1. Introduction
Reliable and predictable behavior of fiber-reinforced polymers (FRPs) is critical to design safe and cost-effective structural FRP components used across various industries, including aerospace, automotive, construction, etc. [
1]. The uncertainty in the performance of FRPs is mainly sourced from defects introduced during the manufacturing process, random microstructure morphology, inherent uncertainties in the properties of FRP constituents, and others [
2]. By using a robust manufacturing process, the uncertainties related to the behavior of FRPs can be mitigated [
3]. Furthermore, comprehending the impact of design parameters such as matrix type, fiber type, and fiber volume fraction (
) on FRP behavior is essential, where by controlling and optimizing these parameters, uncertainties in the mechanical properties of FRPs can potentially be mitigated, leading to an improvement in reliability. For example, the inherent uncertainties in the mechanical properties of natural fibers due to the naturalistic production process of these fibers make the employment of synthetic fibers such as carbon fibers and glass fibers more preferable from the reliability viewpoint because they are produced through controlled process [
4,
5]. Similarly, the matrix type, whether it falls under the category of thermosets or thermoplastics, could heavily affect the overall FRP reliability.
is one of the main design parameters of FRP material. It is evident, both intuitively and from the existing literature [
6,
7], that increasing the
would significantly improve the stiffness and strength in the longitudinal direction in a unidirectional (UD) FRP. However, this does not appear to hold true for the transverse direction, particularly concerning strength, because the failure strength in this direction is primarily dictated by the properties of the matrix and fiber–matrix interface. In the light of the inevitable effect of
on the properties in the transverse direction, pertinent and justifiable questions arise: How and to what extent does the changing of
affects the various mechanical properties of UD FRP? Would this effect be reflected in the reliability and predictability of UD FRP behavior? Does changing
produce a consistent effect across different types of UD FRPs?
To answer these questions, a significant number of experiments under different situations and loading scenarios are required. Conducting these experiments is costly and necessitates a significant investment of time and resources. With advancements in technology and increasing computational power, virtual testing through numerical simulation has emerged as a viable alternative to actual testing. This approach offers cost and time savings and enables us to conduct experiments that simulate real-world scenarios and may be difficult to replicate in a laboratory [
8,
9]. Given the versatility of FRPs, this approach becomes a vital tool for optimizing the design and reliability assessment of the structural FRP components [
2]. However, it is crucial to validate the numerical models through actual testing before utilizing them for virtual testing purposes.
As stated by George Box, “All models are wrong, but some are useful”. In the literature, numerous modeling and virtual testing methods have been developed to aid in the design and optimization of structural FRP components. One of the most effective approaches for developing these virtual testing capabilities is the bottom-up multiscale strategy [
10], see
Figure 1. This strategy includes three length scale models, viz., the macroscopic, mesoscopic, and microscopic levels models, where different geometric features are presented [
11,
12]. The bottom-up multiscale strategy was primarily employed in previous research to build numerical models that simulate the actual behavior of FRP laminates and components (i.e., elastic linear behavior, fracture development, and failure mechanism) under different loading scenarios and for different purposes [
13,
14,
15,
16,
17,
18,
19,
20,
21,
22]. Also, this strategy was successfully employed in many studies [
23,
24,
25] to develop a framework for multiscale uncertainty quantification and to comprehend how uncertainties at smaller scales, such as th nanoscale, microscale, and mesoscale levels, are reflected in or propagated to the macroscale level. It is worth mentioning other multiscale strategies employed in the literature for similar purposes, such as the
Method, Asymptotic Homogenization, Mean Field Approaches, and Transformation Field Analysis, among others. These methods range from purely analytical to highly computational approaches, effectively addressing the complexities of composite behavior, from detailed microstructural characteristics to overall macroscopic responses [
26,
27].
