂ 2ϕ

′′
Πuu (•; du, du) = ϵ[ du] : : ϵ[ du] dV > 0 ∀ du ̸= 0 (A.4a)
Ω ∂ϵ ⊗ ∂ϵ
∂ 2ϕ

Πξ′′ξ (•; dξ , dξ ) = dξ ⋆ ⋆ dξ dV < 0 ∀ dξ ̸= 0 (A.4b)
Ω ∂ξ [∂ξ ]T
∂ 2ϕ

′′
Πkk (•; dk, dk) = dk ⋆⋆ dk dV > 0 ∀ dk ̸= 0 (A.4c)
Ω ∂k[∂k]T
∂ 2φ∗
  
′′
Πκκ (•; dκ di , dκ di ) = − dκ di ⋆ ∆t di di T ⋆ dκ di dV < 0 ∀ dκ di ̸= 0 (A.4d)
Ω ∂κ [∂κ ]
that is, Π is convex in u, k and concave in ξ , κ di . Hence due to the strict coercivity of each diagonal form of the
Hessian (with alternating sign) the extremal properties in (A.1) follow naturally. Below, we give the detailed proof
including the pertinent explicit sensitivity relations.
Step 1: For fixed u, ξ , k, consider
max Π (u, ξ , k, κ̂ di ). (A.5)

κ̂ di
• Stationarity (1st variation)

0 = Πκ′ (•; δκ di )
∂φ ∗ di
  
= k − k − ∆t di (κ ) ⋆ δκ di dV
n
∀δκ di (A.6)
Ω ∂κ
which can be solved for κ di = κ di {k}.
• 2nd variation
d2κ Π = Πκκ
′′
(•; dκ di , dκ di ). (A.7)
From (A.4d), we conclude that d2κ Π < 0 for each dκ di ̸= 0, and the max-property is verified.
def
• Sensitivity: From the implicit relation κ di {k} valid at the stationary point, we define the sensitivity ( dκ di )k =
(κ di )′k {k; dk} that is the solution of the sensitivity (linearized) problem
′′
0 = Πκk (•; δκ di , dk) + Πκκ
′′
(•; δκ di , ( dκ di )k ) ∀δκ di . (A.8)
Eq. (A.8) is solvable for ( dκ di )k for given dk, since the tangent form ′′ (•; δκ di ,
−Πκκ dκ di ) is coercive.
Step 2: For fixed u, ξ , consider
(κ)
min Π (u, ξ , k̂) (A.9)
k̂
where
(κ) def
Π (u, ξ , k) = Π (u, ξ , k, κ {k}).
di
(A.10)
def
• Stationarity (1st variation): Introduce notation (δκ di )k = (κ di )′k {k; δk}
(κ)
0 = Π ′k (•; δk)
= Πk′ (•; δk) + Πκ′ (•; (δκ di )k )
  
=0 from (A.6)
= Πk′ (•; δk)
  
= −κ(ϵ[u], k) + κ di {k} − χ[ξ ] ⋆ δk dV ∀δk, (A.11)
Ω
which can be solved for k = k{u, ξ }.
• 2nd variation
(κ) def (κ)
d2k Π = Π kk (•; dk, dk)
′′
= Πkk (•; dk, dk) + Πkκ

′′ ′′
(•; dk, ( dκ di )k )
= Πkk ′′
(•; dk, dk) − Πκκ
′′
(•; ( dκ di )k , ( dκ di )k }). (A.12)
′′ (•; δk, δκ di ) =
In order to obtain the last identity in (A.12), we used the symmetry of the Hessian of Π : Πkκ
Πκk (•; δκ , δk) and the sensitivity relation (A.8) to give the identity
′′ di
′′
Πkκ (•; dk, ( dκ di )k ) = −Πκκ
′′
(•; ( dκ di )k , ( dκ di )k ). (A.13)
We note that from (A.4c) that ′′ (•;
Πkk dk, dk) > 0 for ′′ (•; ( dκ di ) , ( dκ di ) ) ≥
dk ̸= 0, and from (A.4d) that −Πκκ k k
(κ)
0 for dk ̸= 0 [since either ( dκ di )k = 0 or ( dκ di )k ̸= 0 when dk ̸= 0]. We thus conclude that dk Π > 0 for each
2
dk ̸= 0, and the min-property is thereby verified.

