Critical analyses of RVE concepts in local and peridynamic micromechanics of composites

arXiv:2402.13908 · cond-mat.mtrl-sci · Submitted 2024-02-21 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Condensed Matter. The day in condensed matter: materials, phases and the measurements under them.

Kai: We're looking at the paper: "Critical analyses of RVE concepts in local and peridynamic micromechanics of composites".

Mira: Here's the gist of the paper. A static peridynamic composite material (CM) is analyzed within Computational Analytical Micromechanics (CAM), which establishes a formal similarity between Local Micromechanics (LM) and Peridynamic Micromechanics (PM).

Kai: CAM develops for inhomogeneous body forces with compact support to estimate the effective behavior of CMs using a novel Representative Volume Element (RVE) concept that is independent of constitutive laws or the form of the predicted effective operator, thereby eliminating potential inaccuracies related to size scale, boundary layer, and edge effects.

Mira: The paper discusses key micromechanics concepts such as LM and PM.

Kai: Nonlocal mechanics was developed by Kröner, Eringen, Kunin, Baˇzant and others.

Title and authors: Mira: The paper is titled "Critical analyses of RVE concepts in local and peridynamic micromechanics of composites," written by Valeriy A Buryachenko. It sets up this comparison between the local and peridynamic methods for composite materials, focusing on how the representative volume element plays a central role.

Kai: So, they're looking at both random and periodic structures in static peridynamic composite materials. The paper suggests that local micromechanics and peridynamic micromechanics are formally similar when applied to these composites.

Lev: What does this similarity actually mean for the modeling process? Is it a shortcut or a necessary connection they need to make?

Mira: It allows for the straightforward generalization of local micromechanics methods directly to their peridynamic counterparts. This is important because it connects two different ways of describing material response.

Kai: And they develop this new representative volume element concept specifically for inhomogeneous body forces with compact support. This concept helps them estimate the effective behavior of these composites using a new compressed dataset, which is key.

Lev: So this RVE concept is presented as a way to bypass some of the usual problems that arise when you try to model infinite media with finite samples or boundaries.

Mira: That’s right. The preparation of these datasets is selected by using this fundamentally new RVE concept that doesn't depend on neither on the constitutive law nor the form of the predicted effective operator, which is what makes it a novel approach for estimation.

The paper's summary: Kai: This paper summarizes how they tackle these micromechanics concepts by defining key terms like local micromechanics and peridynamic micromechanics. Nonlocal mechanics itself was developed back in the sixties by researchers like Kröner, Eringen, Kunin, Bažant and others.

Mira: Peridynamics replaces those local partial differential equations with integral equations based on internal forces exerted by surrounding points within a finite horizon. It’s a big shift from classical local theories that only look at immediate neighbors and spatial derivatives of displacement.

Lev: When they talk about the representative volume element, they emphasize that the correct RVE size needs to capture microstructural heterogeneity while avoiding those artificial size scales and edge effects we see in simulations.

Kai: They also introduce a scale separation hypothesis, which is defined by lambda small less than lambda times L where lambda is the typical length scale of microstructural heterogeneity, lambda times Lambda is an applied field scale, and L is the macroscopic length scale.

Mira: And they discuss general integral equations or GIEs as a framework that describes material behavior under nonlocal influences without directly relying on Green's functions or specific constitutive laws.

Lev: The main result they highlight is that effective elastic moduli can be found through statistical averages of fields, leading to representations like the nonlocal effective constitutive law involving h sigma i at x and L tau i at y.

The paper's improvements: Kai: So what are the proposed improvements in this paper? They introduce nonlinear GIEs that can treat both local thermoelastic models and peridynamic models, bridging that gap between the local and nonlocal theories.

Mira: To achieve this, they use datasets from micromechanical problems like Dr or Dp to construct surrogate integral operators using machine learning and neural network techniques. This is a way to build these complex operators without doing the full direct numerical simulation.

Lev: The RVE concept itself is formalized in definition five point three, defining it as the area where effective parameters are stabilized outside it, which enables modeling infinite media using a finite sample. That feels like a really practical tool for implementing machine learning methods.

Kai: They show that by exploiting the dataset D instead of the full direct numerical simulation data D DNS, they can derive four different surrogate operators for things like bulk modulus, stress tensor components, and strain components.

Mira: This approach allows them to get nonlocal operator counterparts of basic local micromechanics concepts, like the effective concentration factor. It's a way to use statistical averages from micromechanical problems instead of solving the full problem every time.

