Introduction
Predicting damage initiation and its progression in structural materials relies heavily on the knowledge of local mechanical stresses present in the material. For particular aging processes, the material damage is initiated at the grain boundaries (GBs), where intergranular microcracks form. With time, these microcracks may grow along the GBs and combine into larger macroscopic cracks, which can eventually compromise the structural integrity of the entire component under load. Since microcracks are invisible to non-destructive inspection techniques, the detection instruments can only reveal the existence of macroscopic cracks, which roughly appear in the final 10% of the component’s lifetime. Having accurate models for predicting the component’s susceptibility to microcracking in its earlier stages is therefore of uttermost importance in many different applications, as this could reduce the costs needed for frequent inspections and replacements. InterGranular Stress-Corrosion Cracking (IGSCC) is one of the most significant ageing-degradation mechanisms. It corresponds to the initiation and propagation of microcracks along the GBs and is common in alloys, that are otherwise typically corrosion-resistant (austenitic stainless steels [1 , 2 , 3 , 4 , 5 ] , zirconium alloys [6 , 7 ] , nickel based alloys [8 , 9 , 10 , 11 ] , high strength aluminum alloys [12 , 13 ] and ferritic steels [14 , 15 ] ). The IGSCC is a multi-level process that includes electro-chemical, micro-mechanical and thermo-mechanical mechanisms. The activation of these mechanisms depends on material properties, corrosive environment and local stress state. It is believed that GB stresses are the driving force of intergranular cracking, therefore they need to be accurately determined, in order to make quantitative predictions about IGSCC initiation. Various approaches to IGSCC-initiation modeling are being considered. One such approach is to treat IGSCC phenomenon on a local GB scale, where GB-normal stresses σ n n \sigma_{nn} σ nn can be studied separately (decoupled) from the environmental effects that degrade the GB strength σ c \sigma_c σ c ; IGSCC ≈ F ( σ n n ) ⋅ F ( σ c ) \approx\mathcal{F}(\sigma_{nn}) \cdot\mathcal{F}(\sigma_c) ≈ F ( σ nn ) ⋅ F ( σ c ) . A GB-normal stress σ n n \sigma_{nn} σ nn is defined here as a component of local stress tensor along the GB-normal direction, i.e., perpendicular to the GB plane.1 Hence, a single stress-based criterion for a local IGSCC initiation can be assumed on every GB: IGSCC gets initiated wherever σ n n > σ c \sigma_{nn} > \sigma_c σ nn > σ c , with both these quantities being local in a sense, that they in principle depend on the position of the GB within the aggregate. The introduced local criterion can be used to evaluate the probability, that a randomly selected GB on a component’s surface, where it is in contact with the corrosive environment, is overloaded (or soon-to-be cracked). This probability can be estimated by calculating a fraction η \eta η of GBs with σ n n > σ c \sigma_{nn} > \sigma_c σ nn > σ c as η = ∫ σ c ∞ PDF ( σ n n ) d σ n n \eta= \int_{\sigma_c}^{\infty} \operatorname{PDF}(\sigma_{nn})\mathrm{d}\sigma_{nn} η = ∫ σ c ∞ PDF ( σ nn ) d σ nn , for the assumed probability-density function PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) . If a fraction of overloaded GBs exceeds a threshold value, η > η f \eta> \eta_f η > η f , a specimen-sized crack may develop, possibly resulting in a catastrophic failure of the component. This approach, based on the accurate knowledge of GB-normal stresses, seems feasible when a GB strength σ c \sigma_c σ c is known and approximately constant within the examined surface section. Unfortunately, this is not the case in real materials.
Measurements have shown that different GBs show different IGSCC sensitivities [5 , 16 , 17 , 18 ] , implying that GB strength σ c \sigma_c σ c depends not only on the material and environmental properties but also strongly on a GB type; σ c = F \sigma_c = \mathcal{F} σ c = F (GB type, material, environment). Here, a GB type denotes a GB microstructure (inter-atomic arrangements in the vicinity of the GB), which affects the GB energy and, eventually, its strength σ c \sigma_c σ c . In the continuum limit, five parameters are needed to define a GB neighborhood: four parameters are required to specify a GB plane in crystallographic systems of the two adjacent grains and one parameter defines a twist rotation between the associated crystal lattices about the plane normal. In principle, the term “GB type” thus refers to GBs with the same GB strength. Sometimes it is convenient to specify a GB type by less than five parameters (e.g., when the values of skipped parameters do not affect the σ n n \sigma_{nn} σ nn distribution). For instance, in the [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GB type, with a GB-plane normal along the [ a b c ] [abc] [ ab c ] direction in one grain and [ d e f ] [def] [ d e f ] direction in the other grain, the twist angle can be assumed random (and thus remains unspecified).
In addition to σ c \sigma_c σ c being a function of GB type, also the distributions of GB-normal stresses should be evaluated for different GB types in order to later perform a meaningful calculation of fraction η \eta η . Hence, PDF ( σ n n ) = F \operatorname{PDF}(\sigma_{nn}) = \mathcal{F} PDF ( σ nn ) = F (GB type, applied stress, material). Since exact general solutions for both the local σ n n \sigma_{nn} σ nn and statistical PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) are too complex to be derived analytically, researchers have restricted themselves to numerical simulations limited to few selected materials and specific (usually uniaxial) loading conditions.
Crystal-plasticity finite element (FE) simulations [19 , 20 , 21 , 22 , 23 , 24 ] and crystal-plasticity fast Fourier transform simulations [25 , 26 ] have been used to obtain intergranular stresses on random GBs in either synthetic or realistic polycrystalline aggregates, providing valuable information for IGSCC initiation in those specific cases. In particular, the fluctuations of intergranular normal stresses (the widths of PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) ) have been found to depend primarily on the elasto-plastic anisotropy of the grains with either cubic [22 , 23 , 24 ] or hexagonal lattice symmetries [24 ] .
Although the computationally demanding simulations can provide accurate results, such an approach deems impractical for a general case and provides little insights into involved physics. Thus, efforts have been made to identify most influential parameters affecting the GB-normal stresses on any single GB type [27 , 28 ] . In the elastic regime of grains with cubic lattice symmetry, Zener elastic anisotropy index A A A [29 ] and effective GB stiffness E 12 E_{12} E 12 , measuring the average stiffness of GB neighborhood along the GB-normal direction, have been identified and demonstrated to be sufficient for quantifying normal-stress fluctuations on any GB type in a given material under uniaxial external loading [28 ] . The empirical relation (still lacking a satisfactory explanation) has been established for the standard deviation of σ n n \sigma_{nn} σ nn distribution evaluated on [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GBs, which is a function of A A A and E 12 E_{12} E 12 . On the contrary, the mean value of the same σ n n \sigma_{nn} σ nn distribution has been shown to be independent of the chosen material and/or the GB type on which it is calculated.
To account for elastic–perfectly plastic grains at applied tensile yield stress, a simple Schmid-Modified Grain-Boundary-Stress model has been proposed [27 ] to investigate the initiation of an intergranular crack, based on a normal stress acting at GB. The model considers combined effects of GB-plane orientation and grain orientations through their Schmid factors. It has been pointed out, that intergranular cracks occur most likely at highly stressed GBs. In other similar studies [5 , 30 , 31 ] , the same model has been used to discuss crack initiation in austenitic stainless steel, concluding that initiation sites coincide with the most highly stressed GBs.
Building upon partial results, limited to either specific loading conditions [27 , 28 ] and/or specific grain-lattice symmetries [28 ] , the goal of this study is to develop a model of GB-normal stresses, that would provide accurate analytic or semi-analytic expressions for σ n n \sigma_{nn} σ nn , with the corresponding statistical measure PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) , depending on a general GB type, general applied stress and general elastic polycrystalline material. Once the knowledge of GB strength σ c \sigma_c σ c becomes available, the resulting expressions will be directly useful in the mechanistic modeling of GB-damage initiation (such as IGSCC2 {}^{2} 2 ) and should therefore become a quick and reliable tool to all the experts dealing with local damage modeling and characterization.
The paper is structured as follows: in Sec. II typical material and GB-type effects on GB-normal stresses are introduced. In Sec. III the perturbative framework for predicting GB-normal stresses is developed, providing analytical and semi-analytical models along with their solutions. In Sec. IV the upgraded models are verified with FE-simulation results. Practical implications are discussed in Sec. V and in Sec. VI some concluding remarks are given. All technical details are deferred to the set of Appendices.
Material and grain boundary type effects on intergranular normal stresses
The anisotropic elasticity of crystals is governed by the generalized Hooke’s law, σ i j = C i j k l ε k l \sigma_{ij} = C_{ijkl}\varepsilon_{kl} σ ij = C ij k l ε k l , where C i j k l C_{ijkl} C ij k l is a 3D fourth-order stiffness tensor. Depending on the symmetry of the underlying grain lattice, C i j k l C_{ijkl} C ij k l can be expressed in terms of two (isotropic), three (cubic), or more (up to 21 for triclinic) independent elastic parameters. All grains in a polycrystalline aggregate are assigned the same elastic material properties, but different, random crystallographic orientations (no texture). However, for practical purposes, we artificially increase the share of GBs of a certain type in our finite element aggregate models, by imposing specific orientations to a relatively small fraction of grains.
In the continuum limit3 , a general GB type is defined by five independent parameters, which specify the orientations of two nearest grains relative to the GB plane. It can be expressed in the form of [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GBs, where their GB plane is the [ a b c ] [abc] [ ab c ] plane in one grain and [ d e f ] [def] [ d e f ] plane in the other grain, with Δ ω \Delta\omega Δ ω denoting a relative twist of the two grain orientations about the GB normal.4
Due to topological constraints, not all GBs can be assigned the same GB character. In practice, a particular
[ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GB type can be ascribed to at most ∼ 17 % \sim17\% ∼ 17% of the GBs in a given aggregate, with the remaining GBs being of random type (i.e., defined by two randomly oriented neighboring grains). A polycrystalline aggregate and two particular GB types are visualized in Figure 1 .
Figure 1. (a) 3D periodic Voronoi aggregate with 4000 grains used in this study. Different grains are denoted by different colors. Finite element mesh is shown for one selected grain. Visualization of two different GBs with fixed GB plane (with normal n n n ) but different crystallographic orientations: (b) [1 ] -[1 ] -30° GB and (c) [111]-[111]-30° GB.
The constitutive equations of the generalized Hooke’s law are solved numerically for a chosen uniform loading with FE solver Abaqus [32 ] . The obtained stresses σ \sigma σ , corresponding to the nearest integration points of a particular GB k k k , are then used to produce a single value σ n n ( k ) \sigma_{nn}(k) σ nn ( k ) as their weighted average. Besides local stresses σ n n ( k ) \sigma_{nn}(k) σ nn ( k ) , first two statistical moments of PDF(σ n n \sigma_{nn} σ nn ), the mean value and standard deviation, are calculated for the distribution of stresses σ n n \sigma_{nn} σ nn on GBs of a chosen, over-represented [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GB type, whose density was artificially boosted (see Appendix A for further details).
Figs. 2 and 3 show typical (strong) effects of different GB types and different materials on both, local stresses σ n n \sigma_{nn} σ nn and the corresponding stress distributions PDF(σ n n \sigma_{nn} σ nn ), for the assumed macroscopic uniaxial tensile loading Σ \Sigma Σ . In Figure 2 , a comparison of different [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GB types is made, with Δ ω \Delta\omega Δ ω assumed random. Each value of GB index refers to a particular GB within the aggregate of fixed grain topology, shown in Figure 1 (a). In this way, the effect of the GB type can be isolated from other contributions. While the mean stress is independent of the GB type (with ⟨ σ n n ⟩ = Σ / 3 \langle\sigma_{nn}\rangle= \Sigma/3 ⟨ σ nn ⟩ = Σ/3 for all types), the stress fluctuations are much larger on the (stiffest) [111]-[111] GBs than on the (softest) [1 ] -[1 ] GBs [28 ] .
Figure 2. (a) Normalized local stress responses σ n n / Σ \sigma_{nn}/\Sigma σ nn /Σ and (b) their statistical distributions PDF ( σ n n / Σ ) \operatorname{PDF}(\sigma_{nn}/\Sigma) PDF ( σ nn /Σ ) in a polycrystalline lithium under macroscopic tensile loading Σ \Sigma Σ for 3 different GB types. Large influence of a chosen GB type and poor prediction capability of the isotropic model (σ n n / Σ = cos 2 θ \sigma_{nn}/\Sigma= \cos^{2}\theta σ nn /Σ = cos 2 θ ) are clearly visible. In panel (a) the results are shown for just 15 randomly selected GBs of each type (GB index).
A similar behavior is observed in Figure 3 , where the effect of different material properties is isolated from other contributions by comparing σ n n \sigma_{nn} σ nn on identical GBs. All stress distributions PDF(σ n n \sigma_{nn} σ nn ) are again centered around ⟨ σ n n ⟩ = Σ / 3 \langle\sigma_{nn}\rangle= \Sigma/3 ⟨ σ nn ⟩ = Σ/3 , while they are at the same time getting considerably wider with increasing grain anisotropy [28 ] .
Figure 3. Similarly as in Figure 2 , but evaluated on [111]-[111] GB type in different materials to demonstrate the effect of their elastic properties. Panel (b) shows how isotropic model begins to fail with the growing anisotropy of the grains.
In most cases depicted in Figure 2 and Figure 3 , a poor prediction capability of the isotropic model5 is observed, implying that local GB stresses are non-trivially dependent on the GB type, material properties and loading conditions.
Perturbative model of grain boundary normal stresses
Assumptions
To develop an accurate prediction for σ n n \sigma_{nn} σ nn (and the corresponding PDF(σ n n \sigma_{nn} σ nn )), a step-by-step approach is taken, inspired by perturbation theory. In this sense, the solution for σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) , starting from the trivial isotropic-grain solution σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) , is refined in each successive step k k k by considering the contribution of more distant grains. To provide an analytic solution, sensible approximations and assumptions are used. For example, following Saint Venant’s principle, the effects of more distant neighborhood on a GB are described in less detail, using only average quantities such as elastic grain anisotropy A u A^{u} A u [33 ] or isotropic bulk stiffness ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ . The strategy for building a perturbative model is shown schematically in Figure 4 . In the simplest approximation (k = 0 k = 0 k = 0 ), the neighborhood of a chosen GB can be modeled as isotropic, in which case the only relevant degree of freedom is the orientation of the GB plane. In the next order iteration (k = 1 k = 1 k = 1 ), the two (anisotropic) grains enclosing the GB are considered, while their combined (axial) strain is assumed equal as if both grains were made from (isotropic) bulk material.
Figure 4. Perturbation-theory based strategy for finding GB-normal stress σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) . In each successive step k k k , a more complex GB neighborhood is taken into account. For simplicity, the scheme presented here is only 2D and subjected to tensile loading Σ \boldsymbol{\Sigma} Σ , but in practise a 3D case for a general uniform loading is considered.
This assumption works well for average grains, but the stiffer or softer the grains in the pair are, the more it starts to fail. To relax that condition, in the next order iteration (k = 2 k = 2 k = 2 ), “buffer” grains are introduced along the GB-normal (axial) direction. Then not only bicrystal pair, but the whole axial chain (containing also the buffer grains) is supposed to deform as if it was made from bulk material. In a similar manner, buffer grains are added also along the transverse directions, forming transverse chains of grains whose axial strain is constrained by the bulk (k = 3 k = 3 k = 3 ).
In the isotropic-grain solution (k = 0 k = 0 k = 0 ), the GB-normal stress is equal to the externally applied stress projected onto a GB plane, σ n n ( 0 ) = Σ z z \sigma_{nn}^{(0)} = \Sigma_{zz} σ nn ( 0 ) = Σ z z , which for uniaxial loading Σ \Sigma Σ translates to σ n n ( 0 ) = Σ cos 2 θ \sigma_{nn}^{(0)} = \Sigma\cos^{2}\theta σ nn ( 0 ) = Σ cos 2 θ , where θ \theta θ is the angle between the GB normal and loading direction. The isotropic-grain solution may be a good initial approximation, but it turns to be a poor solution for moderate and highly anisotropic materials, see Figure 2 and Figure 3 .
To obtain higher-order (k > 0 k > 0 k > 0 ) solutions σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) , the effect of two nearest grains enclosing the GB is considered in more detail, while the effect of more distant, buffer grains is accounted for less rigorously. Instead of a full 3D solution, several partial 1D solutions are obtained simultaneously and properly combined to accurately approximate σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) . Schematically, the corresponding general model can be viewed in Figure 5 , as composed of one axial chain of length L n + 2 L_n + 2 L n + 2 and four lateral chains of length L t + 1 L_t + 1 L t + 1 crossing the two grains, that are adjacent to GB.
Figure 5. A 2D sketch of perturbative model for GB stresses, consisting of two anisotropic grains of unit size, enclosing the GB, and several isotropic buffer grains of variable length, composing one axial chain of length L n + 2 L_n+2 L n + 2 and two (four in 3D) transverse chains of length L t + 1 L_t+1 L t + 1 . Stresses and strains are assumed constant within the grains. Total strain of each chain is prescribed to match that of isotropic bulk of the same length and under the same external loading (Voigt-like assumption on a chain-length scale). 3D coupling of the chains with the surrounding bulk is modeled by assuming variable chain stiffness. External loading Σ \boldsymbol{\Sigma} Σ is dressed by fluctuations f \mathbf{f} f .
The chains are assumed decoupled from each other, but they interact with the surrounding bulk. The bulk is taken as isotropic, with average (bulk) properties, such as elastic stiffness ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ and Poisson’s ratio ⟨ ν ⟩ \langle\nu\rangle ⟨ ν ⟩ . The chain-bulk interaction is, in the first approximation (i.e., without lateral 3D coupling), assumed to be along the chain direction. It constrains the total strain of the chain to that of the isotropic bulk. This boundary condition corresponds to the Voigt-like assumption, but on a chain-length scale.
Buffer grains are assumed isotropic as well, but with elastic stiffness E b E_b E b and Poisson’s ratio ν b \nu_b ν b , both corresponding to the average response of a chain with L n L_n L n (or L t L_t L t ) randomly oriented grains. However, when accounting for the lateral 3D effects (cf. Sec. IIID 1), the chains are allowed to interact also laterally with the bulk and in the limit of long chains both parameters approach those of the bulk, E b ∼ ⟨ E ⟩ E_b \sim\langle E\rangle E b ∼ ⟨ E ⟩ and ν b ∼ ⟨ ν ⟩ \nu_b \sim\langle\nu\rangle ν b ∼ ⟨ ν ⟩ .
The two grains on either side of the GB are assumed anisotropic, with their crystallographic orientations determining the [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] -Δ ω \Delta\omega Δ ω type of the corresponding GB.
Finally, the stresses and strains are considered homogeneous within all the grains. In addition, a general analytical expression for σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) is derived by applying a reduced set of boundary conditions. To facilitate a simple, closed-form solution for σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) , only the conditions for stresses are imposed at the GB, while those for strains are neglected. Hence, the stress equilibrium is fulfilled everywhere in the model, while the strain compatibility at the GB is not guaranteed. These assumptions will be justified a posteriori by comparing the model results with those from numerical simulations.
Analytical models
General setup
Analytical expressions for σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) are presented in the GB coordinate system ( x , y , z ) (x,y,z) ( x , y , z ) with z z z -axis along the GB normal. All quantities expressed in the local-grain coordinate system ( n 1 , n 2 , n 3 ) (n_1,n_2,n_3) ( n 1 , n 2 , n 3 ) , aligned with crystallographic (eigen-)axes, and the laboratory coordinate system ( X , Y , Z ) (X,Y,Z) ( X , Y , Z ) , therefore need to be appropriately transformed using the following (passive) rotations R c r y \mathbf{R}^{\mathrm{cry}} R cry and R l a b \mathbf{R}^{\mathrm{lab}} R lab , respectively (see also Figure 6 ),
Figure 6. Definition of three coordinate systems: laboratory coordinate system ( X , Y , Z ) (X,Y,Z) ( X , Y , Z ) , local-grain coordinate system ( n 1 , n 2 , n 3 ) (n_1,n_2,n_3) ( n 1 , n 2 , n 3 ) , and GB coordinate system ( x , y , z ) (x,y,z) ( x , y , z ) . The latter can be arbitrarily chosen with respect to the twist angle τ \tau τ about the GB normal n ^ ∥ z ^ \hat{\mathbf{n}}\Vert\hat{\mathbf{z}} n ^ ∥ z ^ . Passive rotations R l a b \mathbf{R}^{\mathrm{lab}} R lab and R c r y \mathbf{R}^{\mathrm{cry}} R cry transform external and local-grain quantities, respectively, to the GB coordinate system. Since in the following, two grains will be considered, four coordinate systems will be in use, namely one crystallographic system for each grain, together with associated rotations R c r y , a b c \mathbf{R}^{\mathrm{cry},abc} R cry , ab c and R c r y , d e f \mathbf{R}^{\mathrm{cry},def} R cry , d e f .