def
• Sensitivity: From the implicit relation k{u, ξ }, valid at the stationary point, we define the sensitivities ( dk)u =
def
(k)′u {u, ξ ; du}, and ( dk)ξ = (k)′ξ {u, ξ ; dξ }. By exploiting the stationarity property (A.11), we may establish two
sensitivity equations as follows:
(κ) (κ)
ku (•; δk, kk (•; δk, ( dk)u )
′′ ′′
0=Π du)+ Π ∀δk, du (A.14)
(κ) (κ)
kξ (•; δk, dξ )+ Π kk (•; δk, ( dk)ξ ) ∀δk, dξ .
′′ ′′
0=Π (A.15)
(κ)
Eq. (A.14) is solvable for ( dk)u whereas Eq. (A.15) is solvable for ( dk)ξ since the tangent form Π ′′ (•; δk,
kk dk)
is coercive.
Step 3: For fixed u, consider
(kκ)
min Π (u, ξ̂ ) (A.16)
ξ̂
where
(kκ) def (κ)
Π (u, ξ ) = Π (u, ξ , k{u, ξ })
= Π (u, ξ , k{u, ξ }, κ di {k{u, ξ }}). (A.17)
def
• Stationarity (1st variation): Introduce the notation (δk)ξ = (k)′ξ {u, ξ ; δξ }
(kκ)
ξ (•; δξ )
′
0= Π
(κ) (κ)
= Π ′ξ (•; δξ ) + Π ′k (•; (δk)ξ )
  
=0 from (A.11)
= Πξ′ (•; δξ )
  
= −g(ξ )˙⋆δξ − k{u, ξ } ⋆ χ [δξ ] dV − l (ξ ) (δξ ) = 0 ∀δξ , (A.18)
Ω
which can be solved for ξ = ξ {u}.
• 2nd variation
(kκ) def (kκ)
d2ξ Π = Π ξ ξ (•; dξ , dξ })
′′
(κ) (κ)
ξ ξ (•; dξ , dξ })+ Π ξ k (•; dξ , ( dk)ξ )
′′ ′′
= Π
(κ) (κ)
ξ ξ (•; dξ , dξ })− Π kk (•; ( dk)ξ , ( dk)ξ )
′′ ′′
= Π
(κ)
= Πξ′′ξ (•; dξ , dξ }) − Π kk (•; ( dk)ξ , ( dk)ξ ) .
′′
(A.19)
  
≥0 from (A.12)
(κ) (κ) (κ)
In order to obtain the 3rd identity in (A.19), we used the symmetry of the Hessian of Π : Π ′′ (•; δk, δξ ) =Π
kξ
′′ (•; δξ , δk) and the sensitivity relation (A.15) to give the identity
ξk
(κ) (κ)
ξ k (•; dξ , ( dk)ξ ) = − Π kk (•; ( dk)ξ , ( dk)ξ ).
′′ ′′
Π (A.20)
(kκ)
We note from (A.4b) that Πξ′′ξ (•; dξ , dξ ) < 0 for dξ ̸= 0. We thus conclude that d2ξ Π < 0 for each dξ ̸= 0, and
the max-property is thereby verified.
• Sensitivity: From the implicit relation ξ {u}, valid at the stationary point, we define the sensitivity ( dξ )u =
(ξ )′u {u; du} as the solution of the sensitivity problem
(kκ) (kκ)
ξ u (•; δξ , ξ ξ (•; δξ , ( dξ )u ) ∀δξ , du.
′′ ′′
0= Π du)+ Π (A.21)
(kκ)
Eq. (A.21) is solvable for ( dξ )u since the tangent form Π ′′ (•; δξ ,
ξξ dξ ) is a coercive bilinear form.
Step 4: Consider
(ξ kκ)
min Π (û) (A.22)
û
where
(ξ kκ) def (kκ)
Π (u) = Π (u, ξ {u})
(κ)
= Π (u, ξ {u}, k{u, ξ {u}})
= Π (u, ξ {u}, k{u, ξ {u}}, κ di {k{u, ξ {u}}}). (A.23)
def
• Stationarity (1st variation): Introduce the notation (δξ )u = (ξ )′u {u; δu}
(ξ kκ)
u (•; δu)
′
0= Π
(kκ) (kκ)
u (•; δu) + ξ (•; (δξ )u )
′ ′
= Π Π
  