The fundamental tenet of the framework of the multiscale modeling approaches is the concept of the representative volume element (RVE), which was originally put forward by Hill [
28]. RVE models are mainly utilized to predict the homogenized effective properties and damage initiation of FRP plies. A significant challenge associated with the RVE is that it should be geometrically representative of the actual microstructure morphology [
29], meaning that the RVE must be statistically equivalent to a real microstructure [
30,
31]. The effects of the size of an RVE model with random fiber distribution on the predicted homogenized effective properties and localized damage initiation have been investigated over the recent years using different premises related to the source of the RVEs (i.e., algorithmically generated or reconstructed from real micrograph), microstructural features (i.e., fiber shape and micro-voids), constituent materials, finite element model setup, and damage/failure criteria. Most of the studies have drawn varying conclusions regarding the size of the RVE, which is attributed to the inherent differences in their premises. Therefore, it is preferable to conduct a statistical or sensitivity analysis when selecting the optimum RVE size, especially in situations when the premises differ from those adopted in the existing literature.
In the context of the aforementioned background, this study primarily aims to investigate the effect of the on the reliability and predictability of FRP ply behavior, viz., orthotropic elastic properties and damage initiation strengths, using a validated micromechanical virtual testing approach. To make this approach more realistic and resemble actual testing, different types of uncertainties and randomness are considered when developing the micromechanical RVE models, such as the inherent uncertainties present in the properties of the FRP constituents’ materials, random fiber arrangement, and uncertainty in fiber diameter.
This study is conducted on two different types of FRPs, AS4/8552 and E-glass/MTM57, to see whether the observed results are the same for various types of FRPs. To achieve the purpose of this research, experimental data for the studied FRP types and constituents are required to define their properties and uncertainties in the numerical models, and to verify these numerical models. Also, it is required to build numerical RVE models by developing a fast algorithm with a high jamming limit to generate a random microstructure. This is crucial for the computational affordability of the adopted virtual testing approach and study the cases of high . Design of experiment (DOE) coupled with Monte Carlo simulation are needed to create RVE model samples for virtual testing. This study’s significance lies in enhancing the understanding of the behavior, optimizing the design, and improving the predictability and reliability of the manufactured UD FRP plies, which serve as primary building blocks for structural FRP laminates and components, thus achieving a safer and cost-effective design. In addition to the main concern of this study, subsidiary yet important findings can be derived, such as the effect of the on the RVE size, because to the authors’ knowledge, there is no available information in the literature about the optimum size of the RVE for different values, or whether the has an impact on the RVE size in the first place.
In the rest of this article, the most common sources of uncertainty that affect the performance of FRPs and their modeling methods are outlined to provide insights for enhancing virtual testing realism. Moreover, the considered types of FRP materials, the experimental micromechanical characterization of their constituents’ properties, and the development of RVE models and full factorial DOE incorporated with Monte Carlo simulation are detailed. Lastly, the results related to model validity checking and to the effect of on RVE size and on the reliability and predictability of homogenized properties and damage initiation strengths are discussed and interpreted.
2. Sources of Uncertainty in Structural FRP and Their Modeling Methodologies
The presence of uncertainties at different scales in manufacturing structural FRP components (i.e., the constituent level, ply level, laminate level, and component level) has raised significant concerns regarding the safety of FRP structures. The uncertainty in the performance (i.e., mechanical and thermal properties) of structural FRPs at the ply level (microscale level) results from different uncertainty sources. In this study, the first considered source of uncertainty is the inherent uncertainties in the mechanical properties of FRP constituents (fibers and matrix). The experimental works in the literature [
32,
33,
34,
35,
36] have concluded that the probability density function (PDF) of Weibull distribution is the fittest PDF to model the uncertainties in the mechanical properties of the fibers, specifically fibers’ tensile strength, which is explained by the weakest link theory [
37]. Meanwhile, for the polymer matrix stiffness and strength, the PDF of Gaussian distribution is good enough [
38,
39], especially if the matrix material is isotropic. The second considered source is the variations in the fiber diameter. These variations slightly alter the fiber content in the FRP and the size of the interface between the fibers and matrix. In the existing literature, the uncertainty of fiber diameter was identified and measured under the electronic microscope, and modeled using the PDF of the lognormal distribution [
40].