R c r y = ( h l cos ω h 2 + k 2 h 2 + k 2 + l 2 − k sin ω h 2 + k 2 k l cos ω h 2 + k 2 h 2 + k 2 + l 2 + h sin ω h 2 + k 2 − h 2 + k 2 cos ω h 2 + k 2 + l 2 − h l sin ω h 2 + k 2 h 2 + k 2 + l 2 − k cos ω h 2 + k 2 − k l sin ω h 2 + k 2 h 2 + k 2 + l 2 + h cos ω h 2 + k 2 h 2 + k 2 sin ω h 2 + k 2 + l 2 h h 2 + k 2 + l 2 k h 2 + k 2 + l 2 l h 2 + k 2 + l 2 ) (1) \mathbf{R}^{\mathrm{cry}}=
\begin{pmatrix}
\frac{hl\cos\omega}{\sqrt{h^2+k^2}\sqrt{h^2+k^2+l^2}}-\frac{k\sin\omega}{\sqrt{h^2+k^2}} &
\frac{kl\cos\omega}{\sqrt{h^2+k^2}\sqrt{h^2+k^2+l^2}}+\frac{h\sin\omega}{\sqrt{h^2+k^2}} &
-\frac{\sqrt{h^2+k^2}\cos\omega}{\sqrt{h^2+k^2+l^2}}\\
-\frac{hl\sin\omega}{\sqrt{h^2+k^2}\sqrt{h^2+k^2+l^2}}-\frac{k\cos\omega}{\sqrt{h^2+k^2}} &
-\frac{kl\sin\omega}{\sqrt{h^2+k^2}\sqrt{h^2+k^2+l^2}}+\frac{h\cos\omega}{\sqrt{h^2+k^2}} &
\frac{\sqrt{h^2+k^2}\sin\omega}{\sqrt{h^2+k^2+l^2}}\\
\frac{h}{\sqrt{h^2+k^2+l^2}} &
\frac{k}{\sqrt{h^2+k^2+l^2}} &
\frac{l}{\sqrt{h^2+k^2+l^2}}
\end{pmatrix} \tag*{(1)} R cry = h 2 + k 2 h 2 + k 2 + l 2 h l c o s ω − h 2 + k 2 k s i n ω − h 2 + k 2 h 2 + k 2 + l 2 h l s i n ω − h 2 + k 2 k c o s ω h 2 + k 2 + l 2 h h 2 + k 2 h 2 + k 2 + l 2 k l c o s ω + h 2 + k 2 h s i n ω − h 2 + k 2 h 2 + k 2 + l 2 k l s i n ω + h 2 + k 2 h c o s ω h 2 + k 2 + l 2 k − h 2 + k 2 + l 2 h 2 + k 2 c o s ω h 2 + k 2 + l 2 h 2 + k 2 s i n ω h 2 + k 2 + l 2 l (1)
R l a b = ( cos θ cos ψ cos ϕ − sin ψ sin ϕ cos θ sin ψ cos ϕ + cos ψ sin ϕ − sin θ cos ϕ − cos θ cos ψ sin ϕ − sin ψ cos ϕ − cos θ sin ψ sin ϕ + cos ψ cos ϕ sin θ sin ϕ sin θ cos ψ sin θ sin ψ cos θ ) (2) \mathbf{R}^{\mathrm{lab}}=
\begin{pmatrix}
\cos\theta\cos\psi\cos\phi-\sin\psi\sin\phi&
\cos\theta\sin\psi\cos\phi+\cos\psi\sin\phi&
-\sin\theta\cos\phi\\
-\cos\theta\cos\psi\sin\phi-\sin\psi\cos\phi&
-\cos\theta\sin\psi\sin\phi+\cos\psi\cos\phi&
\sin\theta\sin\phi\\
\sin\theta\cos\psi&
\sin\theta\sin\psi&
\cos\theta
\end{pmatrix} \tag*{(2)} R lab = cos θ cos ψ cos ϕ − sin ψ sin ϕ − cos θ cos ψ sin ϕ − sin ψ cos ϕ sin θ cos ψ cos θ sin ψ cos ϕ + cos ψ sin ϕ − cos θ sin ψ sin ϕ + cos ψ cos ϕ sin θ sin ψ − sin θ cos ϕ sin θ sin ϕ cos θ (2)
While standard notation with three Euler angles ( ψ , θ , ϕ ) (\psi,\theta,\phi) ( ψ , θ , ϕ ) , corresponding to a sequence of rotations R 1 \mathbf{R}_1 R 1 about n ^ 3 \hat{n}_3 n ^ 3 (angle ψ \psi ψ ), R 2 \mathbf{R}_2 R 2 about R 1 n ^ 2 \mathbf{R}_1\hat{n}_2 R 1 n ^ 2 (angle θ \theta θ ) and R 3 \mathbf{R}_3 R 3 about R 2 R 1 n ^ 3 = z ^ \mathbf{R}_2\mathbf{R}_1\hat{n}_3=\hat{z} R 2 R 1 n ^ 3 = z ^ (angle ϕ \phi ϕ ), is used for matrix R l a b \mathbf{R}^{\mathrm{lab}} R lab , the rotation R c r y \mathbf{R}^{\mathrm{cry}} R cry is expressed in terms of ( h , k , l , ω ) (h,k,l,\omega) ( h , k , l , ω ) , where the GB normal corresponds to the [ h k l ] [hkl] [ hk l ] direction6 in the local-grain coordinate system, and ω \omega ω denotes a twist angle about the GB normal. This notation is particularly useful for analyzing the response of [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GBs. In the following, we shall always use ( x , y , z ) (x,y,z) ( x , y , z ) to refer to the axes of the GB coordinate system and ( X , Y , Z ) (X,Y,Z) ( X , Y , Z ) for laboratory system associated with the external loading Σ \boldsymbol{\Sigma} Σ . In this respect,
Σ l a b = ( Σ X X Σ X Y Σ X Z Σ X Y Σ Y Y Σ Y Z Σ X Z Σ Y Z Σ Z Z ) (3) \boldsymbol{\Sigma}^{\mathrm{lab}}=
\begin{pmatrix}
\Sigma_{XX} & \Sigma_{XY} & \Sigma_{XZ}\\
\Sigma_{XY} & \Sigma_{YY} & \Sigma_{YZ}\\
\Sigma_{XZ} & \Sigma_{YZ} & \Sigma_{ZZ}
\end{pmatrix} \tag*{(3)} Σ lab = Σ X X Σ X Y Σ X Z Σ X Y Σ Y Y Σ Y Z Σ X Z Σ Y Z Σ Z Z (3)
and
Σ G B = R l a b Σ l a b ( R l a b ) T = ( Σ x x Σ x y Σ x z Σ x y Σ y y Σ y z Σ x z Σ y z Σ z z ) (4) \boldsymbol{\Sigma}^{\mathrm{GB}}=\mathbf{R}^{\mathrm{lab}}\boldsymbol{\Sigma}^{\mathrm{lab}}(\mathbf{R}^{\mathrm{lab}})^{T}=
\begin{pmatrix}
\Sigma_{xx} & \Sigma_{xy} & \Sigma_{xz}\\
\Sigma_{xy} & \Sigma_{yy} & \Sigma_{yz}\\
\Sigma_{xz} & \Sigma_{yz} & \Sigma_{zz}
\end{pmatrix} \tag*{(4)} Σ GB = R lab Σ lab ( R lab ) T = Σ xx Σ x y Σ x z Σ x y Σ y y Σ y z Σ x z Σ y z Σ z z (4)
To find the solution of perturbative models in Figure 4 , the number of variables needs to match the number of boundary conditions. In isotropic limit, σ i j ( 0 ) = Σ i j \sigma_{ij}^{(0)}=\Sigma_{ij} σ ij ( 0 ) = Σ ij , and thus there are no constraints and no degrees of freedom. In a bicrystal model with (1D) axial constraint, there is only a single unknown ( σ z z ( 1 ) = σ z z ( 2 ) ) (\sigma_{zz}^{(1)}=\sigma_{zz}^{(2)}) ( σ z z ( 1 ) = σ z z ( 2 ) ) , and also a single constraint on the axial strain ( ϵ z z ( 1 ) + ϵ z z ( 2 ) = 2 ϵ z z b u l k ) (\epsilon_{zz}^{(1)}+\epsilon_{zz}^{(2)}=2\epsilon_{zz}^{\mathrm{bulk}}) ( ϵ z z ( 1 ) + ϵ z z ( 2 ) = 2 ϵ z z bulk ) . The situation does not change even when buffer grains are added. For a (3D) model in Figure 5 , the following set of conditions is used, constraining the axial strains of all five 1D chains:
L n ϵ z z ( 1 z = 2 z ) + ϵ z z ( 1 ) + ϵ z z ( 2 ) = ( L n + 2 ) ϵ z z b u l k , L t ϵ x x ( 1 x ) + ϵ x x ( 1 ) = ( L t + 1 ) ϵ x x b u l k , L t ϵ x x ( 2 x ) + ϵ x x ( 2 ) = ( L t + 1 ) ϵ x x b u l k , L t ϵ y y ( 1 y ) + ϵ y y ( 1 ) = ( L t + 1 ) ϵ y y b u l k , L t ϵ y y ( 2 y ) + ϵ y y ( 2 ) = ( L t + 1 ) ϵ y y b u l k . (5) \begin{aligned}
L_n\epsilon_{zz}^{(1z=2z)}+\epsilon_{zz}^{(1)}+\epsilon_{zz}^{(2)}&=(L_n+2)\epsilon_{zz}^{\mathrm{bulk}},\\
L_t\epsilon_{xx}^{(1x)}+\epsilon_{xx}^{(1)}&=(L_t+1)\epsilon_{xx}^{\mathrm{bulk}},\\
L_t\epsilon_{xx}^{(2x)}+\epsilon_{xx}^{(2)}&=(L_t+1)\epsilon_{xx}^{\mathrm{bulk}},\\
L_t\epsilon_{yy}^{(1y)}+\epsilon_{yy}^{(1)}&=(L_t+1)\epsilon_{yy}^{\mathrm{bulk}},\\
L_t\epsilon_{yy}^{(2y)}+\epsilon_{yy}^{(2)}&=(L_t+1)\epsilon_{yy}^{\mathrm{bulk}}. \tag*{(5)}
\end{aligned} L n ϵ z z ( 1 z = 2 z ) + ϵ z z ( 1 ) + ϵ z z ( 2 ) L t ϵ xx ( 1 x ) + ϵ xx ( 1 ) L t ϵ xx ( 2 x ) + ϵ xx ( 2 ) L t ϵ y y ( 1 y ) + ϵ y y ( 1 ) L t ϵ y y ( 2 y ) + ϵ y y ( 2 ) = ( L n + 2 ) ϵ z z bulk , = ( L t + 1 ) ϵ xx bulk , = ( L t + 1 ) ϵ xx bulk , = ( L t + 1 ) ϵ y y bulk , = ( L t + 1 ) ϵ y y bulk . (5)
Strain of each grain is weighted by its length, i.e. , either L n , L t ≥ 0 L_n,L_t\geq0 L n , L t ≥ 0 for buffer grains, or 1 for unit-size GB grains. Superscript label of each strain-tensor component (and similarly for stresses) indicates to which particular grain it corresponds; N = 1 , 2 N=1,2 N = 1 , 2 for GB grains or N i N_i N i for buffer grains in i = x , y , z i=x,y,z i = x , y , z directions.
Applying the generalized Hooke’s law to GB grain N N N , the i j ij ij component of its strain tensor can be written as7
ϵ i j ( N ) = ∑ k , l = 1 3 s i j k l G B , N σ k l ( N ) = ∑ k , l , m , n , o , p = 1 3 R i m c r y , N R i n c r y , N R k o c r y , N R l p c r y , N s m n o p c r y σ k l ( N ) , (6) \begin{aligned}
\epsilon_{ij}^{(N)}=\sum_{k,l=1}^{3}s_{ijkl}^{\mathrm{GB},N}\sigma_{kl}^{(N)} \\
&=\sum_{k,l,m,n,o,p=1}^{3}R_{im}^{\mathrm{cry},N}R_{in}^{\mathrm{cry},N}R_{ko}^{\mathrm{cry},N}R_{lp}^{\mathrm{cry},N}s_{mnop}^{\mathrm{cry}}\,\sigma_{kl}^{(N)}, \tag*{(6)}
\end{aligned} ϵ ij ( N ) = k , l = 1 ∑ 3 s ij k l GB , N σ k l ( N ) = k , l , m , n , o , p = 1 ∑ 3 R im cry , N R in cry , N R k o cry , N R l p cry , N s mn o p cry σ k l ( N ) , (6)
with all stress-tensor components σ k l ( N ) \sigma_{kl}^{(N)} σ k l ( N ) listed in Table I . Note that shear stresses do not appear as variables in either grain, but have their values assigned ( σ i j ( N ) = Σ i j (\sigma_{ij}^{(N)}=\Sigma_{ij} ( σ ij ( N ) = Σ ij for i ≠ j i\ne j i = j ), i.e. , they are set equal to the components of external-stress tensor, rotated to a local GB system; cf. Eq. (4) . Out of the 6 remaining stress components in both grains, two are set equal due to stress-continuity condition ( σ z z ( 1 ) = σ z z ( 2 ) : = σ z z ) (\sigma_{zz}^{(1)}=\sigma_{zz}^{(2)}:=\sigma_{zz}) ( σ z z ( 1 ) = σ z z ( 2 ) := σ z z ) , hence the number of unknowns (five) matches the number of constraints in Eq. (5) .
Table I. Assumed stress components in different grains of the model. Buffer grain label N i N_i N i denotes the corresponding GB grain (N = 1 , 2 N=1,2 N = 1 , 2 ) and the direction of the chain, to which it belongs (i = x , y , z i=x,y,z i = x , y , z ).
Compliance tensor s c r y s^{\mathrm{cry}} s cry , expressed in the local (crystallographic) coordinate system of the grain, is readily transformed to the GB system, where rotation matrices R c r y \mathbf{R}^{\mathrm{cry}} R cry can be different for both grains. Depending on the symmetry of the grain lattice, s c r y s^{\mathrm{cry}} s cry , can be expressed as a function of minimum two (isotropic) and maximum 21 (triclinic) independent elastic parameters. Here, no preference for the underlying symmetry is assumed, thus keeping the approach as general as possible.
To maintain the clarity of the manuscript, only functional dependence of ϵ i i ( N ) \epsilon_{ii}^{(N)} ϵ ii ( N ) is retained here8 (with full analytic expressions for cubic lattice symmetry presented in Appendix B),
ϵ i i ( 1 ) = F ( s c r y ; a , b , c , ω 1 ; σ x x ( 1 ) , σ y y ( 1 ) , σ z z , Σ x y , Σ x z , Σ y z ) , ϵ i i ( 2 ) = F ( s c r y ; d , e , f , ω 2 ; σ x x ( 2 ) , σ y y ( 2 ) , σ z z , Σ x y , Σ x z , Σ y z ) . (7) \begin{aligned}
\epsilon_{ii}^{(1)} &= \mathcal{F}\left(s^{\mathrm{cry}};a,b,c,\omega_1;\sigma_{xx}^{(1)},\sigma_{yy}^{(1)},\sigma_{zz},\Sigma_{xy},\Sigma_{xz},\Sigma_{yz}\right),\\
\epsilon_{ii}^{(2)} &= \mathcal{F}\left(s^{\mathrm{cry}};d,e,f,\omega_2;\sigma_{xx}^{(2)},\sigma_{yy}^{(2)},\sigma_{zz},\Sigma_{xy},\Sigma_{xz},\Sigma_{yz}\right). \tag*{(7)}
\end{aligned} ϵ ii ( 1 ) ϵ ii ( 2 ) = F ( s cry ; a , b , c , ω 1 ; σ xx ( 1 ) , σ y y ( 1 ) , σ z z , Σ x y , Σ x z , Σ y z ) , = F ( s cry ; d , e , f , ω 2 ; σ xx ( 2 ) , σ y y ( 2 ) , σ z z , Σ x y , Σ x z , Σ y z ) . (7)
Generic ( h , k , l , ω ) (h,k,l,\omega) ( h , k , l , ω ) parameters in R c r y \mathbf{R}^{\mathrm{cry}} R cry have been replaced by specific values ( a , b , c , ω 1 ) (a,b,c,\omega_1) ( a , b , c , ω 1 ) and ( d , e , f , ω 2 ) (d,e,f,\omega_2) ( d , e , f , ω 2 ) in GB grains 1 and 2, respectively. This setting corresponds to a well-defined GB type [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω with Δ ω : = ω 2 − ω 1 \Delta\omega:= \omega_2-\omega_1 Δ ω := ω 2 − ω 1 .
Similar expressions apply also to buffer grains N i N_i N i . The only difference is, that there all stress components correspond to projected external loading Σ G B \boldsymbol{\Sigma}^{\mathrm{GB}} Σ GB . The only exception is the axial stress σ i i ( N i ) \sigma_{ii}^{(N_i)} σ ii ( N i ) , which matches σ i i ( N ) \sigma_{ii}^{(N)} σ ii ( N ) in GB grain N N N due to stress-continuity along the chain length. Stress components in each of the grains are summarized in Table I .
Sufficiently far from the GB, the grains can be treated as isotropic. This allows for much simpler expressions for strain components ϵ i i \epsilon_{ii} ϵ ii (with i = x , y , z i=x,y,z i = x , y , z ) in both, buffer grains and the bulk material,
ϵ i i ( N i ) = 1 E b ( σ i i ( N ) − ν b ( tr ( Σ G B ) − Σ i i ) ) , (8) \epsilon_{ii}^{(N_i)}=\frac{1}{E_b}\left(\sigma_{ii}^{(N)}-\nu_b\left(\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{GB}})-\Sigma_{ii}\right)\right), \tag*{(8)} ϵ ii ( N i ) = E b 1 ( σ ii ( N ) − ν b ( tr ( Σ GB ) − Σ ii ) ) , (8)
ϵ i i b u l k = 1 ⟨ E ⟩ ( Σ i i − ⟨ ν ⟩ ( tr ( Σ G B ) − Σ i i ) ) . (9) \epsilon_{ii}^{\mathrm{bulk}}=\frac{1}{\langle E\rangle}\left(\Sigma_{ii}-\langle\nu\rangle\left(\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{GB}})-\Sigma_{ii}\right)\right). \tag*{(9)} ϵ ii bulk = ⟨ E ⟩ 1 ( Σ ii − ⟨ ν ⟩ ( tr ( Σ GB ) − Σ ii ) ) . (9)
With relevant strain components in individual grains defined in Eqs. (7) , (8) and (9) , the set of conditions in Eq. (5) , constraining the axial strains of all five chains, can be solved analytically for all five unknown stresses σ i i ( N ) \sigma_{ii}^{(N)} σ ii ( N ) , including the GB-normal stress σ n n : = σ z z \sigma_{nn}:=\sigma_{zz} σ nn := σ z z . However, the resulting σ n n \sigma_{nn} σ nn has a significant deficiency. It depends on the choice of the local GB coordinate system (the value of twist angle τ \tau τ in Figure 6 ). This dependence originates in the prescribed directions of the four lateral chains, which are directed along the local x x x and y y y axes. A different choice of x x x and y y y axes would produce different lateral constraints, which would result in a different σ n n \sigma_{nn} σ nn . To avoid this ambiguity, a symmetrized lateral boundary condition is derived below.
Symmetrized model
The model is symmetrized by averaging the lateral boundary condition over all possible GB coordinate systems. The twist of the local GB system for an arbitrary angle τ \tau τ about the GB normal (z z z -axis) changes how Σ G B \boldsymbol{\Sigma}^{\mathrm{GB}} Σ GB and σ G B , N \sigma^{\mathrm{GB},N} σ GB , N are expressed in that system. Specifically, the rotation changes the Euler angles ω N \omega_N ω N and ϕ \phi ϕ in transformation matrices (1) and (2), respectively,
ω N → ω N + τ ; ( N = 1 , 2 ) , ϕ → ϕ + τ , (10) \begin{aligned}
\omega_N &\to\omega_N+\tau;\qquad(N=1,2),\\
\phi&\to\phi+\tau, \tag*{(10)}
\end{aligned} ω N ϕ → ω N + τ ; ( N = 1 , 2 ) , → ϕ + τ , (10)
which in turn affect Eqs. (7) , (8) and (9) , and make them τ \tau τ dependent. Since all twist rotations should be equivalent, averaging over τ \tau τ replaces Eq. (5) with new, symmetrized boundary conditions
1 2 π ∫ 0 2 π ( L n ϵ z z ( 1 z = 2 z ) + ϵ z z ( 1 ) + ϵ z z ( 2 ) ) d τ = 1 2 π ∫ 0 2 π ( L n + 2 ) ϵ z z b u l k d τ , \frac{1}{2\pi}\int_{0}^{2\pi}\left(L_n\epsilon_{zz}^{(1_z=2_z)}+\epsilon_{zz}^{(1)}+\epsilon_{zz}^{(2)}\right)\,d\tau=\frac{1}{2\pi}\int_{0}^{2\pi}(L_n+2)\epsilon_{zz}^{\mathrm{bulk}}\,d\tau, 2 π 1 ∫ 0 2 π ( L n ϵ z z ( 1 z = 2 z ) + ϵ z z ( 1 ) + ϵ z z ( 2 ) ) d τ = 2 π 1 ∫ 0 2 π ( L n + 2 ) ϵ z z bulk d τ ,
1 2 π ∫ 0 2 π ( L t ϵ x x ( 1 x ) + ϵ x x ( 1 ) ) d τ = 1 2 π ∫ 0 2 π ( L t + 1 ) ϵ x x b u l k d τ , \frac{1}{2\pi}\int_{0}^{2\pi}\left(L_t\epsilon_{xx}^{(1_x)}+\epsilon_{xx}^{(1)}\right)\,d\tau=\frac{1}{2\pi}\int_{0}^{2\pi}(L_t+1)\epsilon_{xx}^{\mathrm{bulk}}\,d\tau, 2 π 1 ∫ 0 2 π ( L t ϵ xx ( 1 x ) + ϵ xx ( 1 ) ) d τ = 2 π 1 ∫ 0 2 π ( L t + 1 ) ϵ xx bulk d τ ,
1 2 π ∫ 0 2 π ( L t ϵ x x ( 2 x ) + ϵ x x ( 2 ) ) d τ = 1 2 π ∫ 0 2 π ( L t + 1 ) ϵ x x b u l k d τ , \frac{1}{2\pi}\int_{0}^{2\pi}\left(L_t\epsilon_{xx}^{(2_x)}+\epsilon_{xx}^{(2)}\right)\,d\tau=\frac{1}{2\pi}\int_{0}^{2\pi}(L_t+1)\epsilon_{xx}^{\mathrm{bulk}}\,d\tau, 2 π 1 ∫ 0 2 π ( L t ϵ xx ( 2 x ) + ϵ xx ( 2 ) ) d τ = 2 π 1 ∫ 0 2 π ( L t + 1 ) ϵ xx bulk d τ ,
1 2 π ∫ 0 2 π ( L t ϵ y y ( 1 y ) + ϵ y y ( 1 ) ) d τ = 1 2 π ∫ 0 2 π ( L t + 1 ) ϵ y y b u l k d τ , \frac{1}{2\pi}\int_{0}^{2\pi}\left(L_t\epsilon_{yy}^{(1_y)}+\epsilon_{yy}^{(1)}\right)\,d\tau=\frac{1}{2\pi}\int_{0}^{2\pi}(L_t+1)\epsilon_{yy}^{\mathrm{bulk}}\,d\tau, 2 π 1 ∫ 0 2 π ( L t ϵ y y ( 1 y ) + ϵ y y ( 1 ) ) d τ = 2 π 1 ∫ 0 2 π ( L t + 1 ) ϵ y y bulk d τ ,
1 2 π ∫ 0 2 π ( L t ϵ y y ( 2 y ) + ϵ y y ( 2 ) ) d τ = 1 2 π ∫ 0 2 π ( L t + 1 ) ϵ y y b u l k d τ . \frac{1}{2\pi}\int_{0}^{2\pi}\left(L_t\epsilon_{yy}^{(2_y)}+\epsilon_{yy}^{(2)}\right)\,d\tau=\frac{1}{2\pi}\int_{0}^{2\pi}(L_t+1)\epsilon_{yy}^{\mathrm{bulk}}\,d\tau. 2 π 1 ∫ 0 2 π ( L t ϵ y y ( 2 y ) + ϵ y y ( 2 ) ) d τ = 2 π 1 ∫ 0 2 π ( L t + 1 ) ϵ y y bulk d τ .