Conclusion: Lev: So wrapping up, the paper argues that this new approach using compressed statistical average fields, denoted as Dr or Dp, is a way to model infinite media without edge effects. It uses these methods to create surrogate nonlocal operators for basic local micromechanics concepts.

Kai: The main implication is that we can use these cheaper and faster methods instead of relying on direct numerical simulation data for modeling composite materials under inhomogeneous loads. This gives us a more robust way to get estimates of effective behavior.

Mira: They show that the RVE concept, defined in definition five point three, acts as a really working tool for correctly implementing machine learning and neural network techniques when deriving these effective operators without it. It’s the icing on the cake reflecting the micromechanical nature of those derived operators.

Lev: From an engineering standpoint, this means we can use these surrogate models to run simulations that are much faster than solving a full problem, which is crucial for real hardware development where you need quick feedback.

Kai: So, in summary, "Critical analyses of RVE concepts in local and peridynamic micromechanics of composites" presents a way to bridge local and nonlocal theories using machine learning on micromechanical data to get reliable effective behavior estimates.

Mira: It's a solid piece of work because it connects the theoretical framework with practical ways to build these predictive tools for materials science.

Lev: The core idea is making sure that the RVE concept is used correctly in this way, and that those surrogate operators actually work for random or periodic structures under varied loading conditions.

Micromechanics & Composites LLC

cond-mat.mtrl-sci

Submitted: 2024-02-21

Updated: 2024-12-12

Comments: 71 pp., 336 refs. This is a different paper (not a version of the paper Buryachenko V. A. (2024) arxiv.org/abs/2402.13908v4 (109pp., 466 refs.))

Journal ref: J. Period. Nonlocal Modeling, 7: art. 9, 2025

DOI: 10.1007/s42102-025-00132-8

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 81/100

The gist: A static peridynamic composite material (CM) is analyzed within Computational Analytical Micromechanics (CAM)…

Key concepts

Representative Volume Element (RVE)
The correct RVE must capture the material's microstructural differences without introducing artificial scales or edge effects. It is defined as the area where effective parameters stabilize, ensuring the material's response remains independent of boundary conditions, which is crucial for accurate modeling.
Scale Separation Hypothesis
This hypothesis defines scale separation by comparing three length scales: $\Lambda$ (applied field scale), $\lambda$ (microstructural heterogeneity length), and $L$ (macroscopic scale). It assumes that the applied field is much larger than the microstructure ($\Lambda \ll L$), which justifies using simplified, averaged models for effective material behavior.
General Integral Equations (GIEs)
GIEs describe material behavior under nonlocal influences by using integral equations based on internal forces exerted by surrounding points within a finite horizon. This framework allows for modeling nonlocal effects without needing explicit constitutive laws or Green's functions, bridging local and nonlocal theories.
Nonlocal Neural Operators
These are neural network architectures where the connection between layers is modeled by a nonlocal integral operator $K_l$. This structure captures long-range dependencies in the feature space, allowing the system to learn complex, non-local material relationships effectively.

Terminology

Summary

A static peridynamic composite material (CM) is analyzed within Computational Analytical Micromechanics (CAM), which establishes a formal similarity between Local Micromechanics (LM) and Peridynamic Micromechanics (PM). p. 0

Where the paper says it

Solids 2000; 48:175–209) composite materials (CMs) of the random and periodic structures are considered. In the framework of the second background of micromechanics (also called computational analytical micromechanics, CAM), one proved that local micromechanics (LM) and peridynamic micromechanics (PM) are formally similar to each other for CM of both random and periodic structures.

CAM develops for inhomogeneous body forces with compact support to estimate the effective behavior of CMs using a novel Representative Volume Element (RVE) concept that is independent of constitutive laws or the form of the predicted effective operator, thereby eliminating potential inaccuracies related to size scale, boundary layer, and edge effects. p. 0
Where the paper says it

For the inhomogeneous body force with compact support, CAM is developed for the estimation of the effective behavior of CM which is used as a new compressed dataset. The preparation of these datasets is selected by the use of a fundamentally new RVE concept that directly depends on neither constitutive laws of phases nor the form of the predicted effective (surrogate) operator.