=0 from (A.18)
(κ) (κ)
= Π ′u (•; δu) + Π ′k (•; (δk)u )
  
=0 from (A.11)
= Πu′ (•; δu)

= σ (ϵ[u], k) : δϵ dV − l (u) (δu) = 0 ∀δu, (A.24)
Ω
which can be solved for u.
• 2nd variation
(ξ kκ) (ξ kκ)
d2u Π = Π uu (•;
′′
du, du)
(kκ) (kκ)
uu (•; uξ (•; du, ( dξ )u )
′′ ′′
= Π du, du)+ Π
(kκ) (kκ)
uu (•; ξ ξ (•; ( dξ )u , ( dξ )u )
′′ ′′
= Π du, du)− Π
(kκ)
′′
= Πuu (•; du, du) − Π ξ ξ (•; ( dξ )u , ( dξ )u ) .
′′
(A.25)
  
≤0 from (A.19)
(kκ) (kκ)
In order to obtain the 3rd identity result in (A.25), we used the symmetry of the Hessian of Π : Π
(kκ)
′′ (•; δξ , δu) =Π ′′ (•; δu, δξ ) and the sensitivity relation (A.21) to give the identity
ξu uξ
(kκ) (kκ)
uξ (•; du, ( dξ )u ) = − Π ξ ξ (•; ( dξ )u , ( dξ )u ).
′′ ′′
Π (A.26)
(ξ kκ)
′′ (•; du, du) > 0 for du ̸= 0. We thus conclude that d2
We note from (A.4a) that Πuu u Π > 0 for each du ̸= 0,
and the min-property is thereby verified.
Appendix B. Incremental pseudo-elastic strain energy
It was concluded in Section 3.3 that w̄ {ϵ̄} = Πint (z̃{ϵ̄}). Furthermore, Πint (z̃{ϵ̄}) does not depend on ϵ̄ explicitly
but only implicitly via z̃{ϵ̄}. We then obtain, upon differentiation w.r.t. ϵ̄
dw̄ {ϵ̄} = (Πint )′ (•; dz̃{ϵ̄})
= (Πint )′u (•; du{ϵ̄}) + (Πint )′ξ (•; dξ {ϵ̄}) + (Πint )′k (•; dk{ϵ̄}) + (Πint )′κ (•; dκ d {ϵ̄}) (B.1)
where the notation du{ϵ̄} = (u)′ {ϵ̄; dϵ̄}, etc. was introduced to represent sensitivities for a change in the control
variable ϵ̄.
Upon choosing the test functions δu = du − uM in (34a), δξ = dξ in (34c), δk = dk in (34e) and, finally,
δκ d = dκ d in (34f), we obtain the conditions
(u)
(Π )′u (•; du − uM ( dϵ̄)) = (Πint )′u (•; du − uM ( dϵ̄)) − d (λt , du − uM ( dϵ̄)) = 0 (B.2a)
(k)
(Π )′ξ (•; dξ ) = (Πint )′ξ (•; dξ ) + d (λk , dξ ) = 0 (B.2b)
(Π )′k (•; dk) = (Πint )′k (•; dk) = 0 (B.2c)
(Π )′κ (•; dκ d ) = (Πint )′κ (•; dκ d ) = 0. (B.2d)
However, upon differentiating (39) w.r.t. ϵ̄, we note the conditions
(u) (k)
d (λt {ϵ̄}, du − uM ( dϵ̄)) = 0, d (λk {ϵ̄}, dξ ) = 0 (B.3)
which can be combined with (B.2a), (B.2b) to give
(Πint )′u (•; du) = (Πint )′u (•; uM ( dϵ̄)) = ⟨σ : dϵ̄⟩ = ⟨σ ⟩ : dϵ̄ (B.4a)
(Πint )′ξ (•; dξ ) = 0 (B.4b)
where it was used that uM ( dϵ̄) = dϵ̄ · [x − x̄]. Hence, the expression in (B.1) reduces to
dw̄ {ϵ̄} = ⟨σ {ϵ̄}⟩ : dϵ̄ = σ̄ : dϵ̄ (B.5)
which is (41).
References
[1] H.M. Zbib, E. Aifantis, On the gradient-dependent theory of plasticity and shear banding, Acta Mech. 92 (1992) 209.
[2] N.A. Fleck, G.M. Muller, M.F. Ashby, J.W. Hutchinson, Strain gradient plasticity: theory and experiment, Acta Metall. Mater. 42 (1994)
475–487.
[3] G.A. Maugin, Internal variables and dissipative structures, J. Non-Equilib. Thermodyn. 15 (1990) 173–192.
[4] K.C. Valanis, A gradient theory of internal variables, Acta Mech. 116 (1996) 1.
[5] T. Svedberg, K. Runesson, A thermodynamically consistent theory of gradient-regularized plasticity coulpled to damage, Int. J. Plast. 13
(1997) 669–696.
[6] T. Svedberg, K. Runesson, An algorithm for gradient-regularized plasticity coupled to damage based on a dual mixed FE-formulation, Comput.
Methods Appl. Mech. Engrg. 161 (1998) 49–65.