Solving the above set of symmetrized equations for five unknowns σ i i ( N ) \sigma_{ii}^{(N)} σ ii ( N ) (with i = x , y , z i=x,y,z i = x , y , z and N = 1 , 2 N=1,2 N = 1 , 2 ), provides analytical σ n n : = σ z z \sigma_{nn}:=\sigma_{zz} σ nn := σ z z , independent of τ \tau τ . However, for the most general case the resulting expression is too cumbersome to be presented here. Hence, we again resort to its functional dependence
σ n n = F ( s c r y , ⟨ E ⟩ , ⟨ ν ⟩ , E b , ν b ; a , b , c , ω 1 , d , e , f , ω 2 ; L n , L t ; Σ G B ) = F ( s c r y , ⟨ E ⟩ , ⟨ ν ⟩ , E b , ν b ; a , b , c , ω 1 − ϕ , d , e , f , ω 2 − ϕ ; L n , L t ; Σ l a b , θ , ψ ) , (11) \sigma_{nn}=\mathcal{F}(s^{\mathrm{cry}},\langle E\rangle,\langle\nu\rangle,E_b,\nu_b;a,b,c,\omega_1,d,e,f,\omega_2;L_n,L_t;\Sigma^{\mathrm{GB}})
=\mathcal{F}(s^{\mathrm{cry}},\langle E\rangle,\langle\nu\rangle,E_b,\nu_b;a,b,c,\omega_1-\phi,d,e,f,\omega_2-\phi;L_n,L_t;\Sigma^{\mathrm{lab}},\theta,\psi), \tag*{(11)} σ nn = F ( s cry , ⟨ E ⟩ , ⟨ ν ⟩ , E b , ν b ; a , b , c , ω 1 , d , e , f , ω 2 ; L n , L t ; Σ GB ) = F ( s cry , ⟨ E ⟩ , ⟨ ν ⟩ , E b , ν b ; a , b , c , ω 1 − ϕ , d , e , f , ω 2 − ϕ ; L n , L t ; Σ lab , θ , ψ ) , (11)
from which it is clear, that σ n n \sigma_{nn} σ nn does not depend on the choice of the GB coordinate system due to observed ω 1 − ϕ \omega_1-\phi ω 1 − ϕ and ω 2 − ϕ \omega_2-\phi ω 2 − ϕ dependence. The normal stress σ n n \sigma_{nn} σ nn is a complicated function of many parameters9 (e.g. , up to 39 independent parameters in a material with triclinic lattice symmetry and for a most general external loading). However, not all parameters are of same significance, as shown is § 0.3 C, where σ n n \sigma_{nn} σ nn is tested against numerical results. In order to derive a compact, but still meaningful expression, further approximations are needed.
So far, the strategy was based on adding more complexity to the model when getting closer to the GB. In this respect, grains closest to it have been modeled in greater detail (e.g. , employing anisotropic elasticity and mostly unknown loading conditions), while the grains further away required less modeling (e.g. , employing isotropic elasticity and mostly known loading conditions).
With the goal to provide a compact and accurate analytical expression for σ n n \sigma_{nn} σ nn (and the corresponding PDF( σ n n ) (\sigma_{nn}) ( σ nn ) ), few selected limits of the general result, Eq. (11) , are investigated and discussed in more detail. Some of these limits will become very useful later, when a comparison with the numerical results is done in § 0.3 C.
Isotropic limit (k = 0 k=0 k = 0 )
The initial (zeroth order) approximation σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) , representing the exact solution in the isotropic material limit, can be reproduced from Eq. (11) in two ways, either by assuming isotropic properties of the grains (i.e., by taking the appropriate s c r y s^{\mathrm{cry}} s cry ) or taking the limit of very long chains (L n , L t → ∞ L_n,L_t\to\infty L n , L t → ∞ ) with average properties (E b = ⟨ E ⟩ E_b=\langle E\rangle E b = ⟨ E ⟩ , ν b = ⟨ ν ⟩ \nu_b=\langle\nu\rangle ν b = ⟨ ν ⟩ ), in which the chain-strain constraints become ineffective, resulting in stresses equal to external loading,
σ n n ( 0 ) = Σ z z = Σ X X sin 2 θ cos 2 ψ + Σ Y Y sin 2 θ sin 2 ψ + Σ Z Z cos 2 θ + Σ X Y sin 2 θ sin 2 ψ + Σ X Z sin 2 θ cos ψ + Σ Y Z sin 2 θ sin ψ . (12) \sigma_{nn}^{(0)}=\Sigma_{zz}
=\Sigma_{XX}\sin^2\theta\cos^2\psi+\Sigma_{YY}\sin^2\theta\sin^2\psi
+\Sigma_{ZZ}\cos^2\theta+\Sigma_{XY}\sin^2\theta\sin2\psi
+\Sigma_{XZ}\sin2\theta\cos\psi+\Sigma_{YZ}\sin2\theta\sin\psi. \tag*{(12)} σ nn ( 0 ) = Σ z z = Σ X X sin 2 θ cos 2 ψ + Σ Y Y sin 2 θ sin 2 ψ + Σ Z Z cos 2 θ + Σ X Y sin 2 θ sin 2 ψ + Σ X Z sin 2 θ cos ψ + Σ Y Z sin 2 θ sin ψ . (12)
Having a sufficient number of GBs with normals uniformly distributed on a sphere, the corresponding first two statistical moments of PDF( σ n n ( 0 ) ) (\sigma_{nn}^{(0)}) ( σ nn ( 0 ) ) , the mean value and standard deviation, can be straightforwardly expressed as
⟨ σ n n ( 0 ) ⟩ = 1 3 tr ( Σ l a b ) , s ( σ n n ( 0 ) ) = 2 3 5 Σ m i s l a b . (13) \left\langle\sigma_{nn}^{(0)}\right\rangle=\frac{1}{3}\operatorname{tr}(\Sigma^{\mathrm{lab}}),
\qquad
s(\sigma_{nn}^{(0)})=\frac{2}{3\sqrt{5}}\Sigma_{\mathrm{mis}}^{\mathrm{lab}}. \tag*{(13)} ⟨ σ nn ( 0 ) ⟩ = 3 1 tr ( Σ lab ) , s ( σ nn ( 0 ) ) = 3 5 2 Σ mis lab . (13)
where 1 3 tr ( Σ l a b ) \frac{1}{3}\operatorname{tr}(\Sigma^{\mathrm{lab}}) 3 1 tr ( Σ lab ) is a hydrostatic pressure, related to volume change of the aggregate, and Σ m i s l a b \Sigma_{\mathrm{mis}}^{\mathrm{lab}} Σ mis lab corresponds to von Mises external stress, traditionally associated with the yielding of ductile materials. Von Mises stress is related to deviatoric tensor (responsible for volume preserving shape changes of the aggregate),
Σ m i s l a b = 3 2 tr ( ( Σ d e v l a b ) 2 ) , \Sigma_{\mathrm{mis}}^{\mathrm{lab}}=\frac{\sqrt{3}}{\sqrt{2}}\sqrt{\operatorname{tr}\left((\Sigma_{\mathrm{dev}}^{\mathrm{lab}})^2\right)}, Σ mis lab = 2 3 tr ( ( Σ dev lab ) 2 ) ,
Σ d e v l a b : = Σ l a b − 1 3 tr ( Σ l a b ) 1 3 × 3 . (14) \Sigma_{\mathrm{dev}}^{\mathrm{lab}}:=\Sigma^{\mathrm{lab}}-\frac{1}{3}\operatorname{tr}(\Sigma^{\mathrm{lab}})\mathbf{1}_{3\times3}. \tag*{(14)} Σ dev lab := Σ lab − 3 1 tr ( Σ lab ) 1 3 × 3 . (14)
Both, tr ( Σ l a b ) \operatorname{tr}(\Sigma^{\mathrm{lab}}) tr ( Σ lab ) and Σ m i s l a b \Sigma_{\mathrm{mis}}^{\mathrm{lab}} Σ mis lab , are rotational invariants and thus assume identical form in all coordinate systems.
Even though Eq. (13) is derived for isotropic case (k = 0 k=0 k = 0 ), the same functional dependence of the first two statistical moments on Σ l a b \Sigma^{\mathrm{lab}} Σ lab is retained for all orders k k k , suggesting that the loading part can be trivially decoupled from the material and GB-type contributions. In a specific case, when the external stress is of hydrostatic form (i.e. , proportional to identity matrix; Σ l a b : = Σ 0 1 3 × 3 \Sigma^{\mathrm{lab}}:=\Sigma_0\mathbf{1}_{3\times3} Σ lab := Σ 0 1 3 × 3 ), this can be easily confirmed. In that case, there is no effect of grain orientations, since stress tensor is invariant to rotations. Hence, the trivial (hydrostatic) solution applies to the whole aggregate (σ n n ( k ) = Σ 0 \sigma_{nn}^{(k)}=\Sigma_0 σ nn ( k ) = Σ 0 ), resulting in an infinitely narrow stress (and strain) distribution. On the other hand Σ l a b = 0 \Sigma^{\mathrm{lab}}=0 Σ lab = 0 , therefore s ( σ n n ) ∼ Σ m i s l a b s(\sigma_{nn})\sim\Sigma_{\mathrm{mis}}^{\mathrm{lab}} s ( σ nn ) ∼ Σ mis lab applies for any, not just isotropic material.
Axially constrained bicrystal (k = 1 k=1 k = 1 )
The first non-trivial solution σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) corresponds to a bicrystal, embedded axially in the isotropic bulk (L n → 0 L_n\to0 L n → 0 ). As there are no lateral constraints imposed on the two GB grains, this model corresponds to the L t → ∞ L_t\to\infty L t → ∞ limit of the general model shown in Figure 5 . However, to obtain a compact expression for σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) , another simplification is required, which will be justified in § 0.3 C. Since we are interested in the response of [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GBs, which have a well-defined difference of the two twist angles, σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) is obtained by replacing ω 2 \omega_2 ω 2 in Eq. (11) with ω 1 + Δ ω \omega_1+\Delta\omega ω 1 + Δ ω , and averaging it over ω 1 \omega_1 ω 1 :
σ n n ( 1 ) = 1 2 π ∫ 0 2 π ( lim L n → 0 L t → ∞ σ n n ) ω 2 = ω 1 + Δ ω d ω 1 = E 12 Σ z z + E 12 ( ν 12 − ⟨ ν ⟩ ) ( Σ x x + Σ y y ) , (15) \begin{aligned}
\sigma_{nn}^{(1)}=\frac{1}{2\pi}\int_{0}^{2\pi}\left(\lim_{\substack{L_n\to0\\L_t\to\infty}}\sigma_{nn}\right)_{\omega_2=\omega_1+\Delta\omega}\mathrm{d}\omega_1 \\
&=E_{12}\Sigma_{zz}+E_{12}\left(\nu_{12}-\langle\nu\rangle\right)\left(\Sigma_{xx}+\Sigma_{yy}\right), \tag*{(15)}
\end{aligned} σ nn ( 1 ) = 2 π 1 ∫ 0 2 π L n → 0 L t → ∞ lim σ nn ω 2 = ω 1 + Δ ω d ω 1 = E 12 Σ z z + E 12 ( ν 12 − ⟨ ν ⟩ ) ( Σ xx + Σ y y ) , (15)
where
E 12 = 2 ⟨ E ⟩ − 1 E a b c − 1 + E d e f − 1 = 2 ⟨ E ⟩ − 1 s 3333 G B , a b c + s 3333 G B , d e f , (16) E_{12}=\frac{2\langle E\rangle^{-1}}{E_{abc}^{-1}+E_{def}^{-1}}=\frac{2\langle E\rangle^{-1}}{s_{3333}^{\mathrm{GB},abc}+s_{3333}^{\mathrm{GB},def}}, \tag*{(16)} E 12 = E ab c − 1 + E d e f − 1 2 ⟨ E ⟩ − 1 = s 3333 GB , ab c + s 3333 GB , d e f 2 ⟨ E ⟩ − 1 , (16)
ν 12 = ⟨ E ⟩ 4 ( s 3311 G B , a b c + s 3322 G B , a b c + s 3311 G B , d e f + s 3322 G B , d e f ) , \nu_{12}=\frac{\langle E\rangle}{4}\left(s_{3311}^{\mathrm{GB},abc}+s_{3322}^{\mathrm{GB},abc}+s_{3311}^{\mathrm{GB},def}+s_{3322}^{\mathrm{GB},def}\right), ν 12 = 4 ⟨ E ⟩ ( s 3311 GB , ab c + s 3322 GB , ab c + s 3311 GB , d e f + s 3322 GB , d e f ) ,
and
s 33 j j G B , h k l = ∑ m , n , o , p = 1 3 R 3 m c r y , h k l R 3 n c r y , h k l R j o c r y , h k l R j p c r y , h k l s m n o p c r y , (17) s_{33jj}^{\mathrm{GB},hkl}=\sum_{m,n,o,p=1}^{3}R_{3m}^{\mathrm{cry},hkl}R_{3n}^{\mathrm{cry},hkl}R_{jo}^{\mathrm{cry},hkl}R_{jp}^{\mathrm{cry},hkl}s_{mnop}^{\mathrm{cry}}, \tag*{(17)} s 33 j j GB , hk l = m , n , o , p = 1 ∑ 3 R 3 m cry , hk l R 3 n cry , hk l R j o cry , hk l R j p cry , hk l s mn o p cry , (17)
for j = 1 , 2 , 3 j=1,2,3 j = 1 , 2 , 3 and h k l = a b c hkl=abc hk l = ab c or d e f def d e f . This approximation removes (averages out) all the twist-angle degrees of freedom. We will refer to it as the reduced version of the model, intended to mimic the behavior observed in numerical studies. The derived compact expression for σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) is the first main result of this study. It suggests that GB-normal stress is a simple function of the loading part, contained in Σ x x \Sigma_{xx} Σ xx , Σ y y \Sigma_{yy} Σ y y and Σ z z \Sigma_{zz} Σ z z , and the GB-type (and material) part, which is represented compactly by only two (composite) parameters E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 . While E 12 E_{12} E 12 has already been introduced in Ref. [28 ] as an effective GB stiffness, measuring the average stiffness of GB neighborhood along the GB-normal direction, the newly introduced ν 12 \nu_{12} ν 12 can be seen as an effective GB Poisson’s ratio, measuring the average ratio of transverse and axial responses (strains) in both GB grains. Both E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 are unitless and characterize the [ a b c ] [abc] [ ab c ] -[ d e f ] [def] [ d e f ] -Δ ω \Delta\omega Δ ω GB neighborhood in terms of local material and GB-type parameters10
E 12 = F ( s c r y , ⟨ E ⟩ , a , b , c , d , e , f ) , (18) E_{12}=\mathcal{F}(s^{\mathrm{cry}},\langle E\rangle,a,b,c,d,e,f), \tag*{(18)} E 12 = F ( s cry , ⟨ E ⟩ , a , b , c , d , e , f ) , (18)
ν 12 = F ( s c r y , ⟨ E ⟩ , a , b , c , d , e , f ) . (19) \nu_{12}=\mathcal{F}(s^{\mathrm{cry}},\langle E\rangle,a,b,c,d,e,f). \tag*{(19)} ν 12 = F ( s cry , ⟨ E ⟩ , a , b , c , d , e , f ) . (19)
Full analytic expressions for E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 (as well as ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ and ⟨ ν ⟩ \langle\nu\rangle ⟨ ν ⟩ ) depend on the choice of the grain lattice symmetry (expressions for cubic lattice symmetry are given in Appendix B). Note that expressions simplify considerably with more symmetric lattices. In cubic lattices, for example, a GB is fully characterized by E 12 E_{12} E 12 alone, since ν 12 = ⟨ ν ⟩ + 1 2 ( E 12 − 1 − 1 ) \nu_{12}=\langle\nu\rangle+\frac{1}{2}(E_{12}^{-1}-1) ν 12 = ⟨ ν ⟩ + 2 1 ( E 12 − 1 − 1 ) . In isotropic grains, E 12 = 1 E_{12}=1 E 12 = 1 and ν 12 = ⟨ ν ⟩ \nu_{12}=\langle\nu\rangle ν 12 = ⟨ ν ⟩ , which recovers the σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) solution.
Switching to a statistical behavior of infinitely many [ a b c ] [abc] [ ab c ] -[ d e f ] [def] [ d e f ] -Δ ω \Delta\omega Δ ω GBs with randomly oriented GB planes, the first two statistical moments of PDF ( σ n n ( 1 ) ) \operatorname{PDF}(\sigma_{nn}^{(1)}) PDF ( σ nn ( 1 ) ) , the mean value and standard deviation, become
⟨ σ n n ( 1 ) ⟩ = tr ( Σ l a b ) 3 E 12 ( 1 + 2 ( ν 12 − ⟨ ν ⟩ ) ) , \left\langle\sigma_{nn}^{(1)}\right\rangle=\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}})}{3}E_{12}\left(1+2(\nu_{12}-\langle\nu\rangle)\right), ⟨ σ nn ( 1 ) ⟩ = 3 tr ( Σ lab ) E 12 ( 1 + 2 ( ν 12 − ⟨ ν ⟩) ) ,
s ( σ n n ( 1 ) ) = 2 Σ m i s l a b 3 5 E 12 ( 1 − ν 12 + ⟨ ν ⟩ ) 2 . (20) s(\sigma_{nn}^{(1)})=\frac{2\Sigma_{\mathrm{mis}}^{\mathrm{lab}}}{3\sqrt{5}}E_{12}\sqrt{\left(1-\nu_{12}+\langle\nu\rangle\right)^2}. \tag*{(20)} s ( σ nn ( 1 ) ) = 3 5 2 Σ mis lab E 12 ( 1 − ν 12 + ⟨ ν ⟩ ) 2 . (20)
For cubic lattices they simplify to ⟨ σ n n ( 1 ) ⟩ = tr ( Σ l a b ) / 3 \left\langle\sigma_{nn}^{(1)}\right\rangle=\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}})/3 ⟨ σ nn ( 1 ) ⟩ = tr ( Σ lab ) /3 and s ( σ n n ( 1 ) ) = Σ m i s l a b / ( 3 5 ) ∣ 1 − 3 E 12 ∣ s(\sigma_{nn}^{(1)})=\Sigma_{\mathrm{mis}}^{\mathrm{lab}}/(3\sqrt{5})\lvert1-3E_{12}\rvert s ( σ nn ( 1 ) ) = Σ mis lab / ( 3 5 ) ∣ 1 − 3 E 12 ∣ . The fact that the mean stress is equal to Σ / 3 \Sigma/3 Σ/3 for the uniaxial loading Σ \Sigma Σ , while the fluctuation of GB-normal stress in cubic grains is a monotonic function of a single GB parameter E 12 E_{12} E 12 (although the functional dependence differs from that of Eq. (20) ), has already been identified in (realistic) FE simulations [28 ] . However, the observed behavior can now be easily extended to other non-cubic lattices and for general external loading. The accuracy of the derived expressions for local σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) , Eqs. (16)–(17), and statistical PDF ( σ n n ( 1 ) ) \operatorname{PDF}(\sigma_{nn}^{(1)}) PDF ( σ nn ( 1 ) ) , Eq. (20) , is investigated in more detail in § 0.3 C.
Axially constrained chain with L n + 2 L_n+2 L n + 2 grains (k = 2 k=2 k = 2 )
The next-order solution σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) corresponds to a single chain with L n + 2 L_n+2 L n + 2 grains, axially constrained by the isotropic bulk. The reason for adding a buffer grain of length L n > 0 L_n>0 L n > 0 to the bicrystal is to relax the axial strain constraint. In the previous (k = 1 k=1 k = 1 ) iteration, this constraint applies directly to the bicrystal, which produces too large (resp. small) stresses σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) on very stiff (resp. soft) GBs, see § 0.3 C.
Following the same reasoning and steps as in the bicrystal model, the resulting reduced version is derived for a general grain-lattice symmetry and arbitrary external loading
σ n n ( 2 ) : = 1 2 π ∫ 0 2 π ( lim L t → ∞ σ n n ) ∣ ω 2 = ω 1 + Δ ω d ω 1 = 2 + L n 2 E 12 − 1 + L n E 3 − 1 Σ z z + 2 2 E 12 − 1 + L n E 3 − 1 ( ν 12 − ⟨ ν ⟩ − 1 2 L n ( ⟨ ν ⟩ − ν b E 3 − 1 ) ) ( Σ x x + Σ y y ) ≈ 2 + L n 2 E 12 − 1 + L n Σ z z + 2 ( ν 12 − ⟨ ν ⟩ ) 2 E 12 − 1 + L n ( Σ x x + Σ y y ) . (21) \begin{aligned}
\sigma_{nn}^{(2)} := \frac{1}{2\pi}\int_{0}^{2\pi}\left(\lim_{L_t\to\infty}\sigma_{nn}\right)\bigg|_{\omega_2=\omega_1+\Delta\omega}\,\mathrm{d}\omega_1 \\
&= \frac{2+L_n}{2E_{12}^{-1}+L_nE_3^{-1}}\Sigma_{zz}+\frac{2}{2E_{12}^{-1}+L_nE_3^{-1}}\left(\nu_{12}-\langle\nu\rangle-\frac{1}{2}L_n\left(\langle\nu\rangle-\nu_bE_3^{-1}\right)\right)(\Sigma_{xx}+\Sigma_{yy}) \\
&\approx\frac{2+L_n}{2E_{12}^{-1}+L_n}\Sigma_{zz}+\frac{2(\nu_{12}-\langle\nu\rangle)}{2E_{12}^{-1}+L_n}(\Sigma_{xx}+\Sigma_{yy}). \tag*{(21)}
\end{aligned} σ nn ( 2 ) := 2 π 1 ∫ 0 2 π ( L t → ∞ lim σ nn ) ω 2 = ω 1 + Δ ω d ω 1 = 2 E 12 − 1 + L n E 3 − 1 2 + L n Σ z z + 2 E 12 − 1 + L n E 3 − 1 2 ( ν 12 − ⟨ ν ⟩ − 2 1 L n ( ⟨ ν ⟩ − ν b E 3 − 1 ) ) ( Σ xx + Σ y y ) ≈ 2 E 12 − 1 + L n 2 + L n Σ z z + 2 E 12 − 1 + L n 2 ( ν 12 − ⟨ ν ⟩) ( Σ xx + Σ y y ) . (21)
Same definitions for E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 apply as in Eq. (16) , while E 3 : = E b / ⟨ E ⟩ E_3 := E_b/\langle E\rangle E 3 := E b / ⟨ E ⟩ and ν b \nu_b ν b denote, respectively, the normalized elastic stiffness and Poisson’s ratio of the (isotropic) buffer grain. Its response corresponds to the average response of a chain with L n L_n L n randomly oriented grains
E b : = E L n r n d = ⟨ L n ∑ i s 3333 G B , i ⟩ L n , ν b : = ν L n r n d = − ⟨ ∑ i s 1133 G B , i ∑ i s 3333 G B , i ⟩ L n , (22) E_b := E_{L_n}^{\mathrm{rnd}}=\left\langle\frac{L_n}{\sum_i s_{3333}^{\mathrm{GB},i}}\right\rangle_{L_n},
\qquad
\nu_b := \nu_{L_n}^{\mathrm{rnd}}=-\left\langle\frac{\sum_i s_{1133}^{\mathrm{GB},i}}{\sum_i s_{3333}^{\mathrm{GB},i}}\right\rangle_{L_n}, \tag*{(22)} E b := E L n rnd = ⟨ ∑ i s 3333 GB , i L n ⟩ L n , ν b := ν L n rnd = − ⟨ ∑ i s 3333 GB , i ∑ i s 1133 GB , i ⟩ L n , (22)
The averaging ⟨ ⋯ ⟩ L n \langle\cdots\rangle_{L_n} ⟨ ⋯ ⟩ L n is assumed over all possible linear configurations of L n L_n L n grains with random orientations, and the summation index i i i runs over the grains in each chain.
