betaFeedback

WorksPhysics and Astronomy

Learning constitutive relations from experiments: 1. PDE constrained optimization

Abstract

We propose a method to accurately and efficiently identify the constitutive behavior of complex materials through full-field observations. We formulate the problem of inferring constitutive relations from experiments as an indirect inverse problem that is constrained by the balance laws. Specifically, we seek to find a constitutive behavior that minimizes the difference between the experimental observation and the corresponding quantities computed with the model, while enforcing the balance laws. We formulate the forward problem as a boundary value problem corresponding to the experiment, and compute the sensitivity of the objective with respect to model using the adjoint method. The resulting method is robust and can be applied to constitutive models with arbitrary complexity. We focus on elasto-viscoplasticity, but the approach can be extended to other settings. In this part one, we formulate the method and demonstrate it using synthetic data on two problems, one quasistatic and the other dynamic.

Introduction

Continuum mechanics is an approach to solving problems involving complex phenomena directly at the scale of applications (e.g., [20, 9]. It exploits universal laws of physics: the balance of mass, momenta, energy as well as the second law of thermodynamics. However these equations are not closed and a constitutive relation that characterizes the material is necessary to do so. This constitutive relation is typically obtained empirically by conducting experiments. However, we cannot measure the constitutive relation directly and it has to be obtained through the solution of an inverse problem. In fact, we cannot even directly measure the quantities like stress, strain, strain rate and energy density that comprise the constitutive relation. Instead, they have to be inferred from quantities like displacements and total forces that can be measured in the laboratory. Thus, the problem of inferring constitutive relations from experiments, and thereby completing the continuum mechanical formulation, requires the solution of an indirect inverse problem.

The classical approach is to design an experimental setup that is consistent with either uniform states of stress and strain/strain-rate (e.g., uniaxial tension in solids or shear rheometers in complex fluids) or a universal solution (e.g., torsion in solids or viscometric flow in complex fluids); these enable a semi-analytic solution to the indirect inverse problem (e.g. [9]). There are a number of difficulties with this approach. First, we obtain information about the constitutive relation only in simple, idealized states of stress (history) while we seek to deploy them in more complex situations. For example, we may obtain only uniaxial tensile data while we may want to use the model in multiaxial, non-proportional loading scenarios. Second, one has to conduct a large number of tests since each test can only provide a limited amount of information on the constitutive behavior. For example, we may have to repeat a test at various strain rates, or with specimens with different orientation. Third, one may need sterile conditions with precise alignment (e.g., plate impact) to obtain the desired state. This limits the number of tests that one can conduct. Together, these and other shortcomings contribute significantly to the uncertainty and limit the fidelity of the constitutive relations that can be obtained from experiments.

It is also classical to postulate a priori an explicit formula for the constitutive relation with a few constants, and then use above-described experiments to fit the constants (e.g. [9]). Since it is common practice to use a very few constants, the limited experimental information to sufficient to fit them. However, the ability of explicit formulas with a few constant to represent complex behavior is limited. Even in the setting of isotropic, incompressible hyperelasticity, we have a still-evolving menagerie of models (e.g. [37, 28]). It is therefore natural to try to go beyond hypoparametrized explicit formulas to hyper-parametrized neural networks and other approximations that have proven to be enormously successful in a variety of fields. However, this requires significant amounts of data, much more than the traditional approach can deliver.

There is a closely related issue. In history dependent and structured continua, one also has to postulate a priori the internal or state variables that describe the history dependence/internal structure [9]. For example, we introduce plastic strain, accumulated plastic strain and a kinematic hardening variable as internal variables in plasticity. However, the choice of internal variables is unclear in complex anisotropic situations like in composite media or shape memory alloys [25]. Instead, one would like to infer these directly from experimental measurements, further increasing the need for empirical data.

The recent decade has seen a revolution in experimental methods primarily driven by the growth of full-field observations techniques. Digital image correlation (DIC) [13, 51, 23], where one infers deformation by comparing images of a complex (speckle) pattern imprinted on a surface taken before and after deformation. The method has been extended to stereo DIC to obtain out-of-plane deformation [50], the high resolution setting using the scanning electron microscope [26], to dynamics [46] and to digital volume correlation [11, 59]. DIC enables the measurement of the entire deformation and strain field on a surface. This opens the intriguing possibility of breaking away from homogeneous (or idealized) states: instead, one can probe many different strain paths in a single test by working with specimens with complex geometries. However, this requires one to solve the difficult indirect inverse problem. While one can measure the strain field, it is limited to a single surface. More importantly, it is not possible to measure the stress field, and instead, we typically can only measure the total reaction force on a surface.

Another rapidly evolving technique that provides full-field information is high energy diffraction microscopy (HEDM) that uses high energy synchrotron radiation and diffraction to obtain the crystal structure on a three-dimensional voxelated volume [7]. Comparing the obtained crystal structure with the stress-free structure gives us the lattice strain on a voxelated volume. One can convert this to stress by using the elastic modulus of the material. However, there are two issues. First, one only obtains the lattice strain, and not the total strain. Therefore, one still has to solve an indirect inverse problem in inelastic phenomena. Second, this is averaged over the volume of the voxel with some unknown kernel, and therefore subject to errors. For example, the resulting stress fields obtained by this approach are not equilibrated [60].

A variety of approaches have been proposed to solve this indirect inverse problem. The key idea is to use the balance laws, and specifically equilibrium in some form. We review a few contributions to provide context to our work, without attempting a comprehensive review of this rapidly growing literature. One approach is to use a combination of model reduction and Gaussian process regression to obtain the parameters associated with a particular constitutive relation (see [54, 55] for the use of this approach in viscoelastic materials). It is possible to combine Bayesian uncertainty quantification with such approaches (see [57] on toughness). While such approaches are effective in identifying a small number of parameters, they scale exponentially with the number of parameters and are prohibitively expensive for complex materials or neural network representations. While we do not address HEDM in this paper, the ideas presented may be extended to this technique.

Miller, Dawson and collaborators pioneered the concurrent use of finite element analysis with given constitutive relations and experimental observations using high energy x-ray diffraction with the goal of understanding the relationship between single crystal elastic modulus and hardening laws and heterogeneity in stress distribution in polycrystals.[17, 40, 56, 12]. However, they do not seek to obtain the full constitutive relationship.

An emerging approach is to use a physics informed neural network (PINN) where the constitutive relation is represented as a neural network and the (failure of the fields to satisfy) balance laws are used as a part of the objective (loss function) to be minimized as a part of the fitting [21, 18, 53]. The balance laws are not enforced exactly in this approach.

In the the virtual field method (VFM) (for example, [19, 44, 32, 35, 29]), the balance laws are integrated against a set of “virtual fields” to obtain a system of equations for the unknown constants in a postulated constitutive relation. The system is linear when the constitutive relation is linear in the constants as in the case of hyperelasticity, but is typically solved by linearization otherwise. The application of the method requires one to specify kinematically admissible virtual fields, and one may taken them to be the measured strain field. However, the strain field is typically measured on a part of the boundary of the specimen, and has to be extrapolated to obtain it over the entire domain. Further, one needs as many independent full field measurements as there are unknown constitutive relation even for complex domains. Finally, only a finite dimensional approximation of the balance laws is enforced.

A widely used approach is the finite element updating (FEMU) (for example, [27, 42, 36, 38, 22, 49, 30, 48, 14] ). The idea is to formulate the problem as an inverse method which is solved iteratively by using a finite element method to solve the forward problem, and then to update the model parameters using a numerically computed sensitivity. This has been applied to a variety of history dependent phenomena and can be integrated with the inverse problem of digital image correlation [36]. An important consideration here is that the sensitivity, or the gradient of the objective with respect to the objective, is computed numerically. While a number of ideas can be used to speed this up, this approach scales poorly with the number of material parameters. A closely related idea is to formulate the problem as a PDE-constrained optimization problem where the balance laws act as the PDE constraint, and then use the adjoint method to compute the sensitivity. This has been applied to elasticity {citep}[39], viscoelasticity {citep}[15] and Norton-Hoff viscoplasticity {citep}[10]`.

In this work, we build on the last two lines of work, and propose a method to accurately and efficiently identify the constitutive behavior through experiments. We formulate this as an indirect inverse problem that is constrained by the balance laws (PDE constraint). The objective that we seek to minimize is the difference between the experimental observation and the corresponding quantities computed using the constitutive model. We then formulate the forward problem as a boundary value problem corresponding to the experiment, and the problem of computing sensitivity of the solution to the parameters as an adjoint problem. This is a partial differential equation that is linear in space and quasilinear in time. The adjoint equation has widely been used in optimal design in mechanics (e.g., [6, 1]). In this work, we implement both the forward and the adjoint problem using finite element approximation. The cost of solving the adjoint problem is that of a single iteration of the forward problem, and the parameters can be updated at little cost from the solution of the adjoint problem. Thus, the core computation of the method is independent of the number of parameters, and thus suited for complex models with numerous parameters (including neural networks).