[7] T. Liebe, P. Steinmann, Theory and numerics of a thermodynamically consistent framework for geometrically linear gradient plasticity,
Internat. J. Numer. Methods Engrg. 51 (2001) 1437–1467.
[8] T. Liebe, A. Menzel, P. Steinmann, Theory and numerics of a thermodynamically nonlinear gradient plasticity, Internat. J. Engrg. Sci. 41
(2003) 1603–1629.
[9] C. Miehe, A multi-field incremental variational framework for gradient-extended standard dissipative solids, J. Mech. Phys. Solids 59 (2011)
898–923.
[10] F.R. Welschinger, A Variational Framework for Gradient-Extended Dissipative Continua. Application to Damage Emchanics, Fracture, and
Plasticity (Ph.D. thesis), University of Stuttgart, Department of Mechanics, 2010.
[11] C. Miehe, F. Welschinger, M. Hofacker, Thermodynamically consistent phase-field models of fracture. Variational principles and multifield
FE implementations, Internat. J. Numer. Methods Engrg. 83 (2010) 1273–1311.
[12] C. Miehe, Variational gradient plasticity at finite strains. Part I: Mixed potentials for the evolution and update problems of gradient-extended
dissipative solids, Comput. Methods Appl. Mech. Engrg. 268 (2014) 677–703.
[13] S. Mauthe, C. Miehe, Variational gradient plasticity: Local–global updates, regularization and laminate microstructures in single crystals,
in: S. Conti, K. Hackl (Eds.), Analysis and Computation of Microstructure in Finite Plasticity, Lecture Notes in Applied and Computational
Mechanics, Springer, 2015.
[14] P. Cermelli, M.E. Gurtin, On the characterization of the geometrically necessary dislocations in finite plasticity, J. Mech. Phys. Solids 49
(2001) 1539–1568.
[15] A. Acharya, J.L. Bassani, A. Beaudoin, Geometrically necessary dislocations, hardening, and a simple gradient theory of crystal plasticity,
Scr. Mater. 48 (2003) 167–172.
[16] L.P. Evers, W.A.M. Brekelmans, M.G.D. Geers, Scale dependent crystal plasticity framework with dislocation density and grain boundary
effects, Int. J. Solids Struct. 41 (2004) 5209–5230.
[17] S. Yefimov, E. van der Giessen, Multiple slip in a strain-gradient plasticity model motivated by a statistical-mechanics description of
dislocations, Int. J. Solids Struct. 42 (2005) 3375–3394.
[18] V. Levkovitch, B. Svendsen, On the large-deformation- and continuum-based formulation of models for extended crystal plasticity, Int. J.
Solids Struct. 43 (2006) 7246–7267.
[19] N. Ohno, D. Okumura, Higher-order stress and grain size effects due to self-energy of geometrically necessary dislocations, J. Mech. Phys.
Solids 55 (2007) 1879–1898.
[20] U. Borg, A strain gradient crystal plasticity analysis of grain size effects in polycrystals, Eur. J. Mech. A Solids 26 (2007) 313–324.
[21] M. Kuroda, V. Tvergaard, On the formulations of higher-order strain gradient crystal plasticity models, J. Mech. Phys. Solids 56 (2008)
1591–1608.
[22] S. Bargmann, B.D. Reddy, Modeling of polycrystals using a gradient crystal plasticity theory that includes dissipative micro-stresses, Eur. J.
Mech. A Solids 20 (2011) 719–730.
[23] S. Bargmann, M. Ekh, K. Runesson, B. Svendsen, Modeling of polycrystals with gradient crystal plasticity: A comparison of strategies, Phil.