The elastic response of a buffer grain, calculated in this way, is usually softer than that of the bulk (E 3 < 1 E_3<1 E 3 < 1 ). Nevertheless, it is convenient to assume E 3 ≈ 1 E_3\approx1 E 3 ≈ 1 and ν b ≈ ⟨ ν ⟩ \nu_b\approx\langle\nu\rangle ν b ≈ ⟨ ν ⟩ . In fact, this assumption becomes realistic, when the 3D effects are considered, e.g. , the lateral coupling of buffer grain to the neighboring bulk (see § 0.3 D 1).
Assuming E 3 = 1 E_3=1 E 3 = 1 and ν b = ⟨ ν ⟩ \nu_b=\langle\nu\rangle ν b = ⟨ ν ⟩ , the mean value and standard deviation of PDF ( σ n n ( 2 ) ) \operatorname{PDF}(\sigma_{nn}^{(2)}) PDF ( σ nn ( 2 ) ) become
⟨ σ n n ( 2 ) ⟩ = tr ( Σ l a b ) 3 2 + L n + 4 ( ν 12 − ⟨ ν ⟩ ) 2 E 12 − 1 + L n , (23) \left\langle\sigma_{nn}^{(2)}\right\rangle=\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}})}{3}\frac{2+L_n+4(\nu_{12}-\langle\nu\rangle)}{2E_{12}^{-1}+L_n}, \tag*{(23)} ⟨ σ nn ( 2 ) ⟩ = 3 tr ( Σ lab ) 2 E 12 − 1 + L n 2 + L n + 4 ( ν 12 − ⟨ ν ⟩) , (23)
s ( σ n n ( 2 ) ) = 2 Σ m i s l a b 3 5 ( 3 + L n ) − E 12 − 1 2 2 E 12 − 1 + L n , s(\sigma_{nn}^{(2)})=\frac{2\Sigma_{\mathrm{mis}}^{\mathrm{lab}}}{3\sqrt{5}}\frac{\sqrt{\left(3+L_n\right)-E_{12}^{-1\,2}}}{\sqrt{2E_{12}^{-1}+L_n}}, s ( σ nn ( 2 ) ) = 3 5 2 Σ mis lab 2 E 12 − 1 + L n ( 3 + L n ) − E 12 − 1 2 ,
which simplify for cubic lattices to ⟨ σ n n ( 2 ) ⟩ = tr ( Σ l a b ) / 3 \left\langle\sigma_{nn}^{(2)}\right\rangle=\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}})/3 ⟨ σ nn ( 2 ) ⟩ = tr ( Σ lab ) /3 ,
s ( σ n n ( 2 ) ) = 2 Σ m i s l a b 3 5 ( 3 + L n ) − E 12 − 1 2 2 E 12 − 1 + L n . s(\sigma_{nn}^{(2)})=\frac{2\Sigma_{\mathrm{mis}}^{\mathrm{lab}}}{3\sqrt{5}}\sqrt{\frac{(3+L_n)-E_{12}^{-1\,2}}{2E_{12}^{-1}+L_n}}. s ( σ nn ( 2 ) ) = 3 5 2 Σ mis lab 2 E 12 − 1 + L n ( 3 + L n ) − E 12 − 1 2 .
In contrast to the bicrystal model, the σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) expression depends also on the parameter L n L_n L n , which makes it a mixture of bicrystal solution σ n n ( 1 ) \sigma_{nn}^{(1)} σ nn ( 1 ) (reproduced for L n → 0 L_n\to0 L n → 0 ) and isotropic solution σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) (reproduced for L n → ∞ L_n\to\infty L n → ∞ ). However, the effect of L n L_n L n is negligible for GBs with E 12 ∼ 1 E_{12}\sim1 E 12 ∼ 1 and ν 12 ∼ ⟨ ν ⟩ \nu_{12}\sim\langle\nu\rangle ν 12 ∼ ⟨ ν ⟩ . As shown in § 0.3 C, the value L n ∼ 2 L_n\sim2 L n ∼ 2 best replicates the numerical results.
Axially constrained chains with L n + 2 L_n+2 L n + 2 and L t + 1 L_t+1 L t + 1 grains (k = 3 k=3 k = 3 )
The highest-order solution considered in this study is σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) . It corresponds to the complex configuration of chains, shown in Figure 5 . The axial chain consists of L n + 2 L_n+2 L n + 2 grains and the four transverse chains of L t + 1 L_t+1 L t + 1 grains. All the chains are assumed to be axially constrained to the strain of isotropic bulk of equal length. In a similar fashion to previous iterations, the reduced version can be derived for a general grain-lattice symmetry and arbitrary external loading
σ n n ( 3 ) : = 1 2 π ∫ 0 2 π σ n n ∣ ω 2 = ω 1 + Δ ω d ω 1 = A ( 3 ) Σ z z + B ( 3 ) ( Σ x x + Σ y y ) , (24) \begin{aligned}
\sigma_{nn}^{(3)}:=\frac{1}{2\pi}\int_{0}^{2\pi}\sigma_{nn}\bigg|_{\omega_2=\omega_1+\Delta\omega}\,\mathrm{d}\omega_1 \\
&=A^{(3)}\Sigma_{zz}+B^{(3)}(\Sigma_{xx}+\Sigma_{yy}), \tag*{(24)}
\end{aligned} σ nn ( 3 ) := 2 π 1 ∫ 0 2 π σ nn ω 2 = ω 1 + Δ ω d ω 1 = A ( 3 ) Σ z z + B ( 3 ) ( Σ xx + Σ y y ) , (24)
where, assuming E 3 = 1 E_3=1 E 3 = 1 and ν b = ⟨ ν ⟩ \nu_b=\langle\nu\rangle ν b = ⟨ ν ⟩ ,
A ( 3 ) = ( 2 + L n ) ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t t d e f + ⟨ E ⟩ − 1 L t ) + ⟨ ν ⟩ ( ( s t t a b c + ⟨ E ⟩ − 1 L t ) s t l d e f + ( s t t d e f + ⟨ E ⟩ − 1 L t ) s t l a b c ) ( 2 E 12 − 1 + L n ) ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t t d e f + ⟨ E ⟩ − 1 L t ) − 1 2 ⟨ E ⟩ ( ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t l d e f ) 2 + ( s t t d e f + ⟨ E ⟩ − 1 L t ) ( s t l a b c ) 2 ) , A^{(3)}=\frac{(2+L_n)(s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tt}^{def}+\langle E\rangle^{-1}L_t)+\langle\nu\rangle\left((s_{tt}^{abc}+\langle E\rangle^{-1}L_t)s_{tl}^{def}+(s_{tt}^{def}+\langle E\rangle^{-1}L_t)s_{tl}^{abc}\right)}
{(2E_{12}^{-1}+L_n)(s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tt}^{def}+\langle E\rangle^{-1}L_t)-\frac{1}{2}\langle E\rangle\left((s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tl}^{def})^2+(s_{tt}^{def}+\langle E\rangle^{-1}L_t)(s_{tl}^{abc})^2\right)}, A ( 3 ) = ( 2 E 12 − 1 + L n ) ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tt d e f + ⟨ E ⟩ − 1 L t ) − 2 1 ⟨ E ⟩ ( ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tl d e f ) 2 + ( s tt d e f + ⟨ E ⟩ − 1 L t ) ( s tl ab c ) 2 ) ( 2 + L n ) ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tt d e f + ⟨ E ⟩ − 1 L t ) + ⟨ ν ⟩ ( ( s tt ab c + ⟨ E ⟩ − 1 L t ) s tl d e f + ( s tt d e f + ⟨ E ⟩ − 1 L t ) s tl ab c ) ,
B ( 3 ) = − 2 ⟨ ν ⟩ ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t t d e f + ⟨ E ⟩ − 1 L t ) + 1 2 ( 1 + L t − ⟨ ν ⟩ ) ( ( s t t a b c + ⟨ E ⟩ − 1 L t ) s t l d e f + ( s t t d e f + ⟨ E ⟩ − 1 L t ) s t l a b c ) ( 2 E 12 − 1 + L n ) ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t t d e f + ⟨ E ⟩ − 1 L t ) − 1 2 ⟨ E ⟩ ( ( s t t a b c + ⟨ E ⟩ − 1 L t ) ( s t l d e f ) 2 + ( s t t d e f + ⟨ E ⟩ − 1 L t ) ( s t l a b c ) 2 ) . (25) B^{(3)}=-\frac{2\langle\nu\rangle(s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tt}^{def}+\langle E\rangle^{-1}L_t)+\frac{1}{2}(1+L_t-\langle\nu\rangle)\left((s_{tt}^{abc}+\langle E\rangle^{-1}L_t)s_{tl}^{def}+(s_{tt}^{def}+\langle E\rangle^{-1}L_t)s_{tl}^{abc}\right)}
{(2E_{12}^{-1}+L_n)(s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tt}^{def}+\langle E\rangle^{-1}L_t)-\frac{1}{2}\langle E\rangle\left((s_{tt}^{abc}+\langle E\rangle^{-1}L_t)(s_{tl}^{def})^2+(s_{tt}^{def}+\langle E\rangle^{-1}L_t)(s_{tl}^{abc})^2\right)}. \tag*{(25)} B ( 3 ) = − ( 2 E 12 − 1 + L n ) ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tt d e f + ⟨ E ⟩ − 1 L t ) − 2 1 ⟨ E ⟩ ( ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tl d e f ) 2 + ( s tt d e f + ⟨ E ⟩ − 1 L t ) ( s tl ab c ) 2 ) 2 ⟨ ν ⟩ ( s tt ab c + ⟨ E ⟩ − 1 L t ) ( s tt d e f + ⟨ E ⟩ − 1 L t ) + 2 1 ( 1 + L t − ⟨ ν ⟩) ( ( s tt ab c + ⟨ E ⟩ − 1 L t ) s tl d e f + ( s tt d e f + ⟨ E ⟩ − 1 L t ) s tl ab c ) . (25)
for
s t t h k l : = 1 2 ( s 1111 G B , h k l + s 2222 G B , h k l ) + s 1122 G B , h k l , s_{tt}^{hkl}:=\frac{1}{2}\left(s_{1111}^{\mathrm{GB},hkl}+s_{2222}^{\mathrm{GB},hkl}\right)+s_{1122}^{\mathrm{GB},hkl}, s tt hk l := 2 1 ( s 1111 GB , hk l + s 2222 GB , hk l ) + s 1122 GB , hk l ,
s t l h k l : = s 1133 G B , h k l + s 2233 G B , h k l , s_{tl}^{hkl}:=s_{1133}^{\mathrm{GB},hkl}+s_{2233}^{\mathrm{GB},hkl}, s tl hk l := s 1133 GB , hk l + s 2233 GB , hk l ,
s l l h k l : = s 3333 G B , h k l : = E h k l − 1 . (26) s_{ll}^{hkl}:=s_{3333}^{\mathrm{GB},hkl}:=E_{hkl}^{-1}. \tag*{(26)} s l l hk l := s 3333 GB , hk l := E hk l − 1 . (26)
The combinations of compliance-tensor components11 , introduced in Eq. (26) , are related through a material dependent (but GB type independent) linear combination
2 s t t h k l + 2 s t l h k l + s l l h k l = ( s 1111 c r y + s 2222 c r y + s 3333 c r y ) + 2 ( s 1122 c r y + s 1133 c r y + s 2233 c r y ) , (27) 2s^{hkl}_{tt}+2s^{hkl}_{tl}+s^{hkl}_{ll}=(s^{\mathrm{cry}}_{1111}+s^{\mathrm{cry}}_{2222}+s^{\mathrm{cry}}_{3333})+2(s^{\mathrm{cry}}_{1122}+s^{\mathrm{cry}}_{1133}+s^{\mathrm{cry}}_{2233}), \tag*{(27)} 2 s tt hk l + 2 s tl hk l + s l l hk l = ( s 1111 cry + s 2222 cry + s 3333 cry ) + 2 ( s 1122 cry + s 1133 cry + s 2233 cry ) , (27)
which suggests that σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) is a function of (at most) four local GB parameters (in addition to bulk properties ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ , ⟨ ν ⟩ \langle\nu\rangle ⟨ ν ⟩ and chain parameters L n L_n L n , L t L_t L t ). In the L t → ∞ L_t \to\infty L t → ∞ limit, σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) reduces to σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) , see Eq. (21) .
The corresponding first two statistical moments can also be expressed analytically (but they are not shown here for brevity). They have the already familiar loading dependence,
⟨ σ n n ( 3 ) ⟩ ∼ tr ( Σ l a b ) 3 , \left\langle\sigma_{nn}^{(3)} \right\rangle\sim\frac{\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}})}{3}, ⟨ σ nn ( 3 ) ⟩ ∼ 3 tr ( Σ lab ) ,
s ( σ n n ( 3 ) ) ∼ 2 Σ m i s l a b 3 5 . (28) s(\sigma_{nn}^{(3)}) \sim\frac{2\Sigma_{\mathrm{mis}}^{\mathrm{lab}}}{3\sqrt{5}}. \tag*{(28)} s ( σ nn ( 3 ) ) ∼ 3 5 2 Σ mis lab . (28)
Expressions simplify further for higher lattice symmetries. For cubic lattices, for example, A ( 3 ) A^{(3)} A ( 3 ) and B ( 3 ) B^{(3)} B ( 3 ) become (again) only functions of Young’s moduli E a b c E_{abc} E ab c and E d e f E_{def} E d e f along the GB-normal direction (see Appendix § 0.3.3 )12 .
All compact-form solutions σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) , representing the special limits of the general solution, Eq. (11) , are summarized in Table II.
Models validation
In this section, the solutions σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) of derived models are tested against numerical results.13 For demonstration purposes, only cubic elastic materials are chosen for comparison (see Appendix § 0.3.4 for the corresponding elastic properties).
Following the derived expressions, Eqs. (13) , (20) and (23) , the mean value ⟨ σ n n ⟩ \langle\sigma_{nn}\rangle ⟨ σ nn ⟩ and standard deviation s ( σ n n ) s(\sigma_{nn}) s ( σ nn ) of PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) should depend trivially on the external loading Σ l a b \boldsymbol{\Sigma}^{\mathrm{lab}} Σ lab . Using suggested normalization, ⟨ σ n n ⟩ / tr ( Σ l a b ) \langle\sigma_{nn}\rangle/\operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}}) ⟨ σ nn ⟩ / tr ( Σ lab ) and s ( σ n n ) / Σ m i s l a b s(\sigma_{nn})/\Sigma_{\mathrm{mis}}^{\mathrm{lab}} s ( σ nn ) / Σ mis lab , the first two statistical moments become independent of Σ l a b \boldsymbol{\Sigma}^{\mathrm{lab}} Σ lab , which is demonstrated in Figure 7 for 27 different loading configurations.
Figure 7. Effect of external loading can be decoupled from other influences by a suitable choice of normalization factor (a) tr ( Σ l a b ) \operatorname{tr}(\boldsymbol{\Sigma}^{\mathrm{lab}}) tr ( Σ lab ) for mean value ⟨ σ n n ⟩ \langle\sigma_{nn}\rangle ⟨ σ nn ⟩ and (b) Σ m i s l a b \Sigma_{\mathrm{mis}}^{\mathrm{lab}} Σ mis lab for standard deviation s ( σ n n ) s(\sigma_{nn}) s ( σ nn ) . Simulation results are shown for Fe, 27 different external loadings Σ l a b \boldsymbol{\Sigma}^{\mathrm{lab}} Σ lab and three GB types. Non-normalized values are shown in the insets (a) and (b). Panel (c) shows correspondence between Σ l a b \boldsymbol{\Sigma}^{\mathrm{lab}} Σ lab and loading index (1-27).
Very good agreement14 between the prediction and numerical results confirms the validity of the derived expressions being of the form σ n n ( k ) = A ( k ) Σ z z + B ( k ) ( Σ x x + Σ y y ) \sigma_{nn}^{(k)}=A^{(k)}\Sigma_{zz}+B^{(k)}(\Sigma_{xx}+\Sigma_{yy}) σ nn ( k ) = A ( k ) Σ z z + B ( k ) ( Σ xx + Σ y y ) for any k k k . Hence, a tensile loading Σ \boldsymbol{\Sigma} Σ will be used hereafter without the loss of generality.
In Figure 8 the normalized standard deviation s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) is shown for polycrystalline Li (cubic symmetry) as a function of effective GB stiffness parameter E 12 E_{12} E 12 , which is a single characteristic parameter of the [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GB. Results of different models from Table II are compared with the results of finite element simulations. The Li is chosen because of very high elastic anisotropy (A u = 7.97 A^{u}=7.97 A u = 7.97 ), which makes the comparison more challenging.
Figure 8. Standard deviation s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) as a function of effective GB-stiffness parameter E 12 E_{12} E 12 . A comparison is shown between numerical results (FE) and different model predictions from Table II. The results are evaluated for Li on all GBs of a specific type, and thus corresponding to a certain E 12 E_{12} E 12 value. Six GB types are used in total (here, solid lines are unphysical and are meant only to indicate the trend).
Although none of model predictions for s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) are very accurate, some of the models are more appropriate than others. The L n L_n L n -L t L_t L t -chain (full version) model results are grouped into three families (with a given color) with a common axial chain length L n = 0 L_n=0 L n = 0 , 2 or 5. While the response of the L n = 0 L_n=0 L n = 0 (red) family is too steep for all transverse chain lengths L t L_t L t , overestimating the s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) at large E 12 E_{12} E 12 , the response of L n = 2 L_n=2 L n = 2 (green)
Table II. A summary of derived models. Analytical solutions can be written in a compact form only in certain limits.
In contrast to the full version, the reduced version of the model eliminates the twist-angle degrees of freedom, which makes the solution only approximate, but significantly more condensed. Note that both versions provide the same mean value ⟨ σ n n ( k ) ⟩ \left\langle\sigma_{nn}^{(k)}\right\rangle ⟨ σ nn ( k ) ⟩ .
and L n = 5 L_{n}=5 L n = 5 (blue) families is too gradual for L t ≲ 2 L_{t}\lesssim2 L t ≲ 2 , overestimating the s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) at small E 12 E_{12} E 12 . These models are recognized as inappropriate. In addition, all three model families show, for L t = 0 L_{t}=0 L t = 0 , a sudden change in the slope of s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) , which is not observed numerically, suggesting that L t = 0 L_{t}=0 L t = 0 models are also unsuitable. Most favorable are therefore 2 ≲ L n ≲ 5 2\lesssim L_{n}\lesssim5 2 ≲ L n ≲ 5 , L t ≳ 1 L_{t}\gtrsim1 L t ≳ 1 models, which predict s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) consistently below the numerical curve. This systematic underestimation of fluctuations is compensated later in § 0.3 D 2 by accounting for loading fluctuations, which are generated by external loading mediated through the anisotropic bulk surrounding the L n L_{n} L n -L t L_{t} L t -chain model (see last stage in Figure 4 ).
In Figure 8 also the results of the L n L_{n} L n -chain (reduced version) model are shown for comparison. The advantage of the latter is the compact formulation of the σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) and its statistical moments. The resulting s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) curves
show similar dependence of E 12 E_{12} E 12 as the corresponding L n L_{n} L n -L t → ∞ L_{t}\to\infty L t → ∞ (full version) models, however, with slightly reduced fluctuations in the mid-E 12 E_{12} E 12 range18 . Since in either case additional fluctuations need to be added to fit the numerical results, also the validity of the L n L_{n} L n -chain model, with 2 ≲ L n ≲ 5 2\lesssim L_{n}\lesssim5 2 ≲ L n ≲ 5 , is considered appropriate. In Figure 9 a response of a single [112]-[112]-Δ ω \Delta\omega Δ ω GB type is shown in terms of s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) as a function of Δ ω \Delta\omega Δ ω , using numerical simulations and model predictions from Table II for a polycrystalline Li to associate with results of Figure 8 . According to the numerical curve, very small variations in s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) are observed across the whole Δ ω \Delta\omega Δ ω range, which is consistent with previous results [28 ] . Using the same coloring and labeling scheme as in Figure 8 , a very good agreement with simulations is achieved for the L n L_n L n -L t L_t L t -chain model for L n = 2 L_n=2 L n = 2 and L t = 1 L_t=1 L t = 1 (see the inset of Figure 9 ). The other two families of curves produce either too big (L n = 0 L_n=0 L n = 0 , in red) or too small (L n = 5 L_n=5 L n = 5 , in blue) variations across the Δ ω \Delta\omega Δ ω range. Since the twist angle degrees of freedom are integrated out, the response of the L n L_n L n -chain (reduced version) model is independent of Δ ω \Delta\omega Δ ω , which is, by design, also in good agreement with numerical results.
Figure 9. Standard deviation s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) as a function of twist-angle difference Δ ω \Delta\omega Δ ω associated with the [112]-[112]-Δ ω \Delta\omega Δ ω GB type. A comparison is shown between numerical results (FE) and different model predictions from Table II. The properties of Li are used. Note that [112]-[112]-Δ ω \Delta\omega Δ ω GBs correspond to E 12 = 0.77 E_{12}=0.77 E 12 = 0.77 , irrespective of the value of Δ ω \Delta\omega Δ ω . Inset shows the agreement between the FE result and the response of L n L_{n} L n -L t L_{t} L t -chain model for L n = 2 L_{n}=2 L n = 2 and L t = 1 L_{t}=1 L t = 1 (note the artificial shift accounting for the missing loading fluctuations).