The paper discusses key micromechanics concepts such as LM and PM. p. 0

Where the paper says it

Exploitation of the new RVE concept eliminates any potential incorrectness related to both size scale, boundary layer, and edge effects. Keywords: Microstructures; inhomogeneous material; peridynamics; non-local methods; multiscale modeling 1 Introduction The key concepts of micromechanics involve predicting the mechanical behavior of composite materials (CM) based on the mechanical properties of their phases and microstructure.

Nonlocal mechanics was developed by Kröner, Eringen, Kunin, Baˇzant and others. p. 1
Where the paper says it

Notable references include comprehensive texts and books in works such as those by authors [40], [61], [106], [163], [223],[280]. Nonlocal mechanics was initially developed in the 1960s by several researchers such as Kr¨oner, Eringen, Kunin, Baˇzant and others (see for references [61]).

Peridynamics replaces local partial differential equations with integral equations based on internal forces exerted by surrounding points within a finite horizon. p. 1
Where the paper says it

Peridynamics replaces the local partial differential equations (such as the balance of linear momentum) with integral equations, which are free of spatial derivatives of displacement. In this framework, the equilibrium of a material point is governed by the sum of internal forces exerted by the surrounding points within a finite distance (referred to as the horizon).

Central to the discussion are:

  • Representative Volume Element (RVE): The correct RVE size must capture microstructural heterogeneity while avoiding artificial size scales and edge effects, ensuring the material's response is independent of boundary conditions.

  • Scale Separation Hypothesis: A separation of scales is defined by lambda much less than L, where lambda is the typical length scale characteristic of microstructural heterogeneity, lambda is an applied field scale, and L is the macroscopic length scale.

  • General Integral Equations (GIEs): The GIE framework describes material behavior under nonlocal influences without direct reliance on Green's functions or constitutive laws.

  • Effective Elastic Moduli: These are obtained through statistical averages of fields, leading to representations like the nonlocal effective constitutive law: h sigma i(x) = LL(0)h epsilon i(x) + hL tau i(y), where L tau (y):= LL1(y) epsilon(y), which corresponds to the locally elastic constitutive law at the microlevel.

The paper proposes nonlinear GIEs capable of treating both local thermoelastic models and peridynamic models, bridging the gap between local and nonlocal theories. p. 5

Where the paper says it

Buryachenko strongly nonlocal (strain type and displacement type, peridynamics) and weakly nonlocal (strain-gradient, stress-gradient, and higher-order models) forms of these nonlocal operators. Perhaps the most challenging issue is how micromechanics can contribute to the understanding of the bridging mechanism between the coupled scales, which is described by the nonlocal constitutive equations involving the parameters of a relevant effective nonlocal operator.

This approach utilizes datasets derived from micromechanical problems (Dr or Dp) to construct surrogate integral operators using Machine Learning (ML) and Neural Network (NN) techniques. p. 0
Where the paper says it

The preparation of these datasets is selected by the use of a fundamentally new RVE concept that directly depends on neither constitutive laws of phases nor the form of the predicted effective (surrogate) operator. The mentioned datasets with intrinsically incorporated new RVE concept as a necessary ingredient can be implemented into any known machine learning (ML) and neural network (NN) technique used for the prediction of nonlocal surrogate operators.

The RVE concept is formalized in Definition 5. p. 5
Where the paper says it

It leads to the definition of a conceptually new RVE associated with a prescribed effective nonlocal operator (which is usually second order differential operator) for either random ([103], [104], [105]) or periodic ([8], [176], [177], [272])) structure CMs. The significance of RVE concept is drastically increased at the simultaneous analysis of three sorts of nonlocal effects generated by both inhomogeneous applied fields, material nonlocality (horizon), and interactions between inclusions.

3, defining the RVE as the area where effective parameters are stabilized outside it, enabling the modeling of infinite media using a finite sample. p. 1
Where the paper says it

Buryachenko are not feasible. In such cases, finite element analysis (FEA) is widely used in combination with the truncation method modeling the infinite medium by expanding the sample size.

The perspective on data-driven ML in CM modeling was initiated by Silling [264], [315]. p. 5

Where the paper says it

This is the machine learning (ML) and neural network (NN) technique. The prospective direction of data-driven ML technique in the modeling of CMs was initiated by Silling [264], [315] (see also [317]).

A key development involves constructing surrogate nonlocal operators for infinite media based on Direct Numerical Simulation (DNS) data. p. 4
Where the paper says it

Classical definition by Hill [143] implies that these boundary conditions are macroscopically homogeneous and the effective behavior is described by the tensor of effective moduli. The mentioned sample response is estimated from direct numerical simulation (DNS) of microstructural volume elements (MVEs) simulated or extracted, e.