Mag. 90 (2010) 1263–1288.
[24] S. Bargmann, B. Svendsen, M. Ekh, An extended crystal plasticity model for latent hardening in polycrystals, Comput. Mech. 48 (2011)
631–648.
[25] L.H. Poh, R.H.J. Peerlings, M.G.D. Geers, S. Swaddiwudhipong, Homogenization towards a grain-size dependent plasticity theory for single
slip, J. Mech. Phys. Solids 61 (2013) 913–927.
[26] C. Miehe, S. Mauthe, F. Hildebrand, Variational gradient plasticity at finite strains. Part III: Local–global updates and regularization techniques
in multiplicative plasticity for single crystals, Comput. Methods Appl. Mech. Engrg. 268 (2014) 735–762.
[27] M. Ekh, M. Grymer, K. Runesson, T. Svedberg, Gradient crystal plasticity for the computational modeling of polycrystalline metals, Internat.
J. Numer. Methods Engrg. 72 (2007) 197–220.
[28] M.E. Gurtin, A theory of grain boundaries that accounts automatically for grain misorientation and grain-boundary orientation, J. Mech. Phys.
Solids 56 (2008) 640–662.
[29] M. Ekh, S. Bargmann, M. Grymer, Influence of grain boundary conditions on modeling of size-dependence in polycrystals, Acta Mech. 218
(2011) 103–113.
[30] D. Gottschalk, A. Mcbride, B.D. Reddy, A. Javili, P. Wriggers, C.B. Hirschberger, Computational and theoretical aspects of a grain-boundary
model that accounts for grain misorientation and grain-boundary orientation, Comput. Mater. Sci. 111 (2015) 443–459.
[31] P. Van Beers, G.G. McShane, V. Kouznetsova, M.G.D. Geers, Grain boundary interface mechanics in strain gradient crystal plasticity,
J. Mech. Phys. Solids 61 (12) (2013) 2659–2679.
[32] V. Kouznetsova, M.G.D. Geers, W.A.M. Brekelmans, Multi-scale constitutive modelling of heterogeneous materials with a gradient-enhanced
computational homogenization scheme, Internat. J. Numer. Methods Engrg. 54 (2002) 1235–1260.
[33] R. Larsson, S. Diebels, A second order homogenization procedure for multiscale analysis based on micropolar kinematics, Internat. J. Numer.
Methods Engrg. 69 (2007) 2485–2512.
[34] T.J.R. Hughes, G.R. Feijo, L. Mazzei, J.-B. Quincy, The variational multiscale method—a paradigm for computational mechanics, Comput.
Methods Appl. Mech. Engrg. 166 (1–2) (1998) 3–24.
[35] F. Larsson, K. Runesson, F. Su, Variationally consistent computational homogenization of transient heat flow, Internat. J. Numer. Methods
Engrg. 81 (2010) 1659–1686.
[36] Q.-S. Nguyen, On standard dissipative gradient models, Ann. Solid Struct. 1 (2010) 79–86.
[37] M.E. Gurtin, A gradient theory of small–deformation isotropic plasticity that accounts for the Burgers vector and for dissipation due to plastic
spin, J. Mech. Phys. Solids 52 (2004) 2545–2568.
[38] D. Hurtado, Multiscale Modeling of Microcrystalline Materials (Ph.D. thesis), California Institute of Technology, 2011.
[39] F. Larsson, K. Runesson, S. Saroukhani, R. Vafadari, Computational homogenization based on a weak format of micro-periodicity for RVE-
problems, Comput. Methods Appl. Mech. Engrg. 200 (2011) 11–26.
[40] S. Wulfinghoff, E. Bayerschen, T. Böhlke, A gradient plasticity grain boundary yeld theory, Int. J. Plast. 51 (2013) 33–46.