The third considered source is fiber spatial distribution (microstructure morphology). The random spatial arrangement of fibers influences the flow of the polymer matrix. This can result in high accumulation of the matrix in certain areas and low accumulation in others, potentially leading to the formation of voids [
41]. This dispersion in local fiber content causes inhomogeneity in stress and strain distribution, ultimately leading to additional uncertainties and randomness in the mechanical properties of the FRP material. Several algorithms for numerically generating a random microstructure were proposed in the literature to consider the uncertainty associated with fiber spatial distribution. For example, the Random Sequential Adsorption (RSA) method [
42,
43,
44,
45,
46,
47], growth-based algorithms [
48,
49] molecular dynamics-based algorithms [
50,
51], an algorithm based on a constrained optimization formulation [
52], etc. Each of these algorithms has its limitations in terms of jamming limit and execution time. Therefore, several modified versions of these algorithms were proposed by the researchers to overcome these limitations to some extent. In line with this context, this research also contributes by presenting a modified version of an algorithm based on a constrained optimization formulation, as adopted in the study by Pathan et al. [
52]. The modifications to the algorithm mainly aim to speed up the execution time, i.e., reduce the computational cost, without lowering the jamming limit obtained by the original version of this algorithm, because the computational cost is critical for the visibility and practicality of the application of virtual testing and reliability analysis, where hundreds or thousands of numerical models need to be generated and analyzed. The concept, equations, and pseudocode of the algorithm are presented in
Section 4.1.
The uncertainty sourced from the variations in the actual is implicitly accounted for by considering the three prementioned sources of uncertainties. Other sources of uncertainties that stem from defects that occur during the manufacturing process of FRP composites and products, such as voids in the matrix, fiber misalignment, and fiber waviness, are not considered in this study.
4. Developing the Computational Microscale Models
The 3D computational microscale models for virtual testing were developed using the periodic RVE method containing the randomness and uncertainties of the fiber distribution, fiber diameter, , and constituents’ properties. Periodic RVE models are suitable for use only during the elastic regime to analyze uniform stress and strain fields, compute homogenized elastic properties of UD FRP plies, and predict the damage initiation strengths for these plies under various loading conditions. This makes periodic RVE models sufficient for the scope of this research. The development of the periodic RVE models is clarified in the following subsections.
4.1. Microstructure Generation
This research proposes an enhanced version of an algorithm based on a constrained optimization formulation, specifically minimization of the fiber overlapping function, as adopted in the study by Pathan et al. [
52]. The algorithm consists of two main stages:
Initial Guess stage: In this stage, the algorithm starts by randomly generating
diameters (
) of fibers from the corresponding lognormal distribution of fiber diameter (e.g.,
Figure 2a,b), where
is the targeted number of fibers in the RVE model. Then, the size of the RVE (
) is calculated according to the total area of fibers and the targeted
, see Equation (
1). Then, according to uniform distribution, the algorithm proceeds to randomly generate, at the same time, the coordinates (
,
) for the centers of all the
fibers within the range of the RVE dimension (see Equation (
2)). After this, for every fiber that intersects the boundaries of the RVE, a twin fiber is created and located such to guarantee the periodicity of the microstructure. Then, the diameter of each fiber is enlarged with the value of
m.
Figure 5a,c shows examples of initial arrangements of randomly placed fibers; the fibers overlapping are clear.
Optimization stage: To prevent fibers overlapping, the algorithm in this stage will formulate the optimization problem according to Equations (
3)–(
6). These equations represent a feasibility problem; Equation (
3) represents the boundaries of the design space; Equation (
4) is a nonlinear inequality constraint for representing the condition that fibers should not overlap and keep minimum tolerance distance between the fibers (
lf min); while Equation (
5) is also an inequality constraint that guarantees a minimum tolerance distance (
le min) between the fibers and the edges of the RVE.