We formulate and demonstrate the method for both quasistatic and dynamic experiments with synthetic data in the current Part 1. We extend the formulations to include contact in Part 2 and demonstrate it using experimental data from a dynamic indentation test. We demonstrate the method for models formulated as neural networks in Part 3. Together, these show that the method is robust and can be adapted for various experimental situations. While we focus on elasto-viscoplasticity, the ideas are valid for any local constitutive relation.

The paper is organized as follows. We present the method in § 2. We do so for an arbitrary inelastic material described by a rate-dependent internal variable theory in § 2.1, and for a J2 elasto-viscoplastic model with yield in § 2.2. We overview the numerical method in § 2.3 with details provided in the appendix. We demonstrate and validate the method using synthetic data on two examples in § 3: a thick plate with a hole in quasistatic compression in § 3.1 and an extended dynamic impact test in § 3.2. We conclude with some comments in § 4.

Formulation and method

We first present our method for a general history-dependent constitutive relation, and then specialize to elasto-viscoplasticity.

Internal variable theory

Governing equations

We consider an open, bounded domain Ω⊂Rn\Omega\subset\mathbb{R}^{n} occupied by a solid body undergoing a deformation with displacement u(x,t)u(x,t) over time interval (0,T)(0,T). We assume the body is composed of a material of density ρ\rho, and one whose constitutive behavior is described by a set of internal parameters ξ:={ξi}\xi:= \{\xi_i\}, i=1,…,Nξi = 1,\ldots,N_{\xi} that evolve according to a Perzyna-type kinetic law:

σ=S(F,ξ;P),ξ˙=R(F,ξ;P)(1)\sigma= S(F,\xi;P), \\ \dot{\xi} = R(F,\xi;P) \tag*{(1)}

where σ\sigma is the first Piola–Kirchhoff stress, FF is the deformation gradient, and SS and RR are constitutive functions parametrized by material parameters P:={Pi}P := \{P_i\}, i=1,…,Npi = 1,\ldots,N_p. The deformation of the body is then governed by the equations

ρu¨−∇⋅σ=bon Ω,σ=S(∇u,ξ;P),on Ωξ˙=R(∇u,ξ;P)on Ω,σn=fon ∂fΩ,u=u0on ∂uΩ,u∣t=0=u˙∣t=0=ξ∣t=0=0,(2)\begin{aligned} \rho\ddot{u} - \nabla\cdot\sigma&= b && \text{on } \Omega,\\ \sigma&= S(\nabla u,\xi;P), && \text{on } \Omega\\ \dot{\xi} &= R(\nabla u,\xi;P) && \text{on } \Omega,\\ \sigma n &= f && \text{on } \partial_f\Omega,\\ u &= u_0 && \text{on } \partial_u\Omega,\\ u|_{t=0} &= \dot{u}|_{t=0} = \xi|_{t=0} = 0, \tag*{(2)} \end{aligned}

where bb is the applied body force, ff is the applied (dead) traction on part ∂fΩ\partial_f\Omega of the boundary while u0u_0 is the applied displacement on the complement ∂dΩ\partial_d\Omega of the boundary. Note that ∂fΩ\partial_f\Omega typically includes regions that are traction-free. We assume quiescent initial conditions.

Indirect inverse problem of parameter identification

We are given experimental data, usually displacements at certain instances of time on just a portion of the boundary of the domain (some subset of the traction-free portion of ∂fΩ\partial_f\Omega) and some overall force or moment acting on some part of the boundary (some subset of ∂uΩ\partial_u\Omega). We label the experimental data Dexp:={uexp,Mexp}D^{\mathrm{exp}} := \{u^{\mathrm{exp}}, M^{\mathrm{exp}}\}. Here, uexpu^{\mathrm{exp}} is the set of partial displacement measurements, and MexpM^{\mathrm{exp}} is a collection of macroscopic measurable quantities which could consist of net loads or averaged strains. The goal is to find the parameter set PP such that the modeled trajectory and computed macroscopic quantities from solving (2) match the experimental data DexpD^{\mathrm{exp}}. We write this as an optimization problem

inf⁡P∈PO(P,u,ξ,Dexp)subject to {u,ξ} satisfying (2),(3)\inf_{P\in\mathcal{P}} \mathcal{O}(P,u,\xi,D^{\mathrm{exp}})\qquad\text{subject to }\{u,\xi\}\text{ satisfying (2),} \tag*{(3)}

where P\mathcal{P} defines a physical range of parameters and O(P,u,ξ,D)\mathcal{O}(P,u,\xi,D) is the objective or loss function.

We need to choose an objective function O\mathcal{O} that is both computationally efficient and one that attains a minimum when the displacement and forces computed using the solutions {u,ξ}\{u,\xi\} to (2) are equal to the measured values DexpD^{\mathrm{exp}}. A somewhat subtle point in internal variable theories like (1) is that the internal variable is only defined pointwise, and thus does not have a mathematical meaning (specifically a trace) on the boundary. However, the traction is mathematically defined through the interior stress distribution. Thus, we have to interpret the measurement of boundary forces accordingly. In this work we use a finite element discretization, and the governing equations are used in their weak form. Specifically, the displacements are imposed on the nodes, but the strains, stress and internal variables are defined on the quadrature points and inherit their meaning by integration. Therefore, it is natural to write the objective as a volume integral even though the force measurements are made on the boundary. So, we take

O(P,u,ξ,Dexp):=∫0T∫Ωo(P,u,ξ,Dexp) dΩ dt.(4)\mathcal{O}(P,u,\xi,D^{\mathrm{exp}}) := \int_{0}^{T}\int_{\Omega} o(P,u,\xi,D^{\mathrm{exp}})\,d\Omega\,dt. \tag*{(4)}

for some oo that is concentrated near the boundary. We discuss specific choices of oo in Section 3.

Adjoint method for sensitivity

A gradient based optimization approach requires that we compute the sensitivities, that is, the total derivative of the objective with respect to the parameter set PP while enforcing the PDE constraint of the governing equations. We may compute this by applying standard chain rule. However, this would require expressions for dudP\frac{du}{dP} and dξdP\frac{d\xi}{dP}, which are the sensitivity of the solutions of (2) to changes in parameters. These are difficult to compute. Therefore, we use the adjoint method [24, 25] to circumvent the difficulty. We rewrite the objective by adding a term that is zero according to the weak form of the evolution equation,

O=∫0T∫Ωo(P,u,ξ,Dexp) dΩ dt+∫0T{∫Ω[−ρu¨⋅v−σ⋅∇v+b⋅v−ϕ(ξ˙−F)] dΩ+∫∂fΩf⋅v dS} dt,(5)\begin{aligned} \mathcal{O}={}&\int_{0}^{T}\int_{\Omega}o(P,u,\xi,D^{\mathrm{exp}})\,d\Omega\,dt\\ &+\int_{0}^{T}\left\{\int_{\Omega}\left[-\rho\ddot{u}\cdot v-\sigma\cdot\nabla v+b\cdot v-\phi(\dot{\xi}-F)\right]\,d\Omega+\int_{\partial_f\Omega}f\cdot v\,dS\right\}\,dt, \tag*{(5)} \end{aligned}

where vv and ϕ\phi are test functions associated with momentum balance and internal variable evolution. The idea of the adjoint method is to choose the test functions vv and ϕ\phi such that there is no need to explicitly compute the sensitivity of the solutions of (2) to changes in parameters (dudP\frac{d u}{dP} and dξdP\frac{d \xi}{dP}). Differentiating this rewritten objective with respect to PP gives

dOdP=∫0T∫Ω[∂o∂P+∂o∂ududP+∂o∂ξdξdP−ρdu¨dP⋅v−(∂σ∂P+∂σ∂∇u⋅∇(dudP)+∂σ∂ξdξdP)⋅∇v−ϕ(dξ˙dP−∂F∂P−∂F∂∇u⋅∇(dudP)−∂F∂ξdξdP)]dΩ dt.(6)\frac{d\mathcal{O}}{dP}=\int_0^T\int_{\Omega}\left[\frac{\partial o}{\partial P}+\frac{\partial o}{\partial u}\frac{du}{dP}+\frac{\partial o}{\partial\xi}\frac{d\xi}{dP}-\rho\frac{d\ddot{u}}{dP}\cdot v-\left(\frac{\partial\sigma}{\partial P}+\frac{\partial\sigma}{\partial\nabla u}\cdot\nabla\left(\frac{du}{dP}\right)+\frac{\partial\sigma}{\partial\xi}\frac{d\xi}{dP}\right)\cdot\nabla v-\phi\left(\frac{d\dot{\xi}}{dP}-\frac{\partial F}{\partial P}-\frac{\partial F}{\partial\nabla u}\cdot\nabla\left(\frac{du}{dP}\right)-\frac{\partial F}{\partial\xi}\frac{d\xi}{dP}\right)\right]d\Omega\,dt. \tag*{(6)}