While datasets like D DNS contain comprehensive information, the concept of compressed applied to the field PM dataset D r does not necessitate preliminary estimation of the full-field dataset. The goal of micromechanics is to provide estimations via cheaper, faster, and more robust methods than DNS. p. 2
Where the paper says it

This makes the GIE more flexible and generalizable to a broader range of material behaviors providing a fundamental jump in multiscale and multiphysics research with drastically improved accuracy of local field estimations (even to the point of correction of a sign, see [61]). It should be mentioned the stochastic methods where the material prop-

The paper details various methods for constructing surrogate operators:

  1. Combining AMic and CMic: A new nonlocal effective operator similar to (3.322) is obtained by combining results from AMic and CMic methods, applicable to both random and periodic structure CMs with local phase properties. p. 5

    Where the paper says it

    Thus, instead of the classical effective moduli Hill [143] we deal with the effective nonlocal operators (either the integral or differential forms) even for CMs with locally elastic properties of phases. It leads to the definition of a conceptually new RVE associated with a prescribed effective nonlocal operator (which is usually second order differential operator) for either random ([103], [104], [105]) or periodic ([8], [176], [177], [272])) structure CMs.

  2. Nonlocal Neural Operators: The authors explore the use of nonlocal neural operators, where the increment between layers is modeled as a nonlocal (integral) operator K l, capturing long-range dependencies in feature space: z l(x) = A(w l z l-1(x) + b l + (K l (z l-1)(x))). Various architectures like DeepONet, FNO, and LNO have been reviewed. p. 2

    Where the paper says it

    (see [278]). The versatility and effectiveness of peridynamic models has been demonstrated in various practical applications, including damage accumulation, corrosion, the fracture and failure of composites of deterministic structure, crack instability, impact, the fracture of polycrystals, and nanofiber networks (see, e.

  3. Physics-Informed Neural Networks (PINNs): PINNs embed physics equations as constraints in the loss function (L tot = omega g L gov + omega u L u BC + omega f L f BC + omega d L data), ensuring training results align with physical laws. A neural operator can be combined with PINNs to learn complex nonlinearity, multi-material heterogeneity, and nonlocality. p. 5

    Where the paper says it

    Physics-informed neural networks (PINN, see [85], [139], [141],[153], [167], [173], [239], and [244]) have attracted considerable attention because this framework embeds physics equations into the NN as constraints, ensuring the training results correspond to physical laws. A neural operator can be combined with the PINN methods (see [122], [135], [286]) to train a model that can learn complex nonlinearity, multi-material heterogeneity, and nonlocality in physical systems with extremely high generalization accuracy.

  4. Energy-Informed Neural Networks (EINN): EINN loss functions incorporate the total potential of the deformed body and residuals of boundary material points to enforce boundary conditions of a finite-size sample.

The crucial technical advancement is the application of these methods to CAMNN, which incorporates nonlinear micromechanical models (4. p. 0

Where the paper says it

Keywords: Microstructures; inhomogeneous material; peridynamics; non-local methods; multiscale modeling 1 Introduction The key concepts of micromechanics involve predicting the mechanical behavior of composite materials (CM) based on the mechanical properties of their phases and microstructure. The process often requires analyzing the stress field in the constituent phases of the material, typically by using micromechanical models that account for inclusions within a matrix subjected to some homogeneous effective field.

15)-(4.22) into PNO (and PINN) by replacing full-field datasets (D DNS) with compressed statistical average fields (D r or D p). p. 5
Where the paper says it

The heterogeneous PNO (HeteroPNO) approach is proposed in [160] for data-driven constitutive modeling of heterogeneous fiber orientation field in anisotropic materials. Physics-informed neural networks (PINN, see [85], [139], [141],[153], [167], [173], [239], and [244]) have attracted considerable attention because this framework embeds physics equations into the NN as constraints, ensuring the training results correspond to physical laws.

This eliminates edge and size effects. p. 0
Where the paper says it

The mentioned datasets with intrinsically incorporated new RVE concept as a necessary ingredient can be implemented into any known machine learning (ML) and neural network (NN) technique used for the prediction of nonlocal surrogate operators. Exploitation of the new RVE concept eliminates any potential incorrectness related to both size scale, boundary layer, and edge effects.