A feasible solution that could satisfy these constraints can be obtained by minimizing the penalty function (Equation (
6)), which assigns a non-negative penalty (
) for each constraint violation at each fiber, i.e., considering a generic form of the constraints on the coordinates of the
fiber
, the penalty associated with the violation of this constraint is
. Then, the formulated feasibility problem is solved and the feasible solution is obtained by minimizing the penalty function using the
L-BFGS-B algorithm [
67] that is based on the quasi-Newton optimization method. The algorithm will be terminated when the value of the penalty function is equal to zero or the maximum number of iterations is reached. To validate the feasibility of the obtained optimum solution, the original diameter of each fiber is retrieved by subtracting the previously added value of
m, and the value of the penalty function is recalculated. If it is equal to zero, then the solution is feasible; otherwise, the solution will be discarded, and the algorithm will go back to the initial guess stage and start searching for a new solution.
Figure 5b,d illustrate examples of the final arrangements of randomly placed fibers in the RVE. The pseudocode of this algorithm is shown in Algorithm 1, which is scripted and executed in Python.
When the targeted
is high (more than
), the number of feasible solutions decreases, making it difficult to achieve a penalty function of zero value. This is where the added value to the diameter of each fiber plays a role. It helps in speeding up the process of discarding the infeasible solutions and ensuring that the penalty function becomes zero when removing this value from the good solutions with low penalty function values. Therefore, the algorithm exhibited exceptional performance in terms of convergence speed, albeit with a slight trade-off in the jamming limit, which reached approximately
. This value is slightly lower than the
jamming limit in the reference [
52]. However, the convergence speed is significantly higher, as it took less than 5 minutes on average to generate a feasible microstructure of size
with a
of
, which is a lower time in comparison with algorithm of reference [
52], which took 18 min on average to generate a similar microstructure with the same size and
. This is critical and highly beneficial for uncertainty analysis using virtual testing, particularly when a large number of microstructures need to be generated.
Algorithm 1 Pseudocode for generating microstructure |
- 1:
Defining the input parameters: , , PDF of D, , - 2:
While Final_penalty > 0 and rounds < 103: - 3:
D = Generating diameter of fibers - 4:
Calculate L & H (Equation ( 1)) - 5:
Generating random coordinates (Equation ( 2)) - 6:
Generating twin fibers for each fiber intersects the boundaries of RVE - 7:
D = m - 8:
For i in D: - 9:
Assign the i-th fiber coordinate as design variable - 10:
Defining boundaries (Equation ( 3)) - 11:
Defining constraints (Equations ( 4) and ( 5)) - 12:
Defining and calculating the penalty function (Equation ( 6)) - 13:
if penalty > 0: - 14:
Solving the optimization problem (minimize function & L-BFGS-B method) - 15:
Obtaining optimum - 16:
m - 17:
Final_penalty = Recalculate the penalty function (Equation ( 6)) - 18:
rounds += 1 - 19:
if Final_penalty = 0: - 20:
export the fibers coordinates and diameters to text file - 21:
plot the RVE with the randomly distributed fibers
|
4.2. Periodic Boundary Conditions (PBCs) and Displacement Boundary Conditions
The periodic boundary conditions (PBCs) in a general 3D RVE can be expressed in terms of the displacement constraints in vector form, where the relative displacement between the matched nodes on the opposite faces of the RVE (
,
, and
) should be equal to the relative displacement between a set of master nodes (typically
) laying on the opposite faces. According to
Figure 6a, the PBCs can be expressed mathematically as shown in Equations (
7)–(
9), where
L,
H, and
t are the dimensions of the RVE in the
x,
y, and
z directions, respectively.
PBCs were imposed in Abaqus through modifying the keyword of the model to include the equation of PBCs from a text file. While PBCs are necessary to maintain the periodic nature of the RVE, displacement boundary conditions are imposed to generate stresses. Thus, an appropriate set of displacement boundary conditions was applied on the master nodes () for different loading cases.
4.3. Constitutive Models and Discretization of Fiber, Matrix, and Interface
AS4 fibers were modeled as linear, elastic, and transversally isotropic solids. Five independent elastic constants (
,
,
,
,
) and two thermal expansion coefficients (
,
) were defined as clarified in
Table 1 to take the anisotropic behavior between the longitudinal and transverse directions of the AS4 fibers into consideration. Meanwhile, E-glass fibers were modeled as linear isotropic elastic solids with the elastic properties shown in
Table 1. Both AS4 and E-glass fibers were discretized and meshed using 6-
fully integrated wedge elements (
).