We group terms to obtain

dOdP=∫0T∫Ω[(∂o∂P−∂σ∂P⋅∇v+ϕ∂F∂P)−ρv⋅du¨dP−(∇v⋅∂σ∂∇u−ϕ∂F∂∇u)⋅(∇(dudP))+∂o∂ududP−ϕdξ˙dP+(∂o∂ξ−∇v⋅∂σ∂ξ+∂F∂ξϕ)dξdP]dΩ dt.(7)\frac{d\mathcal{O}}{dP}=\int_0^T\int_{\Omega}\left[\left(\frac{\partial o}{\partial P}-\frac{\partial\sigma}{\partial P}\cdot\nabla v+\phi\frac{\partial F}{\partial P}\right)-\rho v\cdot\frac{d\ddot{u}}{dP}-\left(\nabla v\cdot\frac{\partial\sigma}{\partial\nabla u}-\phi\frac{\partial F}{\partial\nabla u}\right)\cdot\left(\nabla\left(\frac{du}{dP}\right)\right)+\frac{\partial o}{\partial u}\frac{du}{dP}-\phi\frac{d\dot{\xi}}{dP}+\left(\frac{\partial o}{\partial\xi}-\nabla v\cdot\frac{\partial\sigma}{\partial\xi}+\frac{\partial F}{\partial\xi}\phi\right)\frac{d\xi}{dP}\right]d\Omega\,dt. \tag*{(7)}

The terms that include du¨dP\frac{d\ddot{u}}{dP} and dξ˙dP\frac{d\dot{\xi}}{dP} are integrated by parts temporally, while the term that includes ∇(dudP)\nabla\left(\frac{du}{dP}\right) is treated with divergence theorem in space. This yields

dOdP=∫0T∫Ω[(∂o∂P−∂σ∂P⋅∇v+ϕ∂F∂P)+[−ρv¨+∂o∂u+∇⋅(∇v⋅∂σ∂∇u−ϕ∂F∂∇u)]⋅dudP+(ϕ˙+∂o∂ξ−∇v⋅∂σ∂ξ+∂F∂ξϕ)dξdP]dΩ dt−∫0T∫∂Ω(∇v⋅∂σ∂∇u−ϕ∂F∂∇u)n⋅dudP dS dt+[∫Ω(−ρv⋅du˙dP+ρv˙⋅dudP−ϕdξdP)dΩ]0T.(8)\begin{aligned} \frac{d\mathcal{O}}{dP}={}&\int_0^T\int_{\Omega}\left[\left(\frac{\partial o}{\partial P}-\frac{\partial\sigma}{\partial P}\cdot\nabla v+\phi\frac{\partial F}{\partial P}\right)+\left[-\rho\ddot{v}+\frac{\partial o}{\partial u}+\nabla\cdot\left(\nabla v\cdot\frac{\partial\sigma}{\partial\nabla u}-\phi\frac{\partial F}{\partial\nabla u}\right)\right]\cdot\frac{du}{dP}\right.\\ &\left.+\left(\dot{\phi}+\frac{\partial o}{\partial\xi}-\nabla v\cdot\frac{\partial\sigma}{\partial\xi}+\frac{\partial F}{\partial\xi}\phi\right)\frac{d\xi}{dP}\right]d\Omega\,dt\\ &-\int_0^T\int_{\partial\Omega}\left(\nabla v\cdot\frac{\partial\sigma}{\partial\nabla u}-\phi\frac{\partial F}{\partial\nabla u}\right)n\cdot\frac{du}{dP}\,dS\,dt\\ &+\left[\int_{\Omega}\left(-\rho v\cdot\frac{d\dot{u}}{dP}+\rho\dot{v}\cdot\frac{du}{dP}-\phi\frac{d\xi}{dP}\right)d\Omega\right]_0^T. \tag*{(8)} \end{aligned}

We eliminate the dependence on dudP\frac{du}{dP} and dξdP\frac{d\xi}{dP} by choosing a particular vv and ϕ\phi such that the respective terms multiplying them become zero. As du˙dP∣t=0=dudP∣t=0=0\left.\frac{d\dot{u}}{dP}\right|_{t=0}=\left.\frac{du}{dP}\right|_{t=0}=0, and dξdP∣t=0=0\left.\frac{d\xi}{dP}\right|_{t=0}=0, we eliminate the contributions from boundary terms by choosing v∣t=T=v˙t=T=0v|_{t=T}=\dot{v}_{t=T}=0 and ϕ∣t=T=0\phi|_{t=T}=0.

We conclude that the expression for the sensitivity is

dOdP=∫0T∫Ω[∂o∂P−∂σ∂P⋅∇v+ϕ∂F∂P]dΩ dt(9)\frac{d\mathcal{O}}{dP}=\int_0^T\int_{\Omega}\left[\frac{\partial o}{\partial P}-\frac{\partial\sigma}{\partial P}\cdot\nabla v+\phi\frac{\partial F}{\partial P}\right]d\Omega\,dt \tag*{(9)}

where the adjoint variables vv and ϕ\phi satisfy the evolution

ρv¨−∇⋅(∇v⋅∂σ∂∇u−ϕ∂F∂∇u)=∂o∂uon Ω,\rho\ddot{v}-\nabla\cdot\left(\nabla v\cdot\frac{\partial\sigma}{\partial\nabla u}-\phi\frac{\partial F}{\partial\nabla u}\right)=\frac{\partial o}{\partial u}\quad\text{on }\Omega,
ϕ˙=∇v⋅∂σ∂ξ−∂F∂ξϕ−∂o∂ξon Ω,\dot{\phi}=\nabla v\cdot\frac{\partial\sigma}{\partial\xi}-\frac{\partial F}{\partial\xi}\phi-\frac{\partial o}{\partial\xi}\quad\text{on }\Omega,
(∇v⋅∂σ∂∇u−ϕ∂F∂∇u)n=0on ∂fΩ,\left(\nabla v\cdot\frac{\partial\sigma}{\partial\nabla u}-\phi\frac{\partial F}{\partial\nabla u}\right)n=0\quad\text{on }\partial_f\Omega,
v=0on ∂dΩ,v=0\quad\text{on }\partial_d\Omega,
v∣t=T=v˙∣t=T=0,ϕ∣t=T=0.(10)v|_{t=T}=\dot{v}|_{t=T}=0,\quad\phi|_{t=T}=0. \tag*{(10)}

We note that the conditions for the adjoint variables are given at the end time t=Tt = T. Thus, the adjoint system should be solved backwards in time starting from t=Tt = T and moving to t=0t = 0.

Putting all of this together, we use the following procedure to solve the optimization problem (3). We start with an initial value for the parameter set PP. We then solve the forward problem (2) for u(t)u(t), ξ(t)\xi(t) followed by the adjoint problem (10) for v(t)v(t), ϕ(t)\phi(t). These are used in (9) to compute the sensitivities. The sensitivities are used to update the parameters and the process is repeated until convergence. The schematic for the algorithm is shown in Figure 1.

Figure 1. Schematic representation of the iterative algorithm to obtain to find material parameters. The forward and adjoint problem are given by (2) and (10) respectively. The sensitivity is given by (9).

Scaling

We now discuss how the numerical cost of the algorithm shown in Figure 1 scales respect to the number of parameters. The forward problem ((2) labeled (F) in the figure) is largely independent of the number of parameters, as is the adjoint problem ((10) labeled (A)). The calculation of the sensitivity ((9) labeled (S)) and the parameter update (labeled (U)) scale linearly, but do not involve the solution of any equations. Thus the overall cost of each iteration is

Cost⁡(NP)=A+BNP,where A≫B(11)\operatorname{Cost}(N_{P}) = A + B N_{P}, \qquad\text{where } A \gg B \tag*{(11)}

where NPN_{P} is the number of parameters. In other words, the cost is linear with a small coefficient. This is in contrast with gradient-free approaches that scale as AαNPA\alpha^{N_{P}} for some α\alpha since NPN_{P} is the dimension of the search space.