However, the RVE concept in Definition 5. p. 5
Where the paper says it

It leads to the definition of a conceptually new RVE associated with a prescribed effective nonlocal operator (which is usually second order differential operator) for either random ([103], [104], [105]) or periodic ([8], [176], [177], [272])) structure CMs. The significance of RVE concept is drastically increased at the simultaneous analysis of three sorts of nonlocal effects generated by both inhomogeneous applied fields, material nonlocality (horizon), and interactions between inclusions.

3 is highlighted as a real working tool for correct implementation of ML and NN techniques, acting as the icing on the cake reflecting the micromechanical nature of effective operators derived without it. Exploiting D instead of D DNS yields four surrogate operators (G gamma(h nu)(x) for gamma = b, sigma, u i, sigma i) instead of a single operator G(h nu)(x). p. 5
Where the paper says it

Perhaps the most challenging issue is how micromechanics can contribute to the understanding of the bridging mechanism between the coupled scales, which is described by the nonlocal constitutive equations involving the parameters of a relevant effective nonlocal operator. Thus, instead of the classical effective moduli Hill [143] we deal with the effective nonlocal operators (either the integral or differential forms) even for CMs with locally elastic properties of phases.

This approach allows the derivation of nonlocal operator counterparts of basic LM concepts like the effective concentration factor. p. 0
Where the paper says it

The prediction of the stress field relies on solving a basic problem that represents a single ellipsoidal inclusion within an infinite, homogeneous matrix subjected to a uniform effective field; this solution has an analytical representation in the form of Eshelby [116] tensor (see for references [236], [333]). For more complex inclusion shapes, analytical solutions like Eshelby’s tensor ∗This is a different paper (not a version of the paper Buryachenko V.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Surrogate operator construction for machine learning systems can be improved by using datasets Dr or Dp, which are defined by statistical averages in micromechanical problems (Eqs. 6.1 and 6.2). This allows ML techniques to obtain a surrogate dataset for any macroscopic loading without solving the micro-problem as in Eq. 4.13, leading to more efficient modeling of infinite media.

  2. The system can be enhanced by using the data-driven kernel L gammahuki = Z K gamma(x - y)(huki(y) - huki(x)) dy (Eq. 6.3). This surrogate model, optimized via the minimization of the loss function in Eq. 6.4, can be used to approximate the effective nonlocal operator for random or periodic structure composites with high accuracy.

  3. The AI system can be adapted to handle body force with compact support b(x) as a prospective tool for data-driven learning of surrogate nonlocal constitutive laws (Section 6). This allows the surrogate model to learn constitutive laws directly from loading parameters rather than just a particular loading parameter, increasing its predictive power.

  4. For periodic structure composites, the system can be configured to use datasets Dp (Eq. 4.41) instead of Dr (Eq. 6.1). This configuration allows for the construction of surrogate operators that are applicable to periodic media with any constitutive laws and any inhomogeneous loading, which is a significant generalization over Dr.

  5. The system can be generalized to handle arbitrary nonlocal elastic properties of phases by using the dataset Dp (Eq. 4.41) in Section 6, as CAM can be looked upon as being proposed for any nonlocal elastic properties of phases used in Eqs. 4.25-30, extending its applicability beyond linear cases.

Abstract

A static peridynamic (proposed by Silling, see J. Mech. Phys. Solids 2000; 48:175--209) composite materials (CMs) of the random and periodic structures are considered. In the framework of the second background of micromechanics (also called computational analytical micromechanics, CAM), one proved that local micromechanics (LM) and peridynamic micromechanics (PM) are formally similar to each other for CM of both random and periodic structures. It allows for the straightforward generalization of LM methods to their PM counterparts. The representative volume element (RVE) playing a central role in the LM is generalized to PM. For the inhomogeneous body force with compact support, CAM is developed for the estimation of the effective behavior of CM which is used as a new compressed dataset. The preparation of these datasets is selected by the use of a fundamentally new RVE concept that directly depends on neither constitutive laws of phases nor the form of the predicted effective (surrogate) operator. The mentioned datasets with intrinsically incorporated new RVE concept as a necessary ingredient can be implemented into any known machine learning (ML) and neural network (NN) technique used for the prediction of nonlocal surrogate operators. Exploitation of the new RVE concept eliminates any potential incorrectness related to both size scale, boundary layer, and edge effects.

Sources

Related papers