Results of Figure 8 and Figure 9 seem to favor the L n L_n L n -L t L_t L t -chain model with L n ∼ 2 L_n\sim2 L n ∼ 2 and L t ∼ 1 L_t\sim1 L t ∼ 1 . This is further corroborated by noting that the overall shift in s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) (by ∼ 0.1 \sim0.1 ∼ 0.1 ), used to fit the simulation results in the inset of Figure 9 , is matching very well the gap at E 12 = 0.77 E_{12}=0.77 E 12 = 0.77 (corresponding to [112]-[112]-Δ ω \Delta\omega Δ ω GB) between the two corresponding curves in Figure 8 .
In the following, the evaluation of derived models is shifted from macro- to mesoscale using a linear correlation property, σ n n ( k ) / Σ = A ( k ) + B ( k ) cos 2 θ \sigma_{nn}^{(k)}/\Sigma=A^{(k)}+B^{(k)}\cos^2\theta σ nn ( k ) /Σ = A ( k ) + B ( k ) cos 2 θ , derived for the external uniaxial loading Σ \Sigma Σ 19 . Statistical analysis employed on a subset of GBs with a fixed angle θ \theta θ (or cos 2 θ \cos^2\theta cos 2 θ ) between the GB normal and uniaxial loading direction is useful because it allows one to test the validity
of individual parts of expressions in σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) (e.g., A ( k ) A^{(k)} A ( k ) and B ( k ) B^{(k)} B ( k ) ). Such analyses are demonstrated in Figure 10 and Figure 11 for polycrystalline Li.
Figure 10. (a) Mean local stress ⟨ σ n n / Σ ⟩ \langle\sigma_{nn}/\Sigma\rangle ⟨ σ nn /Σ ⟩ and (b) corresponding standard deviation s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) , both evaluated numerically as a function of cos 2 θ \cos^2\theta cos 2 θ for a finite range of GB tilt angles, δ ( cos θ ) = 0.05 \delta(\cos\theta)=0.05 δ ( cos θ ) = 0.05 . The averaging range is denoted by horizontal error bars and the averaged values by dots. Lines in panel (a) are linear fits with slope K K K . Twist angle difference Δ ω \Delta\omega Δ ω in [112]-[112]-Δ ω \Delta\omega Δ ω GB has a negligible influence on slope K K K (inset (a)) but a significant effect on the standard deviation (inset (b)).
Figure 11. Slope K K K , obtained from Figure 10 (a), versus effective GB-stiffness parameter E 12 E_{12} E 12 . A comparison is shown between numerical results (FE) and different model predictions from Table II. The properties of Li are used. Note that the L n L_n L n -chain model (reduced version) and L n L_n L n -L t L_t L t -chain model (full version) with L t → ∞ L_t\to\infty L t → ∞ provide identical slopes K = 1 + 3 ( E 12 − 1 ) / ( 2 + L n E 12 ) K=1+3(E_{12}-1)/(2+L_nE_{12}) K = 1 + 3 ( E 12 − 1 ) / ( 2 + L n E 12 ) , assuming E 3 = 1 E_3=1 E 3 = 1 . The lines show no effect of L n L_n L n when E 12 = E 3 = 1 E_{12}=E_3=1 E 12 = E 3 = 1 .
The local mean ⟨ σ n n / Σ ⟩ \langle\sigma_{nn}/\Sigma\rangle ⟨ σ nn /Σ ⟩ and standard deviation20 s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) are shown in Figure 10 as a function of cos 2 θ \cos^2\theta cos 2 θ . Due to finite aggregate size, the mean and standard deviation variation are obtained at given cos 2 θ \cos^2\theta cos 2 θ by averaging over Euler angles ψ \psi ψ and ϕ \phi ϕ on a finite (but small) range of GB tilt angles δ ( cos θ ) = 0.05 \delta(\cos\theta)=0.05 δ ( cos θ ) = 0.05 . The proposed linear trend is nicely reproduced, showing a clear effect of different GB types on the corresponding slopes K K K of fitted lines. In general, slope K K K increases with increasing GB stiffness (parameter E 12 E_{12} E 12 , see also Figure 11 ). However, there is a very weak effect of Δ ω \Delta\omega Δ ω on the corresponding slope K K K when evaluated on the [112]-[112]-Δ ω \Delta\omega Δ ω GB. This suggests that, on average, the GB stiffness (which is independent of Δ ω \Delta\omega Δ ω , see Eq. (43) ) is the main contributor to σ n n \sigma_{nn} σ nn at given cos 2 θ \cos^2\theta cos 2 θ .
It is interesting to note a crossing point in Figure 10 (a) at cos 2 θ = 1 / 3 \cos^2\theta=1/3 cos 2 θ = 1/3 at which σ n n \sigma_{nn} σ nn becomes independent of both material and GB type properties. This point is exactly reproduced by all non-trivial models (σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) , k > 0 k>0 k > 0 ). The value of σ n n \sigma_{nn} σ nn at this point is Σ / 3 \Sigma/3 Σ/3 (actually tr ( Σ ) / 3 \operatorname{tr}(\Sigma)/3 tr ( Σ ) /3 for arbitrary loading).
The simulation results from Figure 10 (a) are analyzed further in Figure 11 where the actual dependence of slope K K K with E 12 E_{12} E 12 is presented and compared with predictions of the models from Table II. While the increasing trend is captured well by all the models, none of the presented curves fit the numerical result very accurately for all E 12 E_{12} E 12 . Actually, this holds true for any combination of L n L_n L n , L t L_t L t values in the L n L_n L n -L t L_t L t -chain model. The most suitable solution is chosen to be that of the L n L_n L n -chain model (reduced version) for L n = 2 L_n = 2 L n = 2 , which matches the true K K K values at the two extreme E 12 E_{12} E 12 points. The same agreement is observed also for other cubic materials (not shown).
The L n L_n L n -L t L_t L t -chain model with L n ∼ 2 L_n \sim2 L n ∼ 2 and L t ∼ 1 L_t \sim1 L t ∼ 1 , which has been selected as the most suitable model at the macroscale (see Figure 8 and Figure 9 ), provides in Figure 11 a very similar K ( E 12 ) K(E_{12}) K ( E 12 ) response as the L n L_n L n -chain model for L n = 2 L_n = 2 L n = 2 . In this sense, both models seem equally well acceptable, however, the latter one will be preferred due to much more compact formulation.
In summary, although the qualitative behavior of σ n n \sigma_{nn} σ nn is well reproduced by the selected model (σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) for L n ∼ 2 L_n \sim2 L n ∼ 2 ) on a wide range of parameters (associated with external loading, material properties and GB type), two ingredients still seem to be missing. The first one is related to the systematic shortage of stress fluctuations observed on the macroscale and the second one is linked to the insufficient agreement of mean stresses on the mesoscale. Both issues are addressed in the next section.
Model upgrades
Variable axial strain constraint and 3D effects
The observed inconsistency in Figure 11 is attributed to the (i) imposed axial strain constraint of the L n L_n L n -chain model and (ii) 3D effects which have been omitted in the model derivation. The 3D effects include primarily a non-zero lateral coupling of the axial grain chain with the elastic bulk. Depending on the relative axial stiffness of the chain with respect to the bulk, this coupling may effectively either increase or decrease the chain stiffness, resulting in larger or lower σ n n \sigma_{nn} σ nn , respectively. To model this in 1D framework, the elastic properties of both GB grains and buffer grain need to be amended21 . Regarding the GB grains, therefore
E 12 → E 12 + δ E 12 , δ E 12 = F ( E 12 , ν 12 , L n , A u ) , (29) E_{12} \to E_{12} + \delta E_{12},
\qquad
\delta E_{12} = \mathcal{F}(E_{12}, \nu_{12}, L_n, A^u), \tag*{(29)} E 12 → E 12 + δ E 12 , δ E 12 = F ( E 12 , ν 12 , L n , A u ) , (29)
and similarly
ν 12 → ν 12 + δ ν 12 , δ ν 12 = F ( E 12 , ν 12 , L n , A u ) . (30) \nu_{12} \to\nu_{12} + \delta\nu_{12},
\qquad
\delta\nu_{12} = \mathcal{F}(E_{12}, \nu_{12}, L_n, A^u). \tag*{(30)} ν 12 → ν 12 + δ ν 12 , δ ν 12 = F ( E 12 , ν 12 , L n , A u ) . (30)
The assumed functional dependence of δ E 12 \delta E_{12} δ E 12 in Eq. (29) can be explained with the help of Figure 12 where the field lines are used to visualize schematically the force field around the two GB grains under tensile loading. While force lines are always parallel in the 1D model (no lateral coupling with the bulk), they concentrate within/outside the stiffer/softer (larger/smaller E 12 E_{12} E 12 ) GB grains in the 3D model. Obviously, the effect gets stronger for E 12 → E 12 , max E_{12} \to E_{12,\max} E 12 → E 12 , m a x or E 12 → E 12 , min E_{12} \to E_{12,\min} E 12 → E 12 , m i n and for increasing material anisotropy A u A^u A u . To account for more (less) field lines in stiffer (softer) GB grains, δ E 12 > 0 \delta E_{12} > 0 δ E 12 > 0 (δ E 12 < 0 \delta E_{12} < 0 δ E 12 < 0 ) should be used in 1D modeling. However, using a non-zero δ E 12 \delta E_{12} δ E 12 (or δ ν 12 \delta\nu_{12} δ ν 12 ) affects also the boundary condition applied on the chain scale in Eq. (11). Since the latter is regulated also by the length of the buffer grain L n L_n L n , both δ E 12 \delta E_{12} δ E 12 and L n L_n L n are coupled as indicated in Eq. (29) .
Figure 12. Schematic view of field lines crossing through the stiff (orange) and soft (green) GB grains in 3D and 1D models.
In a similar way, the properties of the buffer grain (of length L n L_n L n ) are modified due to lateral coupling with the bulk
E 3 → E 3 + δ E 3 ≈ 1 , ν b → ν b + δ ν b ≈ ⟨ ν ⟩ . (31) E_3 \to E_3 + \delta E_3 \approx1,
\qquad
\nu_b \to\nu_b + \delta\nu_b \approx\langle\nu\rangle. \tag*{(31)} E 3 → E 3 + δ E 3 ≈ 1 , ν b → ν b + δ ν b ≈ ⟨ ν ⟩ . (31)
The above mapping follows from the fact that a chain of randomly oriented grains, when coupled laterally to the bulk, should, on average, behave similarly as the bulk itself. An equality in Eq. (31) is achieved for L n → ∞ L_n \to\infty L n → ∞ , while very small deviations are observed at L n = 2 L_n = 2 L n = 2 (see footnote 20). As already mentioned, this modification has been already implemented in Eq. (21) by setting E 3 = 1 E_3 = 1 E 3 = 1 and ν b = ⟨ ν ⟩ \nu_b = \langle\nu\rangle ν b = ⟨ ν ⟩ .
Since E b E_b E b and ν b \nu_b ν b can be evaluated (e.g., numerically) by Eqs. (22) for a given material and L n L_n L n , the corresponding increments can be estimated directly from Eqs. (31) . By design, the same increments should also apply to GB grains, ( δ E 12 , δ ν 12 ) = ( δ E b , δ ν b ) (\delta E_{12}, \delta\nu_{12}) = (\delta E_b, \delta\nu_b) ( δ E 12 , δ ν 12 ) = ( δ E b , δ ν b ) , if ( E 12 , ν 12 ) = ( E b , ν b ) (E_{12}, \nu_{12}) = (E_b, \nu_b) ( E 12 , ν 12 ) = ( E b , ν b ) . Unfortunately, there seems to be no analytical approach to identify the increments for a general pair ( E 12 , ν 12 ) (E_{12}, \nu_{12}) ( E 12 , ν 12 ) . In the following, the functional dependence of δ E 12 \delta E_{12} δ E 12 (and δ E 3 \delta E_3 δ E 3 ) is therefore derived empirically for materials with cubic lattice symmetry, where further simplification is used due to mutual dependence of E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 (ν 12 = ⟨ ν ⟩ + ( E 12 − 1 − 1 ) / 2 \nu_{12} = \langle\nu\rangle+ (E_{12}^{-1} - 1)/2 ν 12 = ⟨ ν ⟩ + ( E 12 − 1 − 1 ) /2 ).
The expression for the reduced version of the L n L_n L n -chain model, Eq. (21) , simplifies for cubic materials and general macroscopic loading to
σ n n ( 2 ) = 2 + L n 2 E 12 − 1 + L n E 3 − 1 Σ z z + 1 2 ( 1 − 2 + L n 2 E 12 − 1 + L n E 3 − 1 ) ( Σ x x + Σ y y ) , (32) \sigma_{nn}^{(2)} = \frac{2+L_n}{2E_{12}^{-1}+L_nE_3^{-1}}\Sigma_{zz}
+\frac{1}{2}\left(1-\frac{2+L_n}{2E_{12}^{-1}+L_nE_3^{-1}}\right)(\Sigma_{xx}+\Sigma_{yy}), \tag*{(32)} σ nn ( 2 ) = 2 E 12 − 1 + L n E 3 − 1 2 + L n Σ z z + 2 1 ( 1 − 2 E 12 − 1 + L n E 3 − 1 2 + L n ) ( Σ xx + Σ y y ) , (32)
which reduces further for uniaxial macroscopic loading Σ \Sigma Σ as
σ n n ( 2 ) / Σ = 1 2 ( 1 − 2 + L n 2 E 12 − 1 + L n E 3 − 1 ) + 3 2 ( − 1 3 + 2 + L n 2 E 12 − 1 + L n E 3 − 1 ) cos 2 θ , (33) \sigma_{nn}^{(2)}/\Sigma= \frac{1}{2}\left(1-\frac{2+L_n}{2E_{12}^{-1}+L_nE_3^{-1}}\right)
+\frac{3}{2}\left(-\frac{1}{3}+\frac{2+L_n}{2E_{12}^{-1}+L_nE_3^{-1}}\right)\cos^2\theta, \tag*{(33)} σ nn ( 2 ) /Σ = 2 1 ( 1 − 2 E 12 − 1 + L n E 3 − 1 2 + L n ) + 2 3 ( − 3 1 + 2 E 12 − 1 + L n E 3 − 1 2 + L n ) cos 2 θ , (33)
where θ \theta θ is an angle between the GB normal and uniaxial loading direction. As discussed before, the model can be upgraded by assuming δ E 12 = F ( E 12 , L n , A u ) \delta E_{12}=F(E_{12},L_n,A^u) δ E 12 = F ( E 12 , L n , A u ) for the two GB grains and δ E 3 = F ( E 3 , L n , A u ) \delta E_3=F(E_3,L_n,A^u) δ E 3 = F ( E 3 , L n , A u ) for the buffer grain. Both increments can be calculated numerically from the requirement that the resulting modified slope (a factor in front of cos 2 θ \cos^2\theta cos 2 θ ),
K = 3 2 ( − 1 3 + 2 + L n 2 ( E 12 + δ E 12 ) − 1 + L n ( E 3 + δ E 3 ) − 1 ) , (34) K=\frac{3}{2}\left(-\frac{1}{3}+\frac{2+L_n}{2(E_{12}+\delta E_{12})^{-1}+L_n(E_3+\delta E_3)^{-1}}\right), \tag*{(34)} K = 2 3 ( − 3 1 + 2 ( E 12 + δ E 12 ) − 1 + L n ( E 3 + δ E 3 ) − 1 2 + L n ) , (34)
is matching the corresponding K F E K^{\mathrm{FE}} K FE slope obtained from the FE simulations for different E 12 E_{12} E 12 values and materials (see Figure 11 where the results for Li are shown)[19 ] .
The results for δ E 12 \delta E_{12} δ E 12 are shown in Figure 13 for L n = 2 L_n = 2 L n = 2 and various materials[20 ] . As anticipated, δ E 12 \delta E_{12} δ E 12 depends strongly on the GB stiffness E 12 E_{12} E 12 , elastic grain anisotropy A u A^u A u and buffer length L n L_n L n (inset of Figure 13 ). Interestingly, for L n = 2 L_n = 2 L n = 2 a (quasi) symmetry is recognized in δ E 12 ( E 12 ) \delta E_{12}(E_{12}) δ E 12 ( E 12 ) curves for all investigated materials[21 ] . The symmetry is lost when L n ≠ 2 L_n \ne2 L n = 2 .
Figure 13. Calculated (dots) and fitted (lines) increments δ E 12 \delta E_{12} δ E 12 as a function of E 12 E_{12} E 12 for the chosen reduced L n L_n L n -chain model (with L n = 2 L_n = 2 L n = 2 ) applied to various materials with cubic lattice symmetry. Fitting function, Eq. (35) , has been chosen based on the observed symmetry δ E 12 ( E 12 ) \delta E_{12}(E_{12}) δ E 12 ( E 12 ) for L n = 2 L_n = 2 L n = 2 . Inset shows that very good agreement is preserved also when the symmetric L n = 2 L_n = 2 L n = 2 fitting function is used for other L n L_n L n (shown for Li).
Based on the observed symmetry in Figure 13 for L n = 2 L_n = 2 L n = 2 , the following empirical fit is proposed for all cubic materials with corresponding elastic anisotropy index A u A^u A u
δ E 12 = C 1 − ∣ E 12 − E ˉ 12 ∣ C 2 , \delta E_{12}=C_1-\lvert E_{12}-\bar{E}_{12}\rvert^{C_2}, δ E 12 = C 1 − ∣ E 12 − E ˉ 12 ∣ C 2 ,
E ˉ 12 = 1 2 ( E 12 , min + E 12 , max ) , (35) \bar{E}_{12}=\frac{1}{2}(E_{12,\min}+E_{12,\max}), \tag*{(35)} E ˉ 12 = 2 1 ( E 12 , m i n + E 12 , m a x ) , (35)
δ E 12 ( E 12 , min ) = 0 , \delta E_{12}(E_{12,\min})=0, δ E 12 ( E 12 , m i n ) = 0 ,
δ E 12 ( E 12 , max ) = 0. \delta E_{12}(E_{12,\max})=0. δ E 12 ( E 12 , m a x ) = 0.
The best agreement with FE results is obtained for
C 1 = 0.08 ( A u ) 0.85 , C_1=0.08(A^u)^{0.85}, C 1 = 0.08 ( A u ) 0.85 ,
C 2 = log C 1 log ( E 12 , max − E 12 , min ) − log 2 . (36) C_2=\frac{\log C_1}{\log(E_{12,\max}-E_{12,\min})-\log2}. \tag*{(36)} C 2 = log ( E 12 , m a x − E 12 , m i n ) − log 2 log C 1 . (36)
It seems quite surprising that the proposed fitting function, Eq. (35) , with only two adjustable parameters (0.08 and 0.85) in Eq. (36) provides such a good agreement for a wide range of (cubic) materials shown in Figure 13 . Good agreement remains also when the L n = 2 L_n=2 L n = 2 fitting function is used for L n ≠ 2 L_n\ne2 L n = 2 models (assuming E 3 + δ E 3 = 1 E_3+\delta E_3=1 E 3 + δ E 3 = 1 in Eq. (34) ) as shown in the inset of Figure 13 . The identified empirical relation represents the second main result of this study.
Stochastic loading fluctuations
So far, the original external loading Σ l a b \Sigma^{\mathrm{lab}} Σ lab (also Σ \Sigma Σ ) has been assigned to all GB models from Table II. However, in reality, this assumption is true only on average. In fact, a GB and its immediate neighbourhood far away from the external surfaces feel an external loading modified by fluctuations, Σ + f \Sigma+f Σ + f , where f f f stands for the fluctuation stress tensor. The fluctuations f f f arise as a consequence of bringing far-away loading Σ \Sigma Σ onto a GB neighbourhood through the elastic bulk of anisotropic grains (see the last stage in Figure 4 ). To account for loading fluctuations in the estimation of σ n n \sigma_{nn} σ nn and PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) , it is assumed for simplicity that fluctuation normal stress f n n f_{nn} f nn is a random variable with Gaussian distribution N ( 0 , s 2 ( f n n ) ) \mathcal{N}(0,s^{2}(f_{nn})) N ( 0 , s 2 ( f nn )) , where the standard deviation depends on the external loading and on grain anisotropy, s ( f n n ) ≈ F ( Σ , A u ) s(f_{nn}) \approx F(\Sigma,A^{u}) s ( f nn ) ≈ F ( Σ , A u ) . Dependence of s ( f n n ) s(f_{nn}) s ( f nn ) on internal GB degrees of freedom (e.g., θ \theta θ , E 12 E_{12} E 12 , ω 1 \omega_{1} ω 1 , ω 2 \omega_{2} ω 2 ) is neglected to a first approximation, which is supported by the results of Figure 10 (b). The latter indeed show that standard deviation s ( σ n n / Σ ) s(\sigma_{nn}/\Sigma) s ( σ nn /Σ ) , evaluated on various [ a b c ] [abc] [ ab c ] -[ d e f ] [def] [ d e f ] GB types at fixed GB tilts cos 2 θ \cos^{2}\theta cos 2 θ with respect to external tensile loading Σ \Sigma Σ , is practically independent of cos 2 θ \cos^{2}\theta cos 2 θ (and thus of σ n n \sigma_{nn} σ nn itself), but slightly dependent on GB type23 . Model of stress fluctuations is derived in Appendix § 2.6 .
Considering that stress fluctuations are independent of stresses themselves, a new update can be proposed as
σ ~ n n ( k ) = σ n n ( k ) + f n n ( k ) , s 2 ( σ ~ n n ( k ) ) = s 2 ( σ n n ( k ) ) + s 2 ( f n n ( k ) ) , PDF ( σ ~ n n ( k ) ) = ( PDF ∗ N ( 0 , s 2 ( f n n ( k ) ) ) ) ( σ n n ( k ) ) , (37) \begin{aligned}
\widetilde{\sigma}_{nn}^{(k)} &= \sigma_{nn}^{(k)} + f_{nn}^{(k)},\\
s^{2}\left(\widetilde{\sigma}_{nn}^{(k)}\right) &= s^{2}\left(\sigma_{nn}^{(k)}\right) + s^{2}\left(f_{nn}^{(k)}\right),\\
\operatorname{PDF}\left(\widetilde{\sigma}_{nn}^{(k)}\right) &= \left(\operatorname{PDF} * \mathcal{N}\left(0,s^{2}\left(f_{nn}^{(k)}\right)\right)\right)\left(\sigma_{nn}^{(k)}\right), \tag*{(37)}
\end{aligned} σ nn ( k ) s 2 ( σ nn ( k ) ) PDF ( σ nn ( k ) ) = σ nn ( k ) + f nn ( k ) , = s 2 ( σ nn ( k ) ) + s 2 ( f nn ( k ) ) , = ( PDF ∗ N ( 0 , s 2 ( f nn ( k ) ) ) ) ( σ nn ( k ) ) , (37)
where symbol ∗ \ast ∗ denotes a convolution.