J2 Elasto-viscoplasticity

Governing equations

We now specialize to a specific example of a J2 plastic material with isotropic power-law hardening and rate dependence in the small strain setting [41, 34]. We denote the displacement as uu and the strain as ε=(∇u+∇uT)/2\varepsilon= (\nabla u + \nabla u^{T})/2. The plastic strain is εp\varepsilon^{p} and the accumulated plastic strain is

q˙=23ε˙p⋅ε˙p.(12)\dot{q} = \sqrt{\frac{2}{3}\dot{\varepsilon}^{p} \cdot\dot{\varepsilon}^{p}}. \tag*{(12)}

The body is linearly elastic before yield, with the stress and stored elastic energy density are

σ=CεeandWe=12εe⋅Cεe(13)\sigma= \mathbb{C}\varepsilon^{e} \qquad\text{and} \qquad W^{e} = \frac{1}{2}\varepsilon^{e} \cdot\mathbb{C}\varepsilon^{e} \tag*{(13)}

respectively where C\mathbb{C} is the elastic modulus and εe=ε−εp\varepsilon^{e} = \varepsilon- \varepsilon^{p} is the elastic strain. We assume a von Mises yield criterion with yield strength σy\sigma_{y}, power law isotropic hardening with potential

Wp(q)=σy[q+nε0pn+1(qε0p)(n+1)/n],(14)W^{p}(q) = \sigma_{y}\left[q + \frac{n\varepsilon_{0}^{p}}{n+1}\left(\frac{q}{\varepsilon_{0}^{p}}\right)^{(n+1)/n}\right], \tag*{(14)}

with ε0p\varepsilon_{0}^{p} the reference plastic strain and nn the hardening exponent, and isotropic power law rate hardening with dissipation potential

ψ(q˙)={g∗(q˙):=mσyε˙0pm+1(q˙ε˙0p)(m+1)/m,q˙≥0,∞,q˙<0,(15)\psi(\dot{q}) = \begin{cases} g^{*}(\dot{q}) := \dfrac{m\sigma_{y}\dot{\varepsilon}_{0}^{p}}{m+1}\left(\dfrac{\dot{q}}{\dot{\varepsilon}_{0}^{p}}\right)^{(m+1)/m}, & \dot{q} \ge0, \\ \infty, & \dot{q} < 0, \end{cases} \tag*{(15)}

where ε˙0p\dot{\varepsilon}_{0}^{p} is the reference plastic strain rate, and mm is the rate-sensitivity exponent.

The governing equations for the fields {u,q,εp}\{u,q,\varepsilon^{p}\} are

ρu¨=∇⋅Cεeon Ω,0∈σM−∂Wp∂q−∂ψ(q˙)on Ω,ε˙p=q˙Mon Ω,u=u0on ∂uΩ,f=Cεe⋅n^on ∂fΩ,u∣t=0=u˙∣t=0=0,q∣t=0=0,εp∣t=0=0,(16)\begin{aligned} \rho\ddot{u} &= \nabla\cdot\mathbb{C}\varepsilon^{e} && \text{on } \Omega,\\ 0 &\in\sigma_{M} - \frac{\partial W^{p}}{\partial q} - \partial\psi(\dot{q}) && \text{on } \Omega,\\ \dot{\varepsilon}^{p} &= \dot{q}M && \text{on } \Omega,\\ u &= u_{0} && \text{on } \partial_{u}\Omega,\\ f &= \mathbb{C}\varepsilon^{e} \cdot\hat{n} && \text{on } \partial_{f}\Omega,\\ u|_{t=0} &= \dot{u}|_{t=0} = 0,\quad q|_{t=0} = 0,\quad\varepsilon^{p}|_{t=0} = 0, \tag*{(16)} \end{aligned}

where u0u_{0} is the applied displacement on ∂uΩ\partial_{u}\Omega, ff is the applied (dead) traction on ∂uΩ\partial_{u}\Omega, σM=23s⋅s\sigma_{M} = \sqrt{\frac{2}{3}s \cdot s} is the Mises stress, s:=Cεe−(1/N)tr⁡(Cεe)s := \mathbb{C}\varepsilon^{e} - (1/N)\operatorname{tr}(\mathbb{C}\varepsilon^{e}) is the stress deviator, and M:=(1/σM)sM := (1/\sigma_{M})s is the flow direction. We have assumed quiescent initial conditions.

The governing equations for quasistatic evolution can be obtained by ignoring the inertial term in the momentum balance, and the initial conditions for the displacement.

Adjoint method for sensitivities

We seek to find the material parameters by solving the optimization problem (3) with (16) replacing (2) as the constraint for a suitable objective OO. We follow § 2.1.3 to find the sensitivity using the adjoint method; the details are provided in Appendix § 6. The sensitivity is

dOdP=∫0T∫Ωq˙γ(−∂2Wp∂q∂P−∂2g∗∂q˙∂P) dΩ dt,(17)\frac{dO}{dP} = \int_{0}^{T}\int_{\Omega}\dot{q}\gamma\left(-\frac{\partial^{2}W^{p}}{\partial q\partial P}-\frac{\partial^{2}g^{*}}{\partial\dot{q}\partial P}\right)\,d\Omega\,dt, \tag*{(17)}

where the adjoint variables vv, γ\gamma, and ζ\zeta, being dual to the forward variables uu, qq, and εp\varepsilon^p, satisfy the evolution relations

ρv¨−∇⋅(C∇v+γ˙∂σM∂ε−q˙ζ⋅∂M∂ε)=−∂o∂uon Ω\rho\ddot{v} - \nabla\cdot\left(\mathbf{C}\nabla v + \dot{\gamma}\frac{\partial\sigma_M}{\partial\varepsilon} - \dot{q}\zeta\cdot\frac{\partial M}{\partial\varepsilon}\right) = -\frac{\partial o}{\partial u} \qquad\text{on } \Omega
ddt[γ(σM−∂Wp∂q−∂g∗∂q˙)−γ˙q∂2g∗∂q˙2−ζ⋅M]=−γ∂2Wp∂q2on Ω\frac{d}{dt}\left[\gamma\left(\sigma_M-\frac{\partial W^p}{\partial q}-\frac{\partial g^*}{\partial\dot{q}}\right)-\dot{\gamma}q\frac{\partial^2 g^*}{\partial\dot{q}^2}-\zeta\cdot M\right] = -\gamma\frac{\partial^2 W^p}{\partial q^2} \qquad\text{on } \Omega
dζdt=∇ζ⋅∂2We∂ε∂εp+γ˙q∂σM∂εp−q˙ζ⋅∂M∂εpon Ω\frac{d\zeta}{dt} = \nabla\zeta\cdot\frac{\partial^2 W^e}{\partial\varepsilon\partial\varepsilon^p}+\dot{\gamma}q\frac{\partial\sigma_M}{\partial\varepsilon^p}-\dot{q}\zeta\cdot\frac{\partial M}{\partial\varepsilon^p} \qquad\text{on } \Omega
v∣t=T=v˙∣t=T=0,γ∣t=T=0,ζ∣t=T=0.(18)v|_{t=T}=\dot{v}|_{t=T}=0,\quad\gamma|_{t=T}=0,\quad\zeta|_{t=T}=0. \tag*{(18)}

Note that these equations are solved backward in time as before.

We obtain the equation in the quasistatic setting by ignoring the inertial (v¨\ddot{v}) term in the first equation, and the final conditions on v˙\dot{v} in the last.

Numerical implementation

We discretize the equations in space using standard P=1P=1 Lagrange polynomial finite elements for the displacement uu and the corresponding adjoint variable vv. The plastic quantities qq and εp\varepsilon^p are spatially discretized at quadrature points. In dynamics, we adopt a mixed scheme for the temporal discretization. For the governing equations (16), we use an explicit central difference update for the displacement, but an implicit backwards Euler scheme for the plastic updates. Similarly, for the adjoint equations (18) we use an explicit central difference method for the adjoint displacement vv, and implicit backwards Euler scheme for the plastic adjoints γ,ζ\gamma,\zeta. In quasistatic evolution, we use an implicit backward Euler scheme for temporal discretization of all variables. Further details are provided in Appendix § 7.