Computational Homogenization of Mesoscale Gradient Viscoplasticity

Uploaded by

Document Informationclick to expand document information

Document Informationclick to expand document information

Copyright:

Available Formats

Computational Homogenization of Mesoscale Gradient Viscoplasticity

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computational Homogenization of Mesoscale Gradient Viscoplasticity

Uploaded by

Copyright:

Available Formats

Available online at www.sciencedirect.

Comput. Methods Appl. Mech. Engrg. 317 (2017) 927–951

Computational homogenization of mesoscale gradient

Keywords: Selective homogenization; Gradient viscoplasticity; Bounds

2. A (sub)class of gradient-enhanced dissipative materials

2.1. Model framework—Thermodynamic balance equations

2.1.2. Canonical constitutive format

def ∂ψ ∂ψ loc def ∂ψ ∂ψ loc def ∂ψ ∂ψ gra

The stationarity condition gives the evolution rule for k expressed as

2.2. Time-discrete variational formulation—The semi-dual format

and we define the semi-dual free energy density as

Boundary conditions6 will be imposed on “displacements” and “tractions” as follows:

Next, we introduce the appropriate solution and test spaces

6 Subscript “pre” denotes “prescribed value”.

π(ϵ, k, ξ , χ, κ di ) = ϕ(ϵ, k, ξ ) − nϕ + κ di ⋆ [k − nk] − χ ⋆ k − ∆t φ ∗ (κ di ). (20)

min max min max Π (û, ξ̂ , k̂, κ̂ di ). (23)

∂ψ loc (ϵ, k{ϵ})

3. Variationally consistent homogenization

uM (x̄, x) = ū(x̄) + ϵ̄(x̄) · [x − x̄] for x ∈ Ω (28)

7 Henceforth, the argument x̄ is suppressed unless there is a risk of confusion.

3.2. SVE-problem—Weak enforcement of micro-(anti)periodicity

K = [L2 (Ω )] M (35c)

3.3. Incremental SVE-potential—Variational problem

Πint (u, ξ , k, κ di ) = ⟨π(ϵ[u], k, ξ , χ[ξ ], κ di )⟩ (36)

The stationary point thus takes the value

3.4.1. Dirichlet boundary condition

= {u ∈ U : ∃ϵ̂ ∈ Rsym s. t. u = ϵ̂ · [x − x̄] on Γ }

3.4.2. Neumann boundary condition

t, = {λt ∈ Lt, : ∃σ̂ ∈ Rsym s.t. λt = σ̂ · n on Γ }

k, = {λk ∈ Lk, : ∃k̂ ∈ R s.t. λk = k̂ on Γ }

3.4.3. Mixed Dirichlet–Neumann boundary conditions

9 For a cubic/square SVE, we obtain x̄ = ⟨x⟩ 1  D

3.4.4. Upper and lower bounds

We may now conveniently summarize the bounds:

4. Operational variational formats

4.1. SVE-problem with Dirichlet boundary conditions (DD-problem)

⟨σ (ϵ̄ + ϵ[uµ ], k) : ϵ[δu]⟩ = 0 D,0

4.2. SVE-problem with Neumann boundary conditions (NN-problem)

From the construction of the spaces LN N

where we introduced the notation for surface average on Γ+ :

⟨σ (ϵ[u], k) : ϵ[δu]⟩ − σ̂ : ⟨⟨[[δu]] ⊗ n⟩⟩ = 0 ,

It is of interest to give a “physical” interpretation of σ̂ and k̂.

⟨σ (ϵ̄ + ϵ[uµ ], k) : ϵ[δu]⟩ = 0 ∀δu ∈ UD,0

⟨σ (ϵ[u], k) : ϵ[δu]⟩ − σ̂ : ⟨⟨[[δu]] ⊗ n⟩⟩ = 0 ,

−⟨g(ξ )˙⋆δξ ⟩ − ⟨k ⋆ χ [δξ ]⟩ = 0 ,

5. Homogenization of polycrystal response

10 Similar arguments apply to the other SVE-problems.

E [MPa] ν [–] l [mm] H gra [MPa] H loc [MPa] σY [MPa] t∗ [s]

where the “plastic multiplier” state function is given as

6.1. 1D bar model

6.2. 2D grain model

7. Conclusions and outlook

Appendix A. Extremal properties of time-incremental problem in the semi-dual format

The task is to verify the extremal properties

min max min max Π (û, ξ̂ , k̂, κ̂ di ) (A.1)

max Π (u, ξ , k, κ̂ di ). (A.5)

• Stationarity (1st variation)

Step 2: For fixed u, ξ , consider