To identify standard deviation s ( f n n ( k ) ) s(f_{nn}^{(k)}) s ( f nn ( k ) ) for tensile Σ \Sigma Σ , the results from Figure 10 (b) can be averaged over different GB tilts and shown in Figure 14 for different materials as a function of A u A^{u} A u . Obtained standard deviation s ( f n n / Σ ) s(f_{nn}/\Sigma) s ( f nn /Σ ) is set to be a measure of local stress fluctuations f n n f_{nn} f nn . As expected, s ( f n n / Σ ) s(f_{nn}/\Sigma) s ( f nn /Σ ) increases with A u A^{u} A u following a simple empirical law s ( f n n / Σ ) = ( 0.070 ± 0.018 ) ( A u ) 0.37 ∓ 0.03 s(f_{nn}/\Sigma)=(0.070\pm0.018)(A^{u})^{0.37\mp0.03} s ( f nn /Σ ) = ( 0.070 ± 0.018 ) ( A u ) 0.37 ∓ 0.03 . The ± \pm ± sign denotes a finite width of s ( f n n / Σ ) s(f_{nn}/\Sigma) s ( f nn /Σ ) domain, which is attributed to GB internal degrees of freedom.
Figure 14. Standard deviation of GB-stress (normal) fluctuations s ( f n n / Σ ) s(f_{nn}/\Sigma) s ( f nn /Σ ) , evaluated numerically on different GB types with fixed GB tilt (and later averaged over different GB tilts, see Figure 10 (b)) and for different materials (A u A^{u} A u ) under tensile loading Σ \Sigma Σ . A shaded region represents the proposed empirical domain of fluctuations. A generalization to arbitrary loading can be done by substituting Σ \Sigma Σ with Σ m i s \Sigma_{\mathrm{mis}} Σ mis on the vertical axis (see Appendix § 2.6 ).
The empirical fit is generalized further to arbitrary loading using the familiar normalization for the second statistical moment (see Appendix § 2.6 for more detail),
s ( f n n ) = Σ m i s ( 0.070 ± 0.018 ) ( A u ) 0.37 ∓ 0.03 . (38) s(f_{nn})=\Sigma_{\mathrm{mis}}(0.070\pm0.018)(A^{u})^{0.37\mp0.03}. \tag*{(38)} s ( f nn ) = Σ mis ( 0.070 ± 0.018 ) ( A u ) 0.37 ∓ 0.03 . (38)
The above relation applies not only to cubic but also to non-cubic materials26 . For example, tensile fluctuations evaluated in calcium sulfate (CaSO4 _4 4 ), with orthorhombic lattice symmetry and A u = 2.78 A^{u}=2.78 A u = 2.78 , fit accurately within the proposed domain in Figure 14 .
Verification of upgraded models
Cubic materials
Statistical response PDF ( σ ~ n n ( 2 ) ) \operatorname{PDF}(\widetilde{\sigma}_{nn}^{(2)}) PDF ( σ nn ( 2 ) ) of the upgraded cubic GB model (using L n = 2 L_{n}=2 L n = 2 and E 3 + δ E 3 = 1 E_{3}+\delta E_{3}=1 E 3 + δ E 3 = 1 in Eq. (32) ),
σ ~ n n ( 2 ) = 2 ( E 12 + δ E 12 ) − 1 + 1 Σ z z + 1 2 ( 1 − 2 ( E 12 + δ E 12 ) − 1 + 1 ) ( Σ x x + Σ y y ) + f n n , (39) \widetilde{\sigma}_{nn}^{(2)}=\frac{2}{(E_{12}+\delta E_{12})^{-1}+1}\Sigma_{zz}+\frac{1}{2}\left(1-\frac{2}{(E_{12}+\delta E_{12})^{-1}+1}\right)(\Sigma_{xx}+\Sigma_{yy})+f_{nn}, \tag*{(39)} σ nn ( 2 ) = ( E 12 + δ E 12 ) − 1 + 1 2 Σ z z + 2 1 ( 1 − ( E 12 + δ E 12 ) − 1 + 1 2 ) ( Σ xx + Σ y y ) + f nn , (39)
where δ E 12 \delta E_{12} δ E 12 is estimated by Eqs. (35) , (36) and s ( f n n ) s(f_{nn}) s ( f nn ) by Eq. (38) , is verified in Figure 15 for polycrystalline Fe under different macroscopic loadings Σ \Sigma Σ . The predicted PDF ( σ ~ n n ( 2 ) ) \operatorname{PDF}(\widetilde{\sigma}_{nn}^{(2)}) PDF ( σ nn ( 2 ) ) distributions are calculated numerically using Monte Carlo sampling of the two27 Euler angles ( θ , ψ ) (\theta,\psi) ( θ , ψ ) , which are used to evaluate Σ x x \Sigma_{xx} Σ xx , Σ y y \Sigma_{yy} Σ y y and Σ z z \Sigma_{zz} Σ z z defined in Eq. (4) . An excellent agreement with simulation results is demonstrated, confirming the accuracy of the proposed model for arbitrary GB type, arbitrary (cubic) material and arbitrary macroscopic loading conditions.
Figure 15. Statistical stress distributions PDF ( σ n n ) \operatorname{PDF}(\sigma_{nn}) PDF ( σ nn ) evaluated on three different GB types in Fe for 12 different macroscopic loadings (grouped into purely diagonal, purely shear and mixed loadings Σ \Sigma Σ ). An excellent agreement is shown between simulation results (solid lines) and upgraded model predictions (dashed lines) for all the cases; cf. Eq. (39) .
In Figure 16 a comparison is shown for a polycrystalline Fe with elongated grains to verify the applicability of the derived models in materials with non-zero morphological texture (but with zero crystallographic texture). In this comparison, the PDF response is calculated on all GBs (random type) using the following simple relation (see Appendix § 2.7 )
Figure 16. Stress distributions PDF(σ n n / Σ \sigma_{nn}/\Sigma σ nn /Σ ) evaluated on all (random) GBs in Fe with grains elongated along the Z Z Z -axis (elongation factor λ z \lambda_z λ z ) for tensile loading Σ \Sigma Σ along the X X X -axis. An excellent agreement is shown between simulation results (black) and model predictions (red) for all three cases; cf. Eqs. (40) –(41) .
PDF r n d ( σ ~ n n ( k ) ) ≈ PDF ( σ ~ n n ( 0 ) ) . (40) \operatorname{PDF}_{\mathrm{rnd}}\left(\widetilde{\sigma}_{nn}^{(k)}\right)\approx\operatorname{PDF}\left(\widetilde{\sigma}_{nn}^{(0)}\right). \tag*{(40)} PDF rnd ( σ nn ( k ) ) ≈ PDF ( σ nn ( 0 ) ) . (40)
The response of random GBs is calculated using the convolution of the isotropic solution PDF(σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) ) and Gaussian distribution N ( 0 , s 2 ( f n n ) \mathcal{N}(0,s^2(f_{nn}) N ( 0 , s 2 ( f nn ) with s ( f n n ) s(f_{nn}) s ( f nn ) from Eq. (38) .28 The distributions are calculated numerically using Monte Carlo sampling of the two Euler angles ( θ , ψ ) (\theta,\psi) ( θ , ψ ) with the following distribution functions (see Appendix G)
f ( cos θ ) = λ z 2 ( 1 1 + ( λ z 2 − 1 ) cos 2 θ ) 3 / 2 , f ( ψ ) = 1 2 π , (41) \begin{aligned}
f(\cos\theta) &= \frac{\lambda_z}{2}\left(\frac{1}{1+(\lambda_z^2-1)\cos^2\theta}\right)^{3/2},\\
f(\psi) &= \frac{1}{2\pi}, \tag*{(41)}
\end{aligned} f ( cos θ ) f ( ψ ) = 2 λ z ( 1 + ( λ z 2 − 1 ) cos 2 θ 1 ) 3/2 , = 2 π 1 , (41)
for − 1 ≤ cos θ ≤ 1 -1 \le\cos\theta\le1 − 1 ≤ cos θ ≤ 1 and 0 ≤ ψ ≤ 2 π 0 \le\psi\le2\pi 0 ≤ ψ ≤ 2 π , with a scaling factor λ z > 0 \lambda_z > 0 λ z > 0 accounting for grain elongation along the Z Z Z -axis (λ z = 1 \lambda_z=1 λ z = 1 denoting no scaling).
Again, an excellent agreement with simulation results is demonstrated in Figure 16 , which confirms the accuracy of the proposed model when applied to materials with arbitrary morphological texture.
In Figure 17 , the accuracy of GB models is furthermore tested on a local GB scale using FE simulations of a polycrystalline Li under tensile loading Σ \Sigma Σ as a reference. In particular, three models (of increasing complexity) are compared: (i) the isotropic model σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) , (ii) the reduced and upgraded version of the L n L_n L n -chain (L n = 2 L_n=2 L n = 2 , L t → ∞ L_t\to\infty L t → ∞ ) model σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) and (iii) the full version of the L n L_n L n -L t L_t L t -chain (L n = 2 L_n=2 L n = 2 , L t = 1 L_t=1 L t = 1 ) model σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) . The accuracy of the models is tested locally by comparing σ n n ( k ) \sigma_{nn}^{(k)} σ nn ( k ) values with FE results σ n n F E \sigma_{nn}^{\mathrm{FE}} σ nn FE evaluated on individual GBs of particular type (three GB types are tested in total).
Figure 17. (a) Local σ n n / Σ \sigma_{nn}/\Sigma σ nn /Σ stress response in Li under macroscopic tensile loading Σ \Sigma Σ . A comparison is shown between FE simulation (lines) and the results of three GB models (shapes). The 50 largest GBs in a 4000-grain aggregate (see Figure 1 ) are shown, to which a specific GB type ([1 ] -[1 ] , [334]-[102] or [111]-[111]) is assigned in each panel. The GBs are sorted (indexed) in descending order with respect to FE results for stresses. (b) Probability distributions (PDF) of discrepancy between the model prediction and FE result for GB stress, demonstrating how accurate the three GB models are (locally). All 1631 special GBs are considered in the PDF (as opposed to 50 shown in (a)). Gaussian distributions with standard deviations s ( f n n / Σ ) s(f_{nn}/\Sigma) s ( f nn /Σ ) from Figure 14 are added for comparison (no fitting has been applied).
According to Figure 17 , both σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) and σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) models are comparable in accuracy, outperforming the simplest σ n n ( 0 ) \sigma_{nn}^{(0)} σ nn ( 0 ) model on softer [1 ] -[1 ] and stiffer [111]-[111] GBs. The uncertainties (deviations from true FE values) in the σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) model (and also σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) model) are of Gaussian type with zero mean and standard deviation (exactly!) equal to s ( f n n ) s(f_{nn}) s ( f nn ) from Figure 14 (see dashed lines in Figure 17 (b)). This confirms the validity (and consistency) of the σ n n ( 2 ) \sigma_{nn}^{(2)} σ nn ( 2 ) model, which is shown to be accurate up to unknown loading fluctuations f n n f_{nn} f nn (which are substantial in Li). The latter are therefore the only29 source of local stress uncertainties (errors), σ n n ( 2 ) − σ n n F E ≈ f n n \sigma_{nn}^{(2)}-\sigma_{nn}^{\mathrm{FE}}\approx f_{nn} σ nn ( 2 ) − σ nn FE ≈ f nn , suggesting that σ n n ( 2 ) ≈ σ n n F E \sigma_{nn}^{(2)}\approx\sigma_{nn}^{\mathrm{FE}} σ nn ( 2 ) ≈ σ nn FE .
Non-cubic materials
To provide accurate stress distributions for non-cubic materials, the evaluation of δ E 12 \delta E_{12} δ E 12 and δ ν 12 \delta\nu_{12} δ ν 12 would need to be derived (see Eqs. (29) and (30) ) to account for variable axial strain constraint and 3D effects missing in the L n L_n L n -chain model. The procedure for that should follow the one described for cubic materials in § 0.3 D 1. However, this is left for future analyses.
In Figure 19 the simulation results for average stress ⟨ σ n n / Σ ⟩ \langle\sigma_{nn}/\Sigma\rangle ⟨ σ nn /Σ ⟩ are presented which are evaluated on 190 [abc]-[def] GBs obtained as combinations of 19 planes defined in Figure 18 for the orthorhombic material CaSO4 _4 4 under tensile loading Σ \Sigma Σ . For comparison, a smooth prediction of ⟨ σ n n ( 2 ) / Σ ⟩ \langle\sigma_{nn}^{(2)}/\Sigma\rangle ⟨ σ nn ( 2 ) /Σ ⟩ from Eq. (23) is shown for arbitrarily chosen L n = 2 L_n=2 L n = 2 . A good qualitative agreement is demonstrated (without fine-tuning of L n L_n L n ), implying that, to a good approximation, only two parameters, E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 , are needed to characterize the response of a general GB, in agreement with the prediction of the GB model.
Figure 18. 19 representative directions [ a b c ] [abc] [ ab c ] , from which GB normals in either GB grain were selected for orthorhombic material (CaSO4 _4 4 ). They correspond to 190 different GB types [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] , considered in our numerical studies. The standard stereographic triangle is shown for reference.
Figure 19. Simulation results (circles) for ⟨ σ n n / Σ ⟩ \langle\sigma_{nn}/\Sigma\rangle ⟨ σ nn /Σ ⟩ in CaSO4 _4 4 under tensile loading Σ \Sigma Σ , evaluated on 190 [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GB types, constructed from the 19 selected directions shown in Figure 18 . For comparison, a smooth prediction (iso-lines) is shown for ⟨ σ n n ( 2 ) / Σ ⟩ \langle\sigma_{nn}^{(2)}/\Sigma\rangle ⟨ σ nn ( 2 ) /Σ ⟩ of the L n = 2 L_n=2 L n = 2 reduced model; cf. Eq. (23) . The inset shows the same results in 3D plot.
Discussion
The derived PDF(σ n n \sigma_{nn} σ nn ) distributions are not only very accurate, as demonstrated for various scenarios (see Figure 15 and Figure 16 ), but also relatively undemanding in computational sense. If they are produced numerically, using Monte Carlo sampling for GB-normal directions, the results can be immediately used for several practical applications. For instance, we could predict the GB-damage
initiation in complex geometries, using the probabilistic approach. If the GB strength σ c \sigma_c σ c of each GB type was known (or measured), and stress field Σ ( r ) \boldsymbol{\Sigma}(\boldsymbol{r}) Σ ( r ) in the investigated component at least roughly estimated (e.g., in FE simulations, using homogeneous material), one can immediately obtain the probability of finding an overloaded GB of a specific type at arbitrary location r \boldsymbol{r} r in the component, P ( r ) = ∫ σ c ∞ P D F ( σ n n ) d σ n n P(\boldsymbol{r}) = \int_{\sigma_c}^{\infty} \mathrm{PDF}(\sigma_{nn})\mathrm{d}\sigma_{nn} P ( r ) = ∫ σ c ∞ PDF ( σ nn ) d σ nn , using Σ ( r ) \boldsymbol{\Sigma}(\boldsymbol{r}) Σ ( r ) as an input for external loading to produce P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) . If that probability exceeded the threshold value, P ( r ) > P f P(\boldsymbol{r}) > P_f P ( r ) > P f , a macroscopic-size crack may develop at r \boldsymbol{r} r , which might result in a catastrophic failure of the component. With such approach, potentially dangerous regions, susceptible to intergranular cracking, can be quickly identified for any component and its loading. A more detailed analysis of such an application will be presented in a separate publication [34 ] . In all the examples presented so far, static elastic loads have been assumed in expressions for σ n n \sigma_{nn} σ nn and P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) . The procedure can be generalized also to dynamic stresses, provided that stress amplitudes remain in the elastic domain and inertia effects are negligible. In this respect, P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) spectra can be used to predict even the initiation of GB-fatigue cracks [35 ] . Following the above procedure for static load and assuming time-dependent evolution of GB strength (due to the build-up of strain localization [35 ] ), the probability P ( r , t ) = ∫ σ c ( t ) ∞ P D F ( σ n n ) d σ n n P(\boldsymbol{r},t) = \int_{\sigma_c(t)}^{\infty} \mathrm{PDF}(\sigma_{nn})\mathrm{d}\sigma_{nn} P ( r , t ) = ∫ σ c ( t ) ∞ PDF ( σ nn ) d σ nn becomes time dependent too. The measurement data can then be used, for example, to estimate how P f P_f P f and GB-strength evolution σ c ( t ) \sigma_c(t) σ c ( t ) change with the number of loading cycles. Although the semi-analytical σ n n \sigma_{nn} σ nn expression, derived for cubic crystal lattices, provides accurate P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) distributions for a wide range of situations, it relies not only on analytical, but also on empirical considerations (estimation of δ E 12 \delta E_{12} δ E 12 and s ( f n n ) s(f_{nn}) s ( f nn ) ). A (quasi) symmetry of δ E 12 ( E 12 ) \delta E_{12}(E_{12}) δ E 12 ( E 12 ) curves was observed for L n = 2 L_n = 2 L n = 2 and all investigated materials. The origin of this feature is not yet understood, it might even be only accidental. Nonetheless, it can be very useful, since it allows us to make the search of the fitting function significantly simpler (Figure 13 ). Due to that, it is important to gain a better understanding of this (quasi) symmetry for cubic lattices (and possibly even non-cubic lattices) in the future. In a similar way, the effect of more distant grains has not been modeled explicitly. Instead it was conveniently packed into an empirical fit of s ( f n n ) s(f_{nn}) s ( f nn ) , which represents the amplitude of GB-stress fluctuations (Figure 14 ). The fact that these fluctuations are more or less independent of stresses, makes the fitting function s ( f n n ) s(f_{nn}) s ( f nn ) relatively simple and, most importantly, the calculation of P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) very accurate (by applying Gaussian broadening with known width s ( f n n ) s(f_{nn}) s ( f nn ) ). However, the accuracy on a local GB scale is limited by the same s ( f n n ) s(f_{nn}) s ( f nn ) (representing the uncertainty of model predictions), and can be substantial in highly anisotropic materials. A possible improvement would necessarily include an exact modeling of the more distant grains (whose structure should probably be considered in a similar level of detail as the two GB grains). Unfortunately, this would probably result in very cumbersome and impractical solutions.
The derivation of GB models was based on two major ideas: the perturbative approach and the Saint Venant’s principle. A possible alternative approach could follow one of the well-known methods, used for calculation of the effective elastic constants of polycrystals from single-crystal and structure properties. For example, in the self-consistent method invented by Kröner [36 ] , an effective stress-strain relation is derived, taking into account the boundary conditions for stresses and strains at the GBs, which are only statistically correct. Analytical results are given for macroscopically isotropic polycrystals, composed of crystal grains with cubic symmetry [37 ] , and also for a general lattice symmetry [36 ] . Replicating such approach, the established relation between the local (single-grain) and macroscopic quantities would need to be modified to account for bi-crystal instead of single-crystal local quantities. While this might be worth trying, it has (at least) one significant shortcoming, common to all multi-scale techniques. It fails to reproduce additional degrees of freedom on a local scale (that would manifest themselves in stress fluctuations and thus in wider P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) distributions), given there are fewer degrees of freedom on a macroscopic scale. Therefore, additional improvements would be needed (as it was done here) to obtain accurate P D F ( σ n n ) \mathrm{PDF}(\sigma_{nn}) PDF ( σ nn ) . A detailed analysis along these lines is left for future work.
Conclusions
In this study, a perturbative model of grain-boundary-normal stresses has been derived for an arbitrary grain-boundary type within a general polycrystalline material, composed of randomly shaped elastic continuum grains with arbitrary lattice symmetry, and under a general uniform external loading. The constructed perturbative models have been solved under reasonable assumptions, needed to obtain compact, yet still accurate analytical and semi-analytical expressions for local grain-boundary-normal stresses and the corresponding statistical distributions. The strategy for deriving the models was based on two central concepts. Using the perturbation principle, the complexity of the model is gradually increased in each successive step, allowing us to first solve and understand simpler variants of the model. Following the Saint Venant’s principle, anisotropic elastic properties of the two grains closest to grain boundary have been considered in full, while the effect of more distant grains has been modeled in much smaller detail, using average quantities such as elastic grain anisotropy or bulk isotropic stiffness parameter.
The following conclusions have been reached from the solutions of derived perturbative models:
The general k k k -th order solution for the local grain-boundary-normal stress is of the following form: σ ~ n n ( k ) = A ( k ) Σ z z + B ( k ) ( Σ x x + Σ y y ) + f n n ( k ) \tilde{\sigma}_{nn}^{(k)} = A^{(k)}\Sigma_{zz} + B^{(k)}\left(\Sigma_{xx} + \Sigma_{yy}\right) + f_{nn}^{(k)} σ ~ nn ( k ) = A ( k ) Σ z z + B ( k ) ( Σ xx + Σ y y ) + f nn ( k ) , where A ( k ) A^{(k)} A ( k ) and B ( k ) B^{(k)} B ( k ) are the analytic functions of grain-boundary type and elastic material properties, Σ i i \Sigma_{ii} Σ ii is a diagonal component of the external loading tensor Σ \boldsymbol{\Sigma} Σ , expressed in a local grain-boundary system, and f n n ( k ) f_{nn}^{(k)} f nn ( k ) is a random variable, representing loading fluctuations.
To a good approximation (k = 2 k = 2 k = 2 ), the response on a chosen grain boundary can be characterized by just two parameters: E 12 E_{12} E 12 measures the average stiffness of grain-boundary neighborhood along the normal direction, while ν 12 \nu_{12} ν 12 is an effective Poisson’s ratio, measuring the average ratio of transverse and axial responses in both adjacent grains.
For an arbitrary lattice symmetry, A ( 2 ) A^{(2)} A ( 2 ) and B ( 2 ) B^{(2)} B ( 2 ) are simple functions, F ( E 12 , ν 12 , ⟨ E ⟩ , ⟨ ν ⟩ , L n ) F(E_{12},\nu_{12},\langle E\rangle,\langle\nu\rangle,L_n) F ( E 12 , ν 12 , ⟨ E ⟩ , ⟨ ν ⟩ , L n ) , where ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ and ⟨ ν ⟩ \langle\nu\rangle ⟨ ν ⟩ denote average elastic bulk properties, and L n ≥ 0 L_n \ge0 L n ≥ 0 is a modeling parameter accounting for the amount of buffer grains. Also higher order solutions (k > 2 k > 2 k > 2 ) have been obtained, but with resulting expressions too cumbersome to be useful in practice.