We solve the optimization problem iteratively starting from an arbitrary initial guess. We update the parameter set using the gradient-based Method of Moving Asymptotes optimization scheme [52]. This technique for solving constrained optimization problems has been used extensively in the optimal design community to solve PDE-constrained design problems over both small [3] and large parameter sets [43, 2]. We allow the optimization to run for a set number of iterations, with this number chosen large enough such that it sufficiently converges. We implement the numerical method in the deal.ii Finite Element Library [5, 4].

Demonstration and validation

We now demonstrate and validate the method described above using two examples with synthetic data. The first example is a quasistatic compression test on a plane-strain specimen while the second is an extended dynamic impact test. In both cases, we generate synthetic data using numerical simulations of a material model with properties similar to that of copper. Specifically, we take the shear modulus to be μ=46.7\mu=46.7 GPa, a Poisson ratio of ν=0.3656\nu=0.3656, and the density to be ρ=8.93×103\rho=8.93 \times10^3 kg/m3^3, with the other parameters P={σy,ε0p,n,ε˙0p,m}P=\{\sigma_y,\varepsilon_0^p,n,\dot{\varepsilon}_0^p,m\} as shown in the first row (PgenP^{\mathrm{gen}}) of Table 1.1 We use these parameters, and the forward problem described in § 2.2.1 above, to simulate the respective experiments to generate the data DexpD^{\mathrm{exp}}. We then use this data to obtain the material parameters using the indirect inversion method described in § 2.2. We initialize these calculations with an initial guess PinitP^{\mathrm{init}} that are significantly different from the ground truth PgenP^{\mathrm{gen}}. The PgenP^{\mathrm{gen}}, PinitP^{\mathrm{init}} and recovered parameters for both experiments are summarized in Table 1.

σy/μ\sigma_y/\muε0p\varepsilon_0^pnnε˙0p\dot{\varepsilon}_0^p (/s/s)mmO\mathcal{O}Oind\mathcal{O}_{\mathrm{ind}}
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00––
PinitP^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.00––
PQSP^{\mathrm{QS}}0.0021270.039323.054.75×1054.75 \times10^51.885.59×10−95.59 \times10^{-9} (2.74×10−4)(2.74 \times10^{-4})0.0069 (1.452)
PDCP^{\mathrm{DC}}0.0017340.028592.874.05×1054.05 \times10^54.501.44×10−41.44 \times10^{-4} (1.81×10−1)(1.81 \times10^{-1})0.021 (1.452)

Table 1. Summary of results. The parameters used to generate the synthetic data, the initial guess and ones obtained through the proposed method in the quasistatic tension (QS) and dynamic compression(DC) experiments. The final objectives are reported in the table with the initial objectives in the parenthesis.

Quasistatic compression test of a plane strain specimen

The first example is a thick plate with an off-center hole subjected to quasistatic compression shown in Figure 2(a). We consider a geometry with L/H=0.8L/H = 0.8, R/H=0.1R/H = 0.1, Loff/H=0.1L_{\mathrm{off}}/H = 0.1, and Hoff/H=0.04H_{\mathrm{off}}/H = 0.04. We assume plane strain conditions. We impose uniform vertical displacement

Figure 2. Results for the quasistatic compression of a thick plate with offset hole. (a) Geometry and boundary conditions. (b) Normalized objective from 20 vs iteration number. (c) Results of an independent uniaxial stress-strain test. (d) Deformed configurations at the final time-step for the ground truth data with the accumulated plasticity qq.

u0(t)=εˉ˙Ht,(19)u_0(t) = \dot{\bar{\varepsilon}}Ht, \tag*{(19)}

with a nominal or macroscopic strain rate εˉ˙\dot{\bar{\varepsilon}} on the top boundary, T\mathcal{T}, while the botton surface is held fixed. The experimental data DexpD^{\mathrm{exp}} consists of the full-field displacement uexp(x,T)=u(x,T)u^{\mathrm{exp}}(x,T) = u(x,T) at the final time TT, and the reaction load history fRexp(t):=∫ΓCσεn dΓf_R^{\mathrm{exp}}(t) := \int_{\Gamma_C} \boldsymbol{\sigma}\boldsymbol{\varepsilon}_n \,\mathrm{d}\Gamma. We emphasize that we only use the full-field displacement at the final time step2. We further assume that we have data from nn tests with varying rates. We define our objective to be

O(P):=∑i=1n(αuL4∫Ω∣ui(x,T)−uexp,i(x,T)∣2 dΩ+αfTiμ2L2∫0Ti∥fRi(t)−fRexp,i∥2 dt)(20)\mathcal{O}(P) := \sum_{i=1}^{n} \left( \frac{\alpha_u}{L^4} \int_{\Omega} \left|u^i(x,T) - u^{\mathrm{exp},i}(x,T)\right|^2 \,\mathrm{d}\Omega+ \frac{\alpha_f}{T_i\mu^2L^2} \int_0^{T_i} \left\|f_R^i(t) - f_R^{\mathrm{exp},i}\right\|^2 \,\mathrm{d}t \right) \tag*{(20)}

where {ui,qi,εp,i}\{u^i, q^i, \varepsilon^{p,i}\} satisfy the governing equation for each i=1,…,ni = 1,\ldots,n

where the superscript indexes each of the tests, and αu\alpha_u and αf\alpha_f are scaling factors that balance the weights of the two objective terms.

We consider a data-set from by three tests (n=3n = 3) on the same geometry and material, but differing in the loading rate. The loading rates span two orders of magnitude, with macroscopic strain rates of 10210^2, 10310^3, and 10410^4/s, reaching a final macroscopic strain of εˉ˙=0.1\dot{\bar{\varepsilon}} = 0.1 over 100 timesteps. The data-set is generated synthetically from forward simulations using the parameters PgenP^{\mathrm{gen}} in Table 1. We then use the proposed approach with the objective (20). Full numerical details of the forward and adjoint problem can be found in Appendix B.1.1 and Appendix B.2.1

The results are shown in Figure 2(b-Figure 2d) and the recovered parameters are listed in Table 1. Figure 2(b) shows the change in the objective as we iterate. It drops rapidly but then gradually stabilizes after a little over 100 iterations decreasing by a roughly a factor 10510^{5}. The values of the recovered parameters after 300 iterations is listed as PQSP^{\mathrm{QS}} in the Table 1. We recover the normalized yield strength σy/μ\sigma_y/\mu to within 10%, but the strain and rate hardening parameters differ significantly. This is true despite the fact that our objective is extremely small, with an objective value on the order of 10−910^{-9}. This means that the experiment with three tests at three strain rates are unable to distinguish between the two sets of parameters, the parameters PgenP^{\mathrm{gen}} used to generate the data and the recovered parameters PrecP^{\mathrm{rec}}. This is further demonstrated in Figure 2(d) which compares the accumulated plastic strain computed using the two sets of parameters, PgenP^{\mathrm{gen}} and PQSP^{\mathrm{QS}}.

Independent objective We found that the parameters we recover PQSP^{\mathrm{QS}} is different from those used to generate the data PgenP^{\mathrm{gen}} even though the difference in the measured quantities (objective) are very small. Therefore, we consider a zero-shot test where we evaluate our results against an independent objective that is not used in the inverse problem. We simulate the material response with the two sets of parameters with an independent uniaxial tension test. The resulting stressstrain curves are shown in Figure 2(c), and they agree well. Thus, even the independent uniaxial stress tension test is also unable to distinguish between the two sets of parameters.

To assess this quantitatively, we define a independent objective to be the average relative root mean square error of the material response for tests performed at nn different strain rates:

Oind=1n∑i=1n(∫(σrec(ε)−σgen(ε))2 dε∫σgen2 dε)1/2.(21)\mathcal{O}_{\mathrm{ind}}=\frac{1}{n}\sum_{i=1}^{n}\left(\frac{\int\left(\sigma_{\mathrm{rec}}(\varepsilon)-\sigma_{\mathrm{gen}}(\varepsilon)\right)^2\,\mathrm{d}\varepsilon}{\int\sigma_{\mathrm{gen}}^2\,\mathrm{d}\varepsilon}\right)^{1/2}. \tag*{(21)}

where σgen\sigma_{\mathrm{gen}} is the uniaxial stress computed using PgenP^{\mathrm{gen}} while σrec\sigma_{\mathrm{rec}} is the uniaxial stress-strain curve generated using the recovered parameters. The value of Oind\mathcal{O}_{\mathrm{ind}} Table 1 confirms the results in Figure 2(c) that the two sets of parameters can not be distinguished in independent uniaxial stress-strain curves.

We conclude this section by studying the robustness of the method to the initial guess, objective and mesh size.