The epoxy matrix behaves as a frictional material, exhibiting brittle behavior under tensile loads while exhibiting pronounced plastic behavior in compression. As the yield surface is influenced by hydrostatic pressure, it can be modeled using a Drucker–Prager yield criterion, assuming isotropic elastic–plastic behavior [
68]. The main features of the Drucker–Prager constitutive model are the pressure-dependent yield surface evolution and the distinction between compressive and tensile damage evolution.
Figure 6b shows a scheme of the damage–plasticity model [
69], which is a modification of the Drucker–Prager plasticity yield surface model [
68]. This model necessitates defining the mechanical response under uniaxial tension and compression in addition to the evolution of the yield surface (plasticity) and material degradation (damage). However, the values of the required parameters (
,
,
,
,
, and
) to define the elastic and plastic behavior for both 8552 and MTM57 epoxy matrices using the Drucker–Prager constitutive model were taken as specified in
Table 2. The matrix was discretized and meshed using eight-node reduced integration linear bricks with hourglass control (C3D8R).
Fiber–matrix interface behavior and failure mechanisms (debonding) were modeled using a classical cohesive zone method [
70,
71,
72]. Cohesive elements were inserted at the fiber–matrix interface. The behavior of these elements was defined by the mixed-mode traction–separation law, and their damage was governed by the quadratic stress criterion (Equation (
10)) [
13,
66].
where
,
, and
are the normal and shear components of the traction vector;
denotes the Macaulay brackets, emphasizing that a purely compressive stress state does not initiate damage [
73]. As seen in
Figure 6c, initially, the cohesive element behavior was linear elastic with very high penalty stiffnesses (
,
). When the damage was initiated, a linear softening was induced to capture the degradation of the stiffness up to complete failure. The energy dissipation was computed through the power law under mixed-mode loading [
69]. However, the values of the required parameters to define the described constitutive model of fiber–matrix interface behavior were taken as specified in
Table 3. It is also worth mentioning that the fiber–matrix interface was meshed using eight-node three-dimensional cohesive elements (
).
Figure 6.
(Color online) (
a) Diagram clarifies the derivation of PBC equations; (
b) a constitutive response diagram of the epoxy matrix [
69]; (
c) a diagram of the traction–separation law of the fiber–matrix interface (normal response and damage variable evolution) [
69].
Figure 6.
(Color online) (
a) Diagram clarifies the derivation of PBC equations; (
b) a constitutive response diagram of the epoxy matrix [
69]; (
c) a diagram of the traction–separation law of the fiber–matrix interface (normal response and damage variable evolution) [
69].
4.4. Loading Cases and Steps
To derive the homogenized elastic properties, thermal expansion coefficients, and damage initiation strengths of the investigated FRP plies, various loading scenarios were applied to the periodic RVE model. Each loading case was designed to capture specific properties.
For determining the homogenized elastic properties—namely, Young’s moduli, Poisson’s ratios, and shear moduli—alongside the homogenized thermal expansion coefficients in the three directions, a loading case of six mechanical steps and one thermal step was applied to the periodic RVE. In each of the six mechanical steps, unit stress was applied to distinct RVE faces, oriented in different directions () under suitable boundary conditions, to obtain the homogenized elastic properties. In the thermal step, a homogeneous uniform temperature field with a magnitude of 1 °C was defined, enabling the determination of homogenized thermal expansion coefficients. It should be noted that, in this loading case, the RVE and its constituents are assumed to behave as a linear elastic material, because the inelastic behavior does not play a role in determining the homogenized elastic properties of the FRP plies.
To obtain the damage initiation strengths under uniaxial transverse tension and compression, two distinct loading cases were devised: one applying a strain and the other a strain in the transverse direction. Similarly, two additional loading cases were devised by applying strain in the transverse and longitudinal directions to obtain the damage initiation strengths under transverse and longitudinal shear. However, a homogeneous uniform thermal step was applied initially to consider the effect of the residual thermal stresses stemming from the mismatch of the thermo-elastic constants of fibers and the matrix on the damage initiation strength of the FRP plies. This thermal step modeled the cooling process from the curing temperature (120 °C) to the service temperature (20 °C).