To account for 3D effects and realistic boundary conditions, a model upgrade has been proposed by assuming E 12 → E 12 + δ E 12 E_{12} \rightarrow E_{12} + \delta E_{12} E 12 → E 12 + δ E 12 and ν 12 → ν 12 + δ ν 12 \nu_{12} \rightarrow\nu_{12} + \delta\nu_{12} ν 12 → ν 12 + δ ν 12 in the expressions for A ( 2 ) A^{(2)} A ( 2 ) and B ( 2 ) B^{(2)} B ( 2 ) , with δ E 12 \delta E_{12} δ E 12 and δ ν 12 \delta\nu_{12} δ ν 12 obtained from fitting the results of numerical simulations. A simple empirical relation for δ E 12 \delta E_{12} δ E 12 (and δ ν 12 \delta\nu_{12} δ ν 12 ) has been derived for materials with cubic crystal lattices.
To account for realistic stresses acting on a grain-boundary model, the external loading has been dressed by fluctuations, Σ → Σ + f \boldsymbol{\Sigma} \rightarrow\boldsymbol{\Sigma} + \mathbf{f} Σ → Σ + f . To a good approximation, the resulting fluctuations of grain-boundary-normal stresses (f n n f_{nn} f nn ), have been found to be independent of stresses. Their distribution is Gaussian, with standard deviation of the form s ( f n n ) ≈ Σ m i s F ( A u ) s(f_{nn}) \approx\Sigma_{\mathrm{mis}}F(A^u) s ( f nn ) ≈ Σ mis F ( A u ) , where F ( A u ) F(A^u) F ( A u ) is an empirical function, that increases with the value of universal elastic anisotropy index A u A^u A u .
A comparison with finite element simulations has demonstrated that the derived semi-analytical expression for a local σ ~ n n ( 2 ) \tilde{\sigma}_{nn}^{(2)} σ ~ nn ( 2 ) is accurate only up to unknown stress fluctuations, i.e. , the uncertainty of model prediction is s ( f n n ) s(f_{nn}) s ( f nn ) . However, the corresponding statistical distributions, PDF ( σ ~ n n ( 2 ) ) \operatorname{PDF}(\tilde{\sigma}_{nn}^{(2)}) PDF ( σ ~ nn ( 2 ) ) , have been shown to be very accurate. Indeed, an excellent agreement with the simulation results has been found for arbitrary grain-boundary types in a general elastic untextured polycrystalline material[27 ] under arbitrary uniform loading.
Acknowledgments
We gratefully acknowledge financial support provided by Slovenian Research Agency (grant P2-0026). We also thank Jérémy Hure for useful discussions and comments that helped us to improve the manuscript.
Finite element aggregate model
Polycrystalline aggregate models are generated upon Voronoi tessellations [38 ] with periodic microstructures in all three spatial directions[39 ] . Finite element meshes are generated with quadratic tetrahedral elements to preserve the geometry of the grains. An example of the model with 4000 grains, used throughout this study, is shown in Figure 1 (a). A general uniform loading Σ l a b \boldsymbol{\Sigma}^{\mathrm{lab}} Σ lab is applied to the aggregate, where Σ i j = ⟨ σ i j ⟩ \Sigma_{ij} = \langle\sigma_{ij}\rangle Σ ij = ⟨ σ ij ⟩ for averages taken over the entire volume of the aggregate model. Since grains are assumed ideally elastic, a unit loading can be applied, using a small strain approximation.
It is important to note that the analysis of each GB type requires a dedicated aggregate model. To isolate the effect of selected GB type, the same grain topology and finite element mesh (see Figure 1 (a)) are used in all of them.
Due to topological constraints, the same GB character cannot be assigned to all GBs. In practice, a chosen [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω GB type can be imposed on at most ∼ 17 % \sim17\% ∼ 17% of GBs[29 ] in a given aggregate, with remaining GBs belonging to a random type (i.e. , the two grains adjacent to the GB are assigned random orientations).
Also, it has been verified that the aggregate size (4000 grains) and finite element mesh density (∼ 5 \sim5 ∼ 5 million total elements) are sufficiently large to produce negligible finite size effects.
The constitutive equations of the generalized Hooke’s law are solved with finite element solver Abaqus [32 ] in a small strain approximation. Numerically calculated stress fields σ i \boldsymbol{\sigma}_i σ i are then used to obtain a single σ n n ( k ) \sigma_{nn}(k) σ nn ( k ) value for each GB k k k of a given type [ a b c ] − [ d e f ] − Δ ω [abc]-[def]-\Delta\omega [ ab c ] − [ d e f ] − Δ ω . In short, σ n n ( k ) \sigma_{nn}(k) σ nn ( k ) is calculated by projecting stress σ i \boldsymbol{\sigma}_i σ i onto a GB normal n ( k ) \mathbf{n}^{(k)} n ( k ) and averaging over all finite elements i i i touching the GB k k k ; σ n n ( k ) = ∑ i A i ( k ) σ i ⋅ n ( k ) / ∑ i A i ( k ) \sigma_{nn}(k)=\sum_i A_i^{(k)}\boldsymbol{\sigma}_i\cdot\mathbf{n}^{(k)}/\sum_i A_i^{(k)} σ nn ( k ) = ∑ i A i ( k ) σ i ⋅ n ( k ) / ∑ i A i ( k ) , where A i ( k ) A_i^{(k)} A i ( k ) is the area of GB-element facet touching the GB, and σ i = 1 / 3 ∑ j = 1 3 σ i , j \boldsymbol{\sigma}_i=1/3\sum_{j=1}^{3}\boldsymbol{\sigma}_{i,j} σ i = 1/3 ∑ j = 1 3 σ i , j is Cauchy stress averaged over three Gauss points j j j of element i i i , that are located closest to the GB plane. Besides local stresses, the first two statistical moments, the mean value and standard deviation of PDF( σ n n ) (\sigma_{nn}) ( σ nn ) , are calculated as ⟨ σ n n ⟩ = ∑ k A k σ n n ( k ) / ∑ k A k \langle\sigma_{nn}\rangle=\sum_k A_k\sigma_{nn}(k)/\sum_k A_k ⟨ σ nn ⟩ = ∑ k A k σ nn ( k ) / ∑ k A k and s ( σ n n ) = ⟨ σ n n 2 ⟩ − ⟨ σ n n ⟩ 2 s(\sigma_{nn})=\sqrt{\langle\sigma_{nn}^{2}\rangle-\langle\sigma_{nn}\rangle^{2}} s ( σ nn ) = ⟨ σ nn 2 ⟩ − ⟨ σ nn ⟩ 2 , respectively, for A k A_k A k denoting the area of GB k k k . The summation k k k is performed over all special GBs of a given type.33
B Analytic expressions for grains with cubic lattice symmetry
In this section, analytic expressions for cubic lattice symmetry are given for completeness. It is assumed that rotation R c r y \mathbf{R}^{\mathrm{cry}} R cry , defined in Eq. (1) , is used for expressing the crystallographic properties of the grain in a local GB coordinate system, whose GB-plane normal (local z z z -axis) is oriented along the [ h k l ] [hkl] [ hk l ] direction of the grain (crystallographic) coordinate system and with ω \omega ω denoting the twist angle about the GB normal.
Strain component in the direction of GB normal can be expressed (for a cubic grain) as
ε z z = ∑ k , l = 1 3 S 33 k l G B σ k l = ∑ k , l , m , n , o , p = 1 3 R 3 m c r y R 3 n c r y R k o c r y R l p c r y s m n o p c r y σ k l = σ x x ( S 0 ( h k ( 2 h k ( h 2 + k 2 + l 2 ) sin 2 ω + l ( k 2 − h 2 ) h 2 + k 2 + l 2 sin 2 ω ) + 2 l 2 ( h 2 k 2 + h 4 + k 4 ) cos 2 ω ) ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 + S 12 ) + σ y y ( S 0 ( l ( h k ( h − k ) ( h + k ) h 2 + k 2 + l 2 sin 2 ω + 2 l ( h 2 k 2 + h 4 + k 4 ) sin 2 ω ) + 2 h k 2 l 2 ( h 2 + k 2 + l 2 ) cos 2 ω ) ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 + S 12 ) + σ z z ( S 0 ( h 4 + k 4 + l 4 ) ( h 2 + k 2 + l 2 ) 2 + S 12 + S 44 2 ) + σ x y 2 S 0 ( ( h 2 k 2 ( h 2 + k 2 ) − l 2 ( h 4 + k 4 ) ) sin 2 ω + h k l ( k 2 − h 2 ) h 2 + k 2 + l 2 cos 2 ω ) ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 + σ x z 2 S 0 ( h k ( k 2 − h 2 ) h 2 + k 2 + l 2 sin ω + l ( − l 2 ( h 2 + k 2 ) + h 4 + k 4 ) cos ω ) h 2 + k 2 ( h 2 + k 2 + l 2 ) 2 + σ y z 2 S 0 ( l ( l 2 ( h 2 + k 2 ) − h 4 − k 4 ) sin ω + h k ( k 2 − h 2 ) h 2 + k 2 + l 2 cos ω ) h 2 + k 2 ( h 2 + k 2 + l 2 ) 2 , (42) \begin{aligned}
\varepsilon_{zz}
&=\sum_{k,l=1}^{3}S^{\mathrm{GB}}_{33kl}\sigma_{kl}
=\sum_{k,l,m,n,o,p=1}^{3}R^{\mathrm{cry}}_{3m}R^{\mathrm{cry}}_{3n}R^{\mathrm{cry}}_{ko}R^{\mathrm{cry}}_{lp}s^{\mathrm{cry}}_{mnop}\sigma_{kl}\\
&=\sigma_{xx}\left(\frac{S_0\left(hk\left(2hk\left(h^2+k^2+l^2\right)\sin^2\omega+l\left(k^2-h^2\right)\sqrt{h^2+k^2+l^2}\sin2\omega\right)+2l^2\left(h^2k^2+h^4+k^4\right)\cos^2\omega\right)}{\left(h^2+k^2\right)\left(h^2+k^2+l^2\right)^2}+S_{12}\right)\\
&\quad+\sigma_{yy}\left(\frac{S_0\left(l\left(hk(h-k)(h+k)\sqrt{h^2+k^2+l^2}\sin2\omega+2l\left(h^2k^2+h^4+k^4\right)\sin^2\omega\right)+2hk^2l^2\left(h^2+k^2+l^2\right)\cos^2\omega\right)}{\left(h^2+k^2\right)\left(h^2+k^2+l^2\right)^2}+S_{12}\right)\\
&\quad+\sigma_{zz}\left(\frac{S_0\left(h^4+k^4+l^4\right)}{\left(h^2+k^2+l^2\right)^2}+S_{12}+\frac{S_{44}}{2}\right)\\
&\quad+\sigma_{xy}\frac{2S_0\left(\left(h^2k^2\left(h^2+k^2\right)-l^2\left(h^4+k^4\right)\right)\sin2\omega+hkl\left(k^2-h^2\right)\sqrt{h^2+k^2+l^2}\cos2\omega\right)}{\left(h^2+k^2\right)\left(h^2+k^2+l^2\right)^2}\\
&\quad+\sigma_{xz}\frac{2S_0\left(hk\left(k^2-h^2\right)\sqrt{h^2+k^2+l^2}\sin\omega+l\left(-l^2\left(h^2+k^2\right)+h^4+k^4\right)\cos\omega\right)}{\sqrt{h^2+k^2}\left(h^2+k^2+l^2\right)^2}\\
&\quad+\sigma_{yz}\frac{2S_0\left(l\left(l^2\left(h^2+k^2\right)-h^4-k^4\right)\sin\omega+hk\left(k^2-h^2\right)\sqrt{h^2+k^2+l^2}\cos\omega\right)}{\sqrt{h^2+k^2}\left(h^2+k^2+l^2\right)^2}, \tag*{(42)}
\end{aligned} ε z z = k , l = 1 ∑ 3 S 33 k l GB σ k l = k , l , m , n , o , p = 1 ∑ 3 R 3 m cry R 3 n cry R k o cry R l p cry s mn o p cry σ k l = σ xx ( ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 S 0 ( hk ( 2 hk ( h 2 + k 2 + l 2 ) sin 2 ω + l ( k 2 − h 2 ) h 2 + k 2 + l 2 sin 2 ω ) + 2 l 2 ( h 2 k 2 + h 4 + k 4 ) cos 2 ω ) + S 12 ) + σ y y ( ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 S 0 ( l ( hk ( h − k ) ( h + k ) h 2 + k 2 + l 2 sin 2 ω + 2 l ( h 2 k 2 + h 4 + k 4 ) sin 2 ω ) + 2 h k 2 l 2 ( h 2 + k 2 + l 2 ) cos 2 ω ) + S 12 ) + σ z z ( ( h 2 + k 2 + l 2 ) 2 S 0 ( h 4 + k 4 + l 4 ) + S 12 + 2 S 44 ) + σ x y ( h 2 + k 2 ) ( h 2 + k 2 + l 2 ) 2 2 S 0 ( ( h 2 k 2 ( h 2 + k 2 ) − l 2 ( h 4 + k 4 ) ) sin 2 ω + hk l ( k 2 − h 2 ) h 2 + k 2 + l 2 cos 2 ω ) + σ x z h 2 + k 2 ( h 2 + k 2 + l 2 ) 2 2 S 0 ( hk ( k 2 − h 2 ) h 2 + k 2 + l 2 sin ω + l ( − l 2 ( h 2 + k 2 ) + h 4 + k 4 ) cos ω ) + σ y z h 2 + k 2 ( h 2 + k 2 + l 2 ) 2 2 S 0 ( l ( l 2 ( h 2 + k 2 ) − h 4 − k 4 ) sin ω + hk ( k 2 − h 2 ) h 2 + k 2 + l 2 cos ω ) , (42)
where S 11 : = s 1111 c r y S_{11}:=s^{\mathrm{cry}}_{1111} S 11 := s 1111 cry , S 12 : = s 1122 c r y S_{12}:=s^{\mathrm{cry}}_{1122} S 12 := s 1122 cry and S 44 : = 4 s 2323 c r y S_{44}:=4s^{\mathrm{cry}}_{2323} S 44 := 4 s 2323 cry are the components of compliance tensor of a grain in Voigt notation and S 0 : = S 11 − S 12 − S 44 / 2 S_0:=S_{11}-S_{12}-S_{44}/2 S 0 := S 11 − S 12 − S 44 /2 . Expressions for the other two diagonal components of strain tensor, ε x x \varepsilon_{xx} ε xx and ε y y \varepsilon_{yy} ε y y , are derived analogously, but are omitted here for brevity.
The effective GB-stiffness parameter E 12 E_{12} E 12 , measuring the average stiffness of GB neighborhood along the GB-normal direction, takes the following form (for cubic grains)
E 12 = 2 ⟨ E ⟩ − 1 s 3333 G B , a b c + s 3333 G B , d e f = ⟨ E ⟩ − 1 S 11 − S 0 ( ( a b ) 2 + ( a c ) 2 + ( b c ) 2 ( a 2 + b 2 + c 2 ) 2 + ( d e ) 2 + ( d f ) 2 + ( e f ) 2 ( d 2 + e 2 + f 2 ) 2 ) . (43) \begin{aligned}
E_{12}=\frac{2\langle E\rangle^{-1}}{s^{\mathrm{GB},abc}_{3333}+s^{\mathrm{GB},def}_{3333}} \\
&=\frac{\langle E\rangle^{-1}}{S_{11}-S_0\left(\frac{(ab)^2+(ac)^2+(bc)^2}{(a^2+b^2+c^2)^2}+\frac{(de)^2+(df)^2+(ef)^2}{(d^2+e^2+f^2)^2}\right)}. \tag*{(43)}
\end{aligned} E 12 = s 3333 GB , ab c + s 3333 GB , d e f 2 ⟨ E ⟩ − 1 = S 11 − S 0 ( ( a 2 + b 2 + c 2 ) 2 ( ab ) 2 + ( a c ) 2 + ( b c ) 2 + ( d 2 + e 2 + f 2 ) 2 ( d e ) 2 + ( df ) 2 + ( e f ) 2 ) ⟨ E ⟩ − 1 . (43)
The effective GB Poisson’s ratio ν 12 \nu_{12} ν 12 , measuring the average ratio of transverse and axial responses (strains) in both GB grains, takes the following form (for cubic grains)
ν 12 = − ⟨ E ⟩ 4 ( s 3311 G B , a b c + s 3322 G B , a b c + s 3311 G B , d e f + s 3322 G B , d e f ) = ⟨ ν ⟩ + 1 2 ( E 12 − 1 − 1 ) . (44) \begin{aligned}
\nu_{12}=-\frac{\langle E\rangle}{4}\left(s^{\mathrm{GB},abc}_{3311}+s^{\mathrm{GB},abc}_{3322}+s^{\mathrm{GB},def}_{3311}+s^{\mathrm{GB},def}_{3322}\right) \\
&=\langle\nu\rangle+\frac{1}{2}\left(E_{12}^{-1}-1\right). \tag*{(44)}
\end{aligned} ν 12 = − 4 ⟨ E ⟩ ( s 3311 GB , ab c + s 3322 GB , ab c + s 3311 GB , d e f + s 3322 GB , d e f ) = ⟨ ν ⟩ + 2 1 ( E 12 − 1 − 1 ) . (44)
where ⟨ ν ⟩ = 1 2 ( 1 − ⟨ E ⟩ ( S 11 + 2 S 12 ) ) \langle\nu\rangle= \frac{1}{2}\left(1-\langle E\rangle(S_{11}+2S_{12})\right) ⟨ ν ⟩ = 2 1 ( 1 − ⟨ E ⟩ ( S 11 + 2 S 12 ) ) .
General grain boundary model solution for cubic crystal lattices
The highest-order (reduced) solution σ n n ( 3 ) \sigma_{nn}^{(3)} σ nn ( 3 ) , derived in Sec. III B 6 for the most general grain-lattice symmetry and external loading, simplifies enormously for cubic lattice symmetry, where
s 1111 c r y = s 2222 c r y = s 3333 c r y = : S 11 , s 1122 c r y = s 1133 c r y = s 2233 c r y = : S 12 (45) s_{1111}^{\mathrm{cry}}=s_{2222}^{\mathrm{cry}}=s_{3333}^{\mathrm{cry}}=:S_{11},
\qquad
s_{1122}^{\mathrm{cry}}=s_{1133}^{\mathrm{cry}}=s_{2233}^{\mathrm{cry}}=:S_{12} \tag*{(45)} s 1111 cry = s 2222 cry = s 3333 cry =: S 11 , s 1122 cry = s 1133 cry = s 2233 cry =: S 12 (45)
⟨ ν ⟩ = 1 2 ( 1 − S 11 + 2 S 12 ⟨ E ⟩ ) (46) \langle\nu\rangle=\frac{1}{2}\left(1-\frac{S_{11}+2S_{12}}{\langle E\rangle}\right) \tag*{(46)} ⟨ ν ⟩ = 2 1 ( 1 − ⟨ E ⟩ S 11 + 2 S 12 ) (46)
which in turn implies
s t l h k l = − s l l h k l + ( S 11 + 2 S 12 ) , s t t h k l = 1 2 ( s l l h k l + ( S 11 + 2 S 12 ) ) (47) s_{tl}^{hkl}=-s_{ll}^{hkl}+(S_{11}+2S_{12}),
\qquad
s_{tt}^{hkl}=\frac{1}{2}\left(s_{ll}^{hkl}+(S_{11}+2S_{12})\right) \tag*{(47)} s tl hk l = − s l l hk l + ( S 11 + 2 S 12 ) , s tt hk l = 2 1 ( s l l hk l + ( S 11 + 2 S 12 ) ) (47)
meaning A c u b ( 3 ) A_{\mathrm{cub}}^{(3)} A cub ( 3 ) and B c u b ( 3 ) B_{\mathrm{cub}}^{(3)} B cub ( 3 ) are only functions of Young‘s moduli E a b c E_{abc} E ab c and E d e f E_{def} E d e f along the GB-normal direction in both grains,
A c u b ( 3 ) = ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) E 12 + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ ) Δ 12 ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ + 4 ( L t + s ′ ) ) Δ 12 (48) A_{\mathrm{cub}}^{(3)}=
\frac{(1+(2L_t+s')E_{12})(L_n+2s'+4(L_t+s')E_{12}+(2L_nL_t+L_ns'-2s'^{2})E_{12})+(L_n+2s')\Delta_{12}}
{(1+(2L_t+s')E_{12})(L_n+2s'+4(L_t+s')+(2L_nL_t+L_ns'-2s'^{2})E_{12})+(L_n+2s'+4(L_t+s'))\Delta_{12}} \tag*{(48)} A cub ( 3 ) = ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ + 4 ( L t + s ′ )) Δ 12 ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) E 12 + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ ) Δ 12 (48)
B c u b ( 3 ) = 1 2 ( 1 − A c u b ( 3 ) ) . B_{\mathrm{cub}}^{(3)}=\frac{1}{2}\left(1-A_{\mathrm{cub}}^{(3)}\right). B cub ( 3 ) = 2 1 ( 1 − A cub ( 3 ) ) .
All the used quantities are dimensionless, with s ′ s' s ′ denoting
s ′ : = ⟨ E ⟩ ( S 11 + 2 S 12 ) = 1 − 2 ⟨ ν ⟩ . (49) s':=\langle E\rangle(S_{11}+2S_{12})=1-2\langle\nu\rangle. \tag*{(49)} s ′ := ⟨ E ⟩ ( S 11 + 2 S 12 ) = 1 − 2 ⟨ ν ⟩ . (49)
E a b c E_{abc} E ab c and E d e f E_{def} E d e f appear only in combinations E 12 E_{12} E 12 and Δ 12 \Delta_{12} Δ 12 , where
Δ 12 : = 4 E a b c − 1 E d e f − 1 ( E a b c − 1 + E d e f − 1 ) 2 − 1 = ( E 12 ⟨ E ⟩ ) 2 E a b c − 1 E d e f − 1 − 1 (50) \begin{aligned}
\Delta_{12}:=\frac{4E_{abc}^{-1}E_{def}^{-1}}{(E_{abc}^{-1}+E_{def}^{-1})^{2}}-1 \\
&=(E_{12}\langle E\rangle)^{2}E_{abc}^{-1}E_{def}^{-1}-1 \tag*{(50)}
\end{aligned} Δ 12 := ( E ab c − 1 + E d e f − 1 ) 2 4 E ab c − 1 E d e f − 1 − 1 = ( E 12 ⟨ E ⟩ ) 2 E ab c − 1 E d e f − 1 − 1 (50)
is a sort of geometric mean of inverse Young’s moduli of both grains, measuring the deviation from a single-grain scenario (i.e., the [ a b c ] − [ a b c ] [abc]-[abc] [ ab c ] − [ ab c ] GB type), in which Δ 12 \Delta_{12} Δ 12 vanishes. If the effective GB stiffness E 12 E_{12} E 12 is a measure of the average stiffness of the [ a b c ] − [ d e f ] [abc]-[def] [ ab c ] − [ d e f ] GB neighborhood along the GB-normal direction, then Δ 12 \Delta_{12} Δ 12 represents an additional, orthogonal degree of freedom, that can break the σ n n \sigma_{nn} σ nn - and s ( σ n n ) s(\sigma_{nn}) s ( σ nn ) -degeneracies of GB types with the same E 12 E_{12} E 12 (but different values of Δ 12 \Delta_{12} Δ 12 ).