Sensitivity to initial guess We consider three tests with different and quite distinct initialization sets {P1init,P2init,P3init}\{P_1^{\mathrm{init}}, P_2^{\mathrm{init}}, P_3^{\mathrm{init}}\}. These are shown, along with their corresponding recovered parameter sets in Table 2. We observe that the method is insensitive to the initial parameter guess, yielding only slight variations in the recovered parameters with the objective remaining on the order of 10−910^{-9} for all of the recovered parameter values.

σy/μ\sigma_y/\muε0P\varepsilon_0^Pnnε˙0P\dot{\varepsilon}_0^P (/s)mmO\mathcal{O}Oind\mathcal{O}_{\mathrm{ind}}
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00
Sensitivity to initial guess
P1initP_1^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.002.74×10−42.74 \times10^{-4}1.452
P1recP_1^{\mathrm{rec}}0.0021270.039323.054.75×1054.75 \times10^51.885.59×10−95.59 \times10^{-9}0.0069
P2initP_2^{\mathrm{init}}0.010.12.02.0×1052.0 \times10^56.06.43×10−46.43 \times10^{-4}3.791
P2recP_2^{\mathrm{rec}}0.0020640.033523.098.16×1058.16 \times10^52.232.81×10−92.81 \times10^{-9}0.0060
P3initP_3^{\mathrm{init}}0.00750.0754.04.0×1054.0 \times10^51.03.49×10−43.49 \times10^{-4}1.91
P3recP_3^{\mathrm{rec}}0.0021420.040393.053.62×1053.62 \times10^51.717.37×10−97.37 \times10^{-9}0.0128
Sensitivity to objective
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00––
P1initP_1^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.00–1.452
αf/αu=0.1\alpha_f/\alpha_u=0.10.0020810.034563.085.34×1055.34 \times10^51.964.44×10−94.44 \times10^{-9}0.0049
αf/αu=1\alpha_f/\alpha_u=10.0021270.039323.054.75×1054.75 \times10^51.885.59×10−95.59 \times10^{-9}0.0069
αf/αu=10\alpha_f/\alpha_u=100.0021320.040093.044.98×1054.98 \times10^51.872.44×10−92.44 \times10^{-9}0.0048
Sensitivity to meshsize
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00––
PinitP^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.00–1.452
h=h0h=h_00.0020360.032333.086.81×1056.81 \times10^52.269.94×10−109.94 \times10^{-10}0.0031
h=0.5h0h=0.5h_00.0021270.039323.054.75×1054.75 \times10^51.885.59×10−95.59 \times10^{-9}0.0060
h=0.25h0h=0.25h_00.0021500.044103.004.56×1054.56 \times10^51.957.50×10−97.50 \times10^{-9}0.0089
h=0.125h0h=0.125h_00.0021290.044103.004.47×1054.47 \times10^52.188.44×10−98.44 \times10^{-9}0.0090

Table 2. Robustness of the proposed method for the quasistatic experiment.

Sensitivity to objective The parameters αu\alpha_u and αf\alpha_f in the objective function represent the weights for the displacement data and the force data, respectively. Table 2 shows the recovered set of parameters for three different orders of αf/αu\alpha_f/\alpha_u, with all of them showing similar performance.

Sensitivity to mesh size To confirm the convergence of solution with respect to the mesh size, we perform the optimization on different mesh refinement levels. Table 2 shows the recovered parameters for four mesh resolutions ranging from 384 to 24,576 elements. As the recovered parameters only differ slightly between the four meshes, we conclude that the solution does not depend on mesh resolution.

We conclude that the method accurately recovers elasto-viscoplastic material parameters from data obtained from the final snapshot and force-history of three quasistatic tests. While the values of the recovered and generate parameters differ slightly, these parameters are unable to distinguish either observations of the test, or the zero-shot independent test, and reflects the degeneracy of the constitutive model.

Dynamic compression of thin specimens

The second example we consider is that of dynamic impact of a cylinder. We consider a thin cylindrical specimen that Ω:=Ω‾×(0,h)\Omega:= \overline{\Omega} \times(0,h) with cross-section Ω‾∈R2\overline{\Omega} \in\mathbb{R}^{2} and thickness of h∈Rh \in\mathbb{R}. Figure 3(a) shows the case where Ω‾\overline{\Omega} is a circular annulus. The cylinder is placed on an anvil and impacted on the top with a striker whose cross section is larger than that of the cylinder. Ignoring an irrelevant rigid translation, this corresponds to the following imposed boundary conditions:

Results for the dynamic compression of circular annulus

Figure 3. Results for the dynamic compression of circular annulus. (a) Geometry and deformation. (b) Applied stretch along the annulus thickness (c) Normalized objective from 24 vs iteration number. (d) The net vertical force applied to the annulus over time. (e) Comparison of accumulated plasticity qq. (f) Results of an independent uniaxial stress-strain test.

u3(X1,X2,0,t)=0,u3(X1,X2,h,t)=(λ(t)−1)h(22)u_{3}(X_{1},X_{2},0,t)=0,\qquad u_{3}(X_{1},X_{2},h,t)=(\lambda(t)-1)h \tag*{(22)}

where λ(t)\lambda(t) is the imposed nominal axial stretch. We assume that the contact with the anvil and the striker is friction-free, and that the specimen is sufficiently small so that we may assume uniform axial strain,

u(X1,X2,X3,t)=uˉ(X1,X2,t)+(λ(t)−1)X3 e3,(23)u(X_{1},X_{2},X_{3},t)=\bar{u}(X_{1},X_{2},t)+(\lambda(t)-1)X_{3}\,e_{3}, \tag*{(23)}

where uˉ:=Ω‾↦R2\bar{u}:=\overline{\Omega}\mapsto\mathbb{R}^{2} is the in-plane displacement. This enables us to reduce this to two space dimensions, see Appendix § 8 for details.

Our experimental data Dexp⁡D^{\exp} consists of the final in plane displacement uˉexp⁡(X1,X2,T)\bar{u}^{\exp}(X_{1},X_{2},T) and the net axial force history fRexp⁡(t)f_{R}^{\exp}(t). Notice that this is more data than is typically collected in a classical dynamic impact experiment, where only fRexp⁡f_{R}^{\exp} is measured. However, we may measure uˉexp⁡(X1,X2,T)\bar{u}^{\exp}(X_{1},X_{2},T) by placing a speckle pattern on the face of the specimen, imaging it before and after the impact, and using digital image correlation. We repeat the test nn times, and consider the following objective

O(P):=∑i=1n(αuL4∫Ω‾∣uˉi(X1,X2,T)−uˉexp⁡,i(X1,X2,T)∣2 dΩˉ+αfTiμ2L4∫0T∣fui−fRexp⁡,i∣2 dT)(24)\mathcal{O}(\mathbf{P}) := \sum_{i=1}^{n}\left(\frac{\alpha_{u}}{L^{4}}\int_{\overline{\Omega}}\left|\bar{u}^{i}(X_{1},X_{2},T)-\bar{u}^{\exp,i}(X_{1},X_{2},T)\right|^{2}\,\mathrm{d}\bar{\Omega}+\frac{\alpha_{f}}{T_{i}\mu^{2}L^{4}}\int_{0}^{T}\left|f_{u}^{i}-f_{R}^{\exp,i}\right|^{2}\,\mathrm{d}T\right) \tag*{(24)}

where {uˉi,qi,εpi}\{\bar{u}^{i},q^{i},\varepsilon_{p}^{i}\} satisfy (50) for i=1,…,ni=1,\ldots,n,

and fuexp⁡,i:=∫Ω‾(Cεˉe,i)33 dSf_{u}^{\exp,i}:=\int_{\overline{\Omega}}(\mathbb{C}\bar{\varepsilon}^{e,i})_{33}\,\mathrm{d}S is the net vertical force. αu\alpha_{u} and αf\alpha_{f} are weights associated with the displacement and force components of the objective. LL, μ\mu and TiT_{i} are characteristic length, shear modulus, and characteristic time respectively.