4.5. RVE Size Validation
The size of the RVE should guarantee that it geometrically represents the actual microstructure, meaning the RVE must be statistically equivalent to the real microstructure. A too-small RVE will not be geometrically representative, while a too-large RVE will be computationally expensive. Achieving a good geometric representation while keeping computational costs reasonable requires optimizing the size of the RVE. The optimum size for the RVE is the smallest size that encompasses all the irregularities affecting stress distribution. This ensures that the obtained homogenized properties of the FRP are independent of the RVE size and accurately represent the macroscopic constitutive response. However, there are no specific procedures to predict the optimum size of the RVE for analyzing a particular FRP. It is usually confirmed by a sensitivity study for the convergence of the following: (i) RVE Micromechanical Features, such as homogenized properties, damage initiation strengths, strain energy, energy density, etc., or (ii) RVE Statistical Microstructure Descriptors, classified into Geometric Descriptors and Statistical Functions.
In this research, the optimum size of the RVE was obtained by evaluating the convergence of the effective properties and damage initiation strength. The RVE size was measured and represented in terms of fiber number () for a specific volume fraction () and a consistent aspect ratio of one (), i.e., square-shaped RVEs. Additionally, it is important to note that in all cases, the thickness of the RVE was set to a constant value of m.
4.6. Solver, Postprocessing, and Implementation
The Abaqus standard solver [
69] was employed to analyze the 3D periodic RVE model. The mesh size was defined using a sensitivity analysis. The obtained analysis results were subjected to postprocessing based on the mechanics of materials equations to calculate the homogenized elastic properties and damage initiation strengths.The development and analysis of the 3D periodic RVE model, incorporating the previously described random microstructure, periodic boundary conditions (PBsC), constitutive models, loading cases, sizes, and postprocessing, were scripted and implemented in Python with the help of Abaqus modules, such as
abaqus,
abaqusConstants, and
caeModules. Utilizing scripting to automate the development and analysis of RVE models offers significant flexibility and ensures an efficient workflow for iterative analysis and experimentation.
7. The Effect of vf on the Size of RVE across Various FRP Types
To investigate the effect of
on the optimum size of the RVE, the convergence of virtual test results for UD AS4/8552 and E-glass/MTM57 plies with different
and RVE sizes was studied and analyzed similarly to the method presented in
Section 6. Due to space constraints, only the curves of damage initiation strengths are presented in
Figure 12, as they fundamentally control the required minimum size of the RVE (i.e.,
) that guarantees convergence and low computational cost, as discussed in
Section 6. From this figure, it is evident that the convergence of
,
,
, and
is achieved at a lower
value for RVEs with a lower
compared to those with a higher
. For instance, in the RVEs of AS4/8552 and E-glass/MTM57 with a
of
, employing an
value of one (unit cell model) or at most three would be sufficient in certain scenarios to predict converged homogenized elastic properties and damage initiation strengths. In contrast, for the RVE of AS4/8552 with a
as high as
, the attainment of converged results necessitates the utilization of an
value of at least 14.
The optimal size of the RVE, in terms of
, for each
is graphically represented in
Figure 13 for AS4/8552 and E-glass/MTM57. This figure confirms the findings discussed earlier, namely that an increase in
corresponds to a concurrent increase in the minimum or optimal size of the RVE (i.e.,
) required to ensure the convergence of homogenized mechanical properties. This outcome is explicable by the fact that increasing the
necessitates a larger RVE with a higher
to encompass the full spectrum of irregularities, such as fiber clustering and resin-rich areas, which significantly influence the predicted stress distribution and mechanical properties. In the same context, when the
is extremely low, it appears that a unit cell model in some scenarios, or a model with at most two or three fibers, suffices to capture the full spectrum of irregularities. These findings remain valid regardless of the FRP type being modeled using the RVE, whether it be AS4/8552, E-glass/MTM57, or any other type.