The GB-normal stress and the corresponding first two statistical moments thus become
σ n n ( 3 ) = A c u b ( 3 ) Σ z z + B c u b ( 3 ) ( Σ x x + Σ y y ) , ⟨ σ n n ( 3 ) ⟩ = 1 3 tr ( Σ l a b ) , \sigma_{nn}^{(3)}=A_{\mathrm{cub}}^{(3)}\Sigma_{zz}+B_{\mathrm{cub}}^{(3)}(\Sigma_{xx}+\Sigma_{yy}),
\qquad
\left\langle\sigma_{nn}^{(3)}\right\rangle=\frac{1}{3}\operatorname{tr}(\Sigma^{\mathrm{lab}}), σ nn ( 3 ) = A cub ( 3 ) Σ z z + B cub ( 3 ) ( Σ xx + Σ y y ) , ⟨ σ nn ( 3 ) ⟩ = 3 1 tr ( Σ lab ) ,
s ( σ n n ( 3 ) ) = 2 Σ m i s l a b 3 5 ∣ ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n − 2 L t + 6 ( L t + s ′ ) E 12 + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n − 2 L t ) Δ 12 ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ + 4 ( L t + s ′ ) ) Δ 12 ∣ (51) s(\sigma_{nn}^{(3)})=\frac{2\Sigma_{\mathrm{mis}}^{\mathrm{lab}}}{3\sqrt{5}}
\left|
\frac{(1+(2L_t+s')E_{12})(L_n-2L_t+6(L_t+s')E_{12}+(2L_nL_t+L_ns'-2s'^{2})E_{12})+(L_n-2L_t)\Delta_{12}}
{(1+(2L_t+s')E_{12})(L_n+2s'+4(L_t+s')+(2L_nL_t+L_ns'-2s'^{2})E_{12})+(L_n+2s'+4(L_t+s'))\Delta_{12}}
\right| \tag*{(51)} s ( σ nn ( 3 ) ) = 3 5 2 Σ mis lab ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n + 2 s ′ + 4 ( L t + s ′ ) + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n + 2 s ′ + 4 ( L t + s ′ )) Δ 12 ( 1 + ( 2 L t + s ′ ) E 12 ) ( L n − 2 L t + 6 ( L t + s ′ ) E 12 + ( 2 L n L t + L n s ′ − 2 s ′ 2 ) E 12 ) + ( L n − 2 L t ) Δ 12 (51)
The relevance of E 12 E_{12} E 12 parameter determining the s ( σ n n ) s(\sigma_{nn}) s ( σ nn ) for materials with cubic lattice symmetry was correctly identified already in [28 ] . The new parameter Δ 12 \Delta_{12} Δ 12 in Eq. (51) represents a higher-order correction, which can, at least qualitatively, explain the (small) spread of s ( σ n n ) s(\sigma_{nn}) s ( σ nn ) values, observed numerically for GB types with the same E 12 E_{12} E 12 (see Fig. 13(a) in [28 ] ).
Material elastic properties
Elastic constants of single crystals with cubic symmetry, together with their aggregate properties, are listed in Table III for several representative materials. The materials are ordered according to their universal elastic anisotropy index A u A^{u} A u [33 ] , where A u = 0 A^{u}=0 A u = 0 corresponds to an isotropic crystal.
In Table IV the elastic constants of a single crystal with orthorhombic symmetry (CaSO4 _4 4 ) and its aggregate properties are gathered.
Table IV. Elastic constants of a single crystal with orthorhombic symmetry (C a S O 4 \mathrm{CaSO}_{4} CaSO 4 ) [41 ] and its aggregate properties. C i j C_{ij} C ij and ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ are in units of GPa.
Model of stress fluctuations
The stress applied to the GB model (see, e.g. , Figure 5 ) is assumed to be equal to the external stress, modified by fluctuations, Σ + f ( r ) \mathbf{\Sigma}+\mathbf{f}(r) Σ + f ( r ) , where f ( r ) \mathbf{f}(r) f ( r ) is the fluctuation stress tensor at position r r r . In a large aggregate, where crystallographic orientations of the grains are uncorrelated with their shapes (assuming zero morphological texture), the average fluctuation stress tensor should go towards zero, i.e. ,
⟨ f i j ( r ) ⟩ r = ⟨ ( R f ( r ) R T ) i j ⟩ r = 0 , (52) \left\langle f_{ij}(r)\right\rangle_r=\left\langle\left(\mathbf{R}\mathbf{f}(r)\mathbf{R}^{T}\right)_{ij}\right\rangle_r=0, \tag*{(52)} ⟨ f ij ( r ) ⟩ r = ⟨ ( Rf ( r ) R T ) ij ⟩ r = 0 , (52)
when averaged over all GBs of a chosen type (and thus having a specific value of E 12 E_{12} E 12 , ν 12 \nu_{12} ν 12 , ...) and for a fixed GB-normal direction n n n (see Figure 20 ). This should be true in any coordinate system (R \mathbf{R} R denotes an arbitrary rotation matrix).
Figure 20. The fluctuation stress tensor f ( r ) \mathbf{f}(r) f ( r ) at position r r r is defined as the difference between the actual loading stress Σ ( r ) \mathbf{\Sigma}(r) Σ ( r ) , acting on a given GB neighborhood at position r r r , and the external loading stress Σ \mathbf{\Sigma} Σ , i.e. , f ( r ) = Σ ( r ) − Σ \mathbf{f}(r)=\mathbf{\Sigma}(r)-\mathbf{\Sigma} f ( r ) = Σ ( r ) − Σ .
Since fluctuations are induced by external loading Σ \mathbf{\Sigma} Σ and strain incompatibility of the grains (which correlates with the elastic anisotropy index A u A^{u} A u ), it seems reasonable to assume that the corresponding standard deviation s s s depends on Σ \mathbf{\Sigma} Σ and A u A^{u} A u , possibly on GB model intrinsic parameters (e.g. , E 12 E_{12} E 12 , ν 12 \nu_{12} ν 12 , ...), but not on the global aggregate (or external loading) rotation R \mathbf{R} R ,
s ( f i j ( r ) ) = s ( ( R f ( r ) R T ) i j ) = F ( Σ , A u , E 12 , … ) . (53) s(f_{ij}(r))=s\left((\mathbf{R}\mathbf{f}(r)\mathbf{R}^{T})_{ij}\right)=\mathcal{F}(\mathbf{\Sigma},A^{u},E_{12},\ldots). \tag*{(53)} s ( f ij ( r )) = s ( ( Rf ( r ) R T ) ij ) = F ( Σ , A u , E 12 , … ) . (53)
caption: Elastic constants C i j C_{ij} C ij (in Voigt notation) of single crystals with cubic symmetry [40 ] and their aggregate properties. C i j C_{ij} C ij and ⟨ E ⟩ \langle E\rangle ⟨ E ⟩ are in units of GPa. Fe is assumed in gamma phase.
In addition, the rotational invariance of s s s suggests that Σ \mathbf{\Sigma} Σ dependence in Eq. (53) can be expressed solely in terms of Σ \mathbf{\Sigma} Σ invariants
F ( Σ ) → F ( tr ( Σ ) , det ( Σ ) , tr ( Σ d e v 2 ) ) . (54) \mathcal{F}(\mathbf{\Sigma})\to\mathcal{F}\left(\operatorname{tr}(\mathbf{\Sigma}),\det(\mathbf{\Sigma}),\operatorname{tr}(\mathbf{\Sigma}_{\mathrm{dev}}^{2})\right). \tag*{(54)} F ( Σ ) → F ( tr ( Σ ) , det ( Σ ) , tr ( Σ dev 2 ) ) . (54)
In the limit where the external loading is of hydrostatic form, Σ = Σ h y d \mathbf{\Sigma}=\mathbf{\Sigma}_{\mathrm{hyd}} Σ = Σ hyd (and Σ d e v = 0 \mathbf{\Sigma}_{\mathrm{dev}}=0 Σ dev = 0 ), a trivial solution is obtained with no stress fluctuations, f = 0 \mathbf{f}=0 f = 0 , implying that Σ d e v \mathbf{\Sigma}_{\mathrm{dev}} Σ dev is the only relevant loading contribution in Eq. (54) . To account for this limit, the following fluctuation stress tensor is finally proposed at position r r r ,
f ( r ) = η ( r ) R ( r ) Σ d e v R ( r ) T , (55) \mathbf{f}(r)=\eta(r)\mathbf{R}(r)\mathbf{\Sigma}_{\mathrm{dev}}\mathbf{R}(r)^{T}, \tag*{(55)} f ( r ) = η ( r ) R ( r ) Σ dev R ( r ) T , (55)
where
R ( r ) = R ( α 1 ( r ) , α 2 ( r ) , α 3 ( r ) ) (56) \mathbf{R}(r)=\mathbf{R}\left(\alpha_{1}(r),\alpha_{2}(r),\alpha_{3}(r)\right) \tag*{(56)} R ( r ) = R ( α 1 ( r ) , α 2 ( r ) , α 3 ( r ) ) (56)
is a random rotation matrix with corresponding Euler angles α j ( r ) \alpha_j(r) α j ( r ) , and η ( r ) \eta(r) η ( r ) is a random number with the assumed Gaussian distribution N ( 0 , s 2 ( η ) ) \mathcal{N}(0,s^{2}(\eta)) N ( 0 , s 2 ( η )) with s ( η ) = F ( A u , E 12 , … ) s(\eta)=\mathcal{F}(A^{u},E_{12},\ldots) s ( η ) = F ( A u , E 12 , … ) .34
In Eq. (55) , the fluctuation stress tensor is modeled as the deviatoric part of the external loading35 rescaled and rotated by a random amount to account for uncertainty of far-away grains that blur the external loading.
Using f ( r ) \mathbf{f}(r) f ( r ) defined in Eq. (55) and a general expression σ n n ( k ) = A ( k ) Σ z z + B ( k ) ( Σ x x + Σ y y ) \sigma_{nn}^{(k)}=A^{(k)}\Sigma_{zz}+B^{(k)}(\Sigma_{xx}+\Sigma_{yy}) σ nn ( k ) = A ( k ) Σ z z + B ( k ) ( Σ xx + Σ y y ) , the GB-normal-stress fluctuations evaluate to (using notation f n n ( k ) = Δ σ n n ( k ) f_{nn}^{(k)}=\Delta\sigma_{nn}^{(k)} f nn ( k ) = Δ σ nn ( k ) )
f n n ( k ) ( r ) = A ( k ) f z z ( r ) + B ( k ) ( f x x ( r ) + f y y ( r ) ) = ( A ( k ) − B ( k ) ) f z z ( r ) , (57) \begin{aligned}
f_{nn}^{(k)}(r)=A^{(k)}f_{zz}(r)+B^{(k)}\left(f_{xx}(r)+f_{yy}(r)\right) \\
&=\left(A^{(k)}-B^{(k)}\right)f_{zz}(r), \tag*{(57)}
\end{aligned} f nn ( k ) ( r ) = A ( k ) f z z ( r ) + B ( k ) ( f xx ( r ) + f y y ( r ) ) = ( A ( k ) − B ( k ) ) f z z ( r ) , (57)
where the tr ( f ( r ) ) = 0 \operatorname{tr}(\mathbf{f}(r))=0 tr ( f ( r )) = 0 property of Eq. (55) has been used. The first two statistical moments follow as
⟨ f n n ( k ) ( r ) ⟩ r = ( A ( k ) − B ( k ) ) ∫ η PDF ( η ) d η ∭ Σ d e v , z z PDF ( α 1 , α 2 , α 3 ) d α 1 d α 2 d α 3 = 0 , (58) \begin{aligned}
\left\langle f_{nn}^{(k)}(\mathbf{r}) \right\rangle_{\mathbf{r}} = \left(A^{(k)}-B^{(k)}\right)\int\eta\,\operatorname{PDF}(\eta)\,d\eta\iiint\Sigma_{\mathrm{dev},zz}\operatorname{PDF}(\alpha_1,\alpha_2,\alpha_3)\,d\alpha_1d\alpha_2d\alpha_3 \\
&=0, \tag*{(58)}
\end{aligned} ⟨ f nn ( k ) ( r ) ⟩ r = ( A ( k ) − B ( k ) ) ∫ η PDF ( η ) d η ∭ Σ dev , z z PDF ( α 1 , α 2 , α 3 ) d α 1 d α 2 d α 3 = 0 , (58)
and
s ( f n n ( k ) ( r ) ) = ⟨ ( f n n ( k ) ( r ) ) 2 ⟩ r = ∣ A ( k ) − B ( k ) ∣ ∫ η 2 PDF ( η ) d η ∭ Σ d e v , z z 2 PDF ( α 1 , α 2 , α 3 ) d α 1 d α 2 d α 3 = 2 Σ m i s 3 5 ∣ A ( k ) − B ( k ) ∣ s ( η ) , (59) \begin{aligned}
s\left(f_{nn}^{(k)}(\mathbf{r})\right)=\sqrt{\left\langle\left(f_{nn}^{(k)}(\mathbf{r})\right)^2\right\rangle_{\mathbf{r}}} \\
&=\left|A^{(k)}-B^{(k)}\right|\sqrt{\int\eta^2\operatorname{PDF}(\eta)\,d\eta\iiint\Sigma_{\mathrm{dev},zz}^{2}\operatorname{PDF}(\alpha_1,\alpha_2,\alpha_3)\,d\alpha_1d\alpha_2d\alpha_3} \\
&=\frac{2\Sigma_{\mathrm{mis}}}{3\sqrt{5}}\left|A^{(k)}-B^{(k)}\right|s(\eta), \tag*{(59)}
\end{aligned} s ( f nn ( k ) ( r ) ) = ⟨ ( f nn ( k ) ( r ) ) 2 ⟩ r = A ( k ) − B ( k ) ∫ η 2 PDF ( η ) d η ∭ Σ dev , z z 2 PDF ( α 1 , α 2 , α 3 ) d α 1 d α 2 d α 3 = 3 5 2 Σ mis A ( k ) − B ( k ) s ( η ) , (59)
with PDF ( η ) \operatorname{PDF}(\eta) PDF ( η ) corresponding to Gaussian distribution N ( 0 , s 2 ( η ) ) \mathcal{N}(0,s^2(\eta)) N ( 0 , s 2 ( η )) and PDF ( α 1 , α 2 , α 3 ) \operatorname{PDF}(\alpha_1,\alpha_2,\alpha_3) PDF ( α 1 , α 2 , α 3 ) to random orientation distribution. For crystal lattices with cubic symmetry, ∣ A ( k ) − B ( k ) ∣ \left|A^{(k)}-B^{(k)}\right| A ( k ) − B ( k ) simplifies to ∣ 3 A ( k ) − 1 ∣ / 2 \left|3A^{(k)}-1\right|/2 3 A ( k ) − 1 /2 .
In overall, the derived expression for s s s suggests that the loading contribution to GB-normal-stress fluctuations can be trivially decoupled as
s ( f n n ( k ) ( r ) ) = Σ m i s F ( A μ , E 12 , … ) . (60) s\left(f_{nn}^{(k)}(\mathbf{r})\right)=\Sigma_{\mathrm{mis}}\,\mathcal{F}(A^{\mu},E_{12},\ldots). \tag*{(60)} s ( f nn ( k ) ( r ) ) = Σ mis F ( A μ , E 12 , … ) . (60)
Macroscopic response of random grain boundaries
The upgraded models for local stresses σ ~ n n ( k ) \widetilde{\sigma}_{nn}^{(k)} σ nn ( k ) and their macroscopic manifestation PDF ( σ ~ n n ( k ) ) \operatorname{PDF}(\widetilde{\sigma}_{nn}^{(k)}) PDF ( σ nn ( k ) ) are typically used for GBs of a certain GB type, corresponding to fixed values of E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 (together with δ E 12 \delta E_{12} δ E 12 and δ ν 12 \delta\nu_{12} δ ν 12 ). The response of random GBs can therefore be estimated by integration over all GB types, hence taking into account all GBs in a given aggregate,
PDF r n d ( σ ~ n n ( k ) ) = ∬ w ( E 12 , ν 12 ) PDF ( σ ~ n n ( k ) ) d E 12 d ν 12 , (61) \operatorname{PDF}_{\mathrm{rnd}}(\widetilde{\sigma}_{nn}^{(k)})=\iint w(E_{12},\nu_{12})\operatorname{PDF}(\widetilde{\sigma}_{nn}^{(k)})\,dE_{12}d\nu_{12}, \tag*{(61)} PDF rnd ( σ nn ( k ) ) = ∬ w ( E 12 , ν 12 ) PDF ( σ nn ( k ) ) d E 12 d ν 12 , (61)
where w ( E 12 , ν 12 ) w(E_{12},\nu_{12}) w ( E 12 , ν 12 ) represents the distribution function of GB types in an aggregate and PDF ( σ ~ n n ( k ) ) \operatorname{PDF}(\widetilde{\sigma}_{nn}^{(k)}) PDF ( σ nn ( k ) ) the macroscopic response of a specific GB type; cf. Eq. (37) .
For aggregates with zero crystallographic texture, the response of random GBs can also be obtained more elegantly. There, the average σ n n \sigma_{nn} σ nn of all GBs with the same GB normal n n n (but arbitrary E 12 E_{12} E 12 and ν 12 \nu_{12} ν 12 ) should be equal to the external loading Σ \boldsymbol{\Sigma} Σ , projected onto that GB plane, i.e., ⟨ σ n n ⟩ = σ n n ( 0 ) = Σ z z \langle\sigma_{nn}\rangle=\sigma_{nn}^{(0)}=\Sigma_{zz} ⟨ σ nn ⟩ = σ nn ( 0 ) = Σ z z . This is true because crystallographic orientations of grains are not correlated with orientations of GB planes, hence randomly distributed grain orientations are expected along every GB-normal direction n n n , providing an average (bulk) response. Thus,
PDF r n d ( σ ~ n n ( k ) ) ≈ PDF ( σ ~ n n ( 0 ) ) , (62) \operatorname{PDF}_{\mathrm{rnd}}(\widetilde{\sigma}_{nn}^{(k)})\approx\operatorname{PDF}(\widetilde{\sigma}_{nn}^{(0)}), \tag*{(62)} PDF rnd ( σ nn ( k ) ) ≈ PDF ( σ nn ( 0 ) ) , (62)
or, equivalently,
PDF r n d ( σ ~ n n ( k ) ) ≈ PDF ( σ ~ n n ( k ≥ 1 ) ) ∣ E 12 + d E 12 = 1 E 3 + d E 3 = 1 ν 12 + d ν 12 = ⟨ ν ⟩ . (63) \operatorname{PDF}_{\mathrm{rnd}}(\widetilde{\sigma}_{nn}^{(k)})\approx\left.\operatorname{PDF}(\widetilde{\sigma}_{nn}^{(k\geq1)})\right|_{\substack{E_{12}+dE_{12}=1\\E_3+dE_3=1\\\nu_{12}+d\nu_{12}=\langle\nu\rangle}}. \tag*{(63)} PDF rnd ( σ nn ( k ) ) ≈ PDF ( σ nn ( k ≥ 1 ) ) E 12 + d E 12 = 1 E 3 + d E 3 = 1 ν 12 + d ν 12 = ⟨ ν ⟩ . (63)
In practice, the (approximate) PDF of stress response in a random aggregate can be obtained simply by convoluting the isotropic solution PDF( σ n n ( 0 ) ) (\sigma_{nn}^{(0)}) ( σ nn ( 0 ) ) with the Gaussian distribution N ( 0 , s 2 ( f n n ) \mathcal{N}(0,s^2(f_{nn}) N ( 0 , s 2 ( f nn ) , taking s ( f n n ) s(f_{nn}) s ( f nn ) from Eq. (38) .
Grain-boundary-normal distribution in aggregates with elongated grains
In aggregates with zero morphological texture (i.e., with no preferred direction for grain shapes), the GB normals, n = ( sin θ cos ψ , sin θ sin ψ , cos θ ) n=(\sin\theta\cos\psi,\sin\theta\sin\psi,\cos\theta) n = ( sin θ cos ψ , sin θ sin ψ , cos θ ) , are uniformly distributed on a sphere, with corresponding distribution functions for the two angles
f ( cos θ ) = 1 2 ; ( − 1 ≤ cos θ ≤ 1 ) , f(\cos\theta)=\frac{1}{2};\qquad(-1\leq\cos\theta\leq1), f ( cos θ ) = 2 1 ; ( − 1 ≤ cos θ ≤ 1 ) ,
f ( ψ ) = 1 2 π ; ( 0 ≤ ψ ≤ 2 π ) . (64) f(\psi)=\frac{1}{2\pi};\qquad(0\leq\psi\leq2\pi). \tag*{(64)} f ( ψ ) = 2 π 1 ; ( 0 ≤ ψ ≤ 2 π ) . (64)
To generate an aggregate with grains elongated along the z z z -axis, a simple geometrical scaling can be applied to the initially isotropic aggregate
x → x , y → y , z → λ z z ; ( λ z > 0 ) . (65) x\to x,\qquad y\to y,\qquad z\to\lambda_z z;\qquad(\lambda_z>0). \tag*{(65)} x → x , y → y , z → λ z z ; ( λ z > 0 ) . (65)
As a result of such transformation, the two distribution functions become
f ( cos θ ) = λ z 2 ( 1 1 + ( λ z 2 − 1 ) cos 2 θ ) 3 / 2 ; ( − 1 ≤ cos θ ≤ 1 ) , f(\cos\theta)=\frac{\lambda_z}{2}\left(\frac{1}{1+(\lambda_z^2-1)\cos^2\theta}\right)^{3/2};\qquad(-1\leq\cos\theta\leq1), f ( cos θ ) = 2 λ z ( 1 + ( λ z 2 − 1 ) cos 2 θ 1 ) 3/2 ; ( − 1 ≤ cos θ ≤ 1 ) ,
f ( ψ ) = 1 2 π ; ( 0 ≤ ψ ≤ 2 π ) . (66) f(\psi)=\frac{1}{2\pi};\qquad(0\leq\psi\leq2\pi). \tag*{(66)} f ( ψ ) = 2 π 1 ; ( 0 ≤ ψ ≤ 2 π ) . (66)
The derived distributions are used to predict the stress response PDF ( σ n n ( k ) ) \operatorname{PDF}(\sigma_{nn}^{(k)}) PDF ( σ nn ( k ) ) of any GB type within the elongated aggregate{sup}33 (see {xref}Sec. IV`).
References