We apply our method to a specimen whose cross-section is a circular annulus shown in Figure 3(a). We generate synthetic data for n=3n=3 tests imposed thickness strains shown in Figure 3(b) with the initial strain rate in the range ε˙=102−104 s−1\dot{\varepsilon}=10^{2}-10^{4}\,\mathrm{s}^{-1} with the parameters Pgen\mathbf{P}^{\mathrm{gen}}. We initialize our iterative optimization approach with the parameters Pinit\mathbf{P}^{\mathrm{init}}. αu\alpha_{u} and αf\alpha_{f} are chosen such that the contribution of both parts of the objective are approximately equal, αf/αu=1\alpha_{f}/\alpha_{u}=1. Figure 3(c) shows the evolution of the objective with iteration. The objective decreases rapidly at first and then stabilizes at a value with a factor greater than 10310^{3} than the initial value. The resulting parameters are shown in Table 1 as PDCP^{\mathrm{DC}}. As in the previous quasistatic compression test, there is good agreement with the yield strength, but not for the hardening parameters. Despite the fact that the objective is very small, the experimental observations agree well, as shown in Figure 3(d,Figure 3e). Figure 3(d) compares the experimental reaction force with the one simulated with the recovered parameters, and we observe excellent agreement. Figure 3(e) compares the accumulated plastic strain computed with the original parameters PgenP^{\mathrm{gen}} with those computed with the recovered parameters PDCP^{\mathrm{DC}}. Again, we see good agreement. Finally, we conduct the zero-shot test of comparing the response in uniaxial tensile tests. Figure 3(f) shows the results of the independent stress-strain test, and we again see very good agreement.

Finally, we demonstrate the robustness of our method.

Sensitivity to initial guess We consider three tests with different and quite distinct initialization sets {P1init,P2init,P3init}\{P_1^{\mathrm{init}}, P_2^{\mathrm{init}}, P_3^{\mathrm{init}}\}. These are shown, along with their corresponding recovered parameter sets in Table 3. We observe that the method is insensitive to the initial parameter guess, yielding only slight variations in the recovered parameters with the objective remaining on the order of 10−910^{-9} for all of the recovered parameter values.

σy/μ\sigma_y/\muε0p\varepsilon_0^pnnε˙0p\dot{\varepsilon}_0^p (/s)mmO\mathcal{O}Oind\mathcal{O}_{\mathrm{ind}}
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00
Sensitivity to initial guess
P1initP_1^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.001.81×10−11.81 \times10^{-1}1.452
P1recP_1^{\mathrm{rec}}0.0017340.028592.874.05×1054.05 \times10^54.501.44×10−41.44 \times10^{-4}0.021
P2initP_2^{\mathrm{init}}0.010.12.02.0×1052.0 \times10^56.05.55×10−15.55 \times10^{-1}3.791
P2recP_2^{\mathrm{rec}}0.0017930.066232.112.40×1052.40 \times10^54.508.57×10−48.57 \times10^{-4}0.079
P3initP_3^{\mathrm{init}}0.00750.0754.04.0×1054.0 \times10^51.02.12×10−12.12 \times10^{-1}1.91
P3recP_3^{\mathrm{rec}}0.00220.078122.706.02×1056.02 \times10^51.092.93×10−42.93 \times10^{-4}0.090
Sensitivity to objective
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00––
P1initP_1^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.001.81×10−11.81 \times10^{-1}1.452
αf/αu=0.1\alpha_f/\alpha_u = 0.10.0017260.035832.713.46×1053.46 \times10^54.581.47×10−41.47 \times10^{-4}0.050
αf/αu=1\alpha_f/\alpha_u = 10.0017340.028592.874.05×1054.05 \times10^54.501.44×10−41.44 \times10^{-4}0.021
αf/αu=10\alpha_f/\alpha_u = 100.0017780.039292.693.51×1053.51 \times10^54.810.87×10−40.87 \times10^{-4}0.043
Sensitivity to geometry
PgenP^{\mathrm{gen}}0.0019350.022453.235.00×1055.00 \times10^52.00––
P1initP_1^{\mathrm{init}}0.0050000.050005.003.00×1053.00 \times10^55.001.81×10−11.81 \times10^{-1}1.452
Circular0.0017340.028592.874.05×1054.05 \times10^54.501.44×10−41.44 \times10^{-4}0.021
Ellipse0.0019210.045902.553.10×1053.10 \times10^53.141.81×10−51.81 \times10^{-5}0.020
Flower hole0.0017310.035062.733.71×1053.71 \times10^54.261.32×10−41.32 \times10^{-4}0.021
Flower0.0018000.027682.933.93×1053.93 \times10^53.048.49×10−58.49 \times10^{-5}0.023

Table 3. Robustness of the proposed method for the dynamic compression experiment.

Sensitivity to objective The parameters αu\alpha_u and αf\alpha_f in the objective function represent the weights for the displacement data and the force data respectively. Table 3 shows the recovered set of parameters for three different orders of αf/αu\alpha_f/\alpha_u, with all of them showing similar performance.

Sensitivity to mesh size To confirm the convergence of solution with respect to the mesh size, we perform the optimization for three different mesh sizes. Figure 4 shows the objective versus optimization iterations for different mesh sizes. Since the objective remains the same for all mesh sizes, we conclude that it is safe to assume the solution does not depend on the size of the mesh.

Figure 4. Objective curves with respect to iterations for different mesh sizes.

Sensitivity to configuration The geometry of the specimen can significantly impact the inversion methodology. Certain geometries can induce a wide range of strains during compression, providing datasets with richer information about the material properties. To investigate this, we performed the inversion on different cross-sectional geometries, specifically circular, elliptical, and flower-shaped configurations. The results for the elliptic annulus, disc with flower-shaped hole and flower-shaped specimen are shown in Figure 5, Figure 6 and Figure 7, respectively. The recovered parameters for these configurations are compared in Table 3. Due to the geometry of the ellipse configuration, the average forces experienced is much higher than the other two configurations, leading to more efficient recovery of parameters and a lower overall objective value. Furthermore, the lack of angular symmetry of the strain field in ellipse provides a richer dataset resulting in a more efficient inversion.

Results for the dynamic compression of elliptic annulus

Figure 5. Results for the dynamic compression of elliptic annulus. (a) Normalized objective from 242^4 vs iteration number. (b) The net vertical force applied to the annulus over time. (c) Comparison of accumulated plasticity qq. (d) Results of an independent uniaxial stress-strain test.

Results for the dynamic compression of flower-shaped specimen

Figure 6. Results for the dynamic compression of flower-shaped specimen . (a) Normalized objective from 24 vs iteration number. (b) The net vertical force applied to the annulus over time. (c) Comparison of accumulated plasticity qq. (d) Results of an independent uniaxial stress-strain test.

Figure 7. Results for the dynamic compression of disc with flower-shaped hole. (a) Normalized objective from 24 vs iteration number. (b) The net vertical force applied to the annulus over time. (c) Comparison of accumulated plasticity qq. (d) Results of an independent uniaxial stress-strain test.

Conclusion

Material characterization through a constitutive model is necessary to close the balance laws and allow for continuum modeling. Traditional approaches rely on experimental configurations that yield uniform states of stress and strain (rate), and can therefore be easily inverted to characterize the constitutive law. However, these approaches only probe highly selected and idealized states and trajectories, thereby leading to large uncertainties in the complex states and trajectories that arise in the applications of the resulting models. Further, some of these configurations require careful alignment to achieve the idealized states, further limiting the amount of data that can be collected. In this work, we have introduced an alternate approach that uses a combination of PDE-constrained optimization and full-field observation techniques to obtain material models from complex experiments.

We formulate the problem of learning the constitutive model as an indirect inverse problem. We assume that the model is given in a parametrized form: this can be either classical where one has a few parameters or a hyperparametrized formulation like a neural network. We recast this inverse problem as a PDE-constrained optimization where we seek to find the parameters that minimize the difference between experimentally measured quantities, and the corresponding quantities computed using the parameters subject to the balance laws (PDE-constraint). We follow an iterative gradient-based approach where we use the adjoint method to compute the sensitivities. This results in an adjoint partial differential equation (linear in space and quasilinear in time) that is no more expensive to solve than the forward problem. The resulting algorithm is shown in Figure 1. As noted in the introduction, our work builds on prior work in the literature on the finite-element updating and PDE-constrained optimization.

In this paper, we describe the formulation, first for a general history dependent material and then for the specific model of a J2J_2 elasto-viscoplasticity. We then demonstrate it with synthetic data in two experimental configurations: quasistatic compression of a thick plate with a hole, and extended dynamic compression. In forthcoming work, we demonstrate the method against experimental data (part 2).

We point out a number of salient features of this approach. First, the method scales linearly (see § 2.1.4) with the number of parameters in a model with a small pre-factor to the linear term. Therefore, this method is ideally suited for complex models with a large number of parameters including a hyperparametrized formulation like a neural network. We demonstrate this in forthcoming work (part 3).

Second, the method is able to recover complex behavior with a relatively little data. In the quasistatic test described in § 3.1, we only used the strain field at the final time instead of a series of snapshots. This is because this test used a complex geometry and full-field information, and thus probed multiple trajectories through the adjoint method. At the same time, the computational cost of the method is relatively independent of the amount of data: so we could have easily incorporated multiple snapshots.

Third, the approach is versatile. We have demonstrated this by applying it to two different configurations. Further, the approach is capable of addressing non-smooth problems. We applied the method here to elasto-viscoplasticity including a yield criterion. This is difficult to do with a method that uses a numerical derivative. In forthcoming work (part 2), we extend the approach to indentation including contact.

Finally, the formulation does not make any assumptions about isotropy or homogeneity. Our current work addresses learning single crystal behavior from observations on polycrystals.

Acknowledgement

We are delighted to acknowledge many useful discussions with Professors Ravi Ravichandran and Andrew Stuart. We are grateful for the financial support of the Army Research Laboratory (W911NF22-2-0120) and the Army Research Office (W911NF-22-1-0269).

Adjoint problem in elasto-viscoplascitity

We consider an objective of integral from

O(u,q,εp,P)=∫0T∫Ωo(u,q,εp,P) dΩ.(25)\mathcal{O}(u,q,\varepsilon^{p},P)=\int_{0}^{T}\int_{\Omega}o(u,q,\varepsilon^{p},P)\,d\Omega. \tag*{(25)}

We follow § 2.1.3 to use the adjoint method to compute the sensitivity of O\mathcal{O} with respect to the parameters PP. So we augment to objective using the governing equations

O=∫0T∫Ω{o+ρu¨⋅v+∂We∂ε⋅∇v−b⋅v+γ˙[σ‾M−∂Wp∂q−∂g∗∂q˙]+ζ⋅(ε˙p−q˙M)} dΩdt+∫0T∫∂Ω(f⋅v) dS dt.(26)\begin{aligned} \mathcal{O}={}&\int_{0}^{T}\int_{\Omega}\left\{o+\rho\ddot{u}\cdot v+\frac{\partial W^{e}}{\partial\varepsilon}\cdot\nabla v-b\cdot v+\dot{\gamma}\left[\overline{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial g^{*}}{\partial\dot{q}}\right]+\zeta\cdot\left(\dot{\varepsilon}^{p}-\dot{q}M\right)\right\}\,d\Omega dt\\ &+\int_{0}^{T}\int_{\partial\Omega}(f\cdot v)\,dS\,dt. \tag*{(26)} \end{aligned}

where the fields vv, γ\gamma and ζ\zeta which correspond to the displacement, plastic hardening and plastic strain, respectively are to be determined. Also, we have used the Kuhn-Tucker condition to replace (16)2(16)_{2} to include the irreversibility of the accumulated plastic strain. We differentiate this augmented objective with respect to the parameters PP,

dOdP=∫0T∫Ω{∂o∂P+∂ρ∂Pu¨⋅v+∂2We∂ε∂P⋅∇v+γ˙(∂σ‾M∂P−∂2Wp∂q∂P−∂2g∗∂q˙∂P)+∂o∂uδPu+ρv⋅δPu¨+(∇v⋅∂2We∂ε∂ε+γ˙∂σ‾M∂ε−ζ˙⋅∂M∂ε)⋅∇δPu+(∂o∂q−γ˙∂2Wp∂q2)δPq+(−γ˙∂2g∗∂q˙2+γ[σ‾M−∂Wp∂q−∂g∗∂q˙]−ζ⋅M)δPq˙+ζ⋅δPε˙p+(∂o∂εp+∇v⋅∂2We∂ε∂εp+γ˙∂σ‾M∂εp−ζ˙⋅∂M∂εp)⋅δPεp} dΩ dt.(27)\begin{aligned} \frac{d\mathcal{O}}{dP}={}&\int_{0}^{T}\int_{\Omega}\left\{\frac{\partial o}{\partial P}+\frac{\partial\rho}{\partial P}\ddot{u}\cdot v+\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial P}\cdot\nabla v+\dot{\gamma}\left(\frac{\partial\overline{\sigma}_{M}}{\partial P}-\frac{\partial^{2}W^{p}}{\partial q\partial P}-\frac{\partial^{2}g^{*}}{\partial\dot{q}\partial P}\right)\right.\\ &+\frac{\partial o}{\partial u}\delta_{P}u+\rho v\cdot\delta_{P}\ddot{u}+\left(\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon}+\dot{\gamma}\frac{\partial\overline{\sigma}_{M}}{\partial\varepsilon}-\dot{\zeta}\cdot\frac{\partial M}{\partial\varepsilon}\right)\cdot\nabla\delta_{P}u\\ &+\left(\frac{\partial o}{\partial q}-\dot{\gamma}\frac{\partial^{2}W^{p}}{\partial q^{2}}\right)\delta_{P}q+\left(-\dot{\gamma}\frac{\partial^{2}g^{*}}{\partial\dot{q}^{2}}+\gamma\left[\overline{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial g^{*}}{\partial\dot{q}}\right]-\zeta\cdot M\right)\delta_{P}\dot{q}\\ &+\zeta\cdot\delta_{P}\dot{\varepsilon}^{p}\\ &\left.+\left(\frac{\partial o}{\partial\varepsilon^{p}}+\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon^{p}}+\dot{\gamma}\frac{\partial\overline{\sigma}_{M}}{\partial\varepsilon^{p}}-\dot{\zeta}\cdot\frac{\partial M}{\partial\varepsilon^{p}}\right)\cdot\delta_{P}\varepsilon^{p}\right\}\,d\Omega\,dt. \tag*{(27)} \end{aligned}

We then integrate by parts, and enforce quiescent conditions on vv at the final time TT to remove the boundary terms. Then, localizing gives the sensitivities as

dOdP=∫0T∫Ω{∂o∂P+∂ρ∂Pu¨⋅v+∂2We∂ε∂P⋅∇v+γ˙(∂σ‾M∂P−∂2Wp∂q∂P−∂2g∗∂q˙∂P)} dΩ dt,(28)\frac{d\mathcal{O}}{dP}=\int_{0}^{T}\int_{\Omega}\left\{\frac{\partial o}{\partial P}+\frac{\partial\rho}{\partial P}\ddot{u}\cdot v+\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial P}\cdot\nabla v+\dot{\gamma}\left(\frac{\partial\overline{\sigma}_{M}}{\partial P}-\frac{\partial^{2}W^{p}}{\partial q\partial P}-\frac{\partial^{2}g^{*}}{\partial\dot{q}\partial P}\right)\right\}\,d\Omega\,dt, \tag*{(28)}

if the adjoint variables satisfy the evolution

0=∫Ω[ρv¨⋅δPu+∂o∂u⋅δPu+(∇v⋅∂2We∂ε∂ε+γ˙∂σ‾M∂ε−ζ˙⋅∂M∂ε)⋅∇δPu] dΩ∀δPu∈K0,0=\int_{\Omega}\left[\rho\ddot{v}\cdot\delta_{P}u+\frac{\partial o}{\partial u}\cdot\delta_{P}u+\left(\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon}+\dot{\gamma}\frac{\partial\overline{\sigma}_{M}}{\partial\varepsilon}-\dot{\zeta}\cdot\frac{\partial M}{\partial\varepsilon}\right)\cdot\nabla\delta_{P}u\right]\,d\Omega\quad\forall\delta_{P}u\in\mathcal{K}_{0},
ddt[γ(σ‾M−∂Wp∂q−∂g∗∂q˙)−γ˙∂2g∗∂q˙2−ζ⋅M]=∂o∂q−γ˙∂2Wp∂q2on Ω,\frac{d}{dt}\left[\gamma\left(\overline{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial g^{*}}{\partial\dot{q}}\right)-\dot{\gamma}\frac{\partial^{2}g^{*}}{\partial\dot{q}^{2}}-\zeta\cdot M\right]=\frac{\partial o}{\partial q}-\dot{\gamma}\frac{\partial^{2}W^{p}}{\partial q^{2}}\qquad\text{on }\Omega,
dζdt=∂o∂εp+∇v⋅∂2We∂ε∂εp+γ˙∂σ‾M∂εp−ζ˙⋅∂M∂εpon Ω,\frac{d\zeta}{dt}=\frac{\partial o}{\partial\varepsilon^{p}}+\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon^{p}}+\dot{\gamma}\frac{\partial\overline{\sigma}_{M}}{\partial\varepsilon^{p}}-\dot{\zeta}\cdot\frac{\partial M}{\partial\varepsilon^{p}}\qquad\text{on }\Omega,
v∣t=T=0,v˙∣t=T=0,γ∣t=T=0,ζ∣t=T=0.(29)v|_{t=T}=0,\qquad\dot{v}|_{t=T}=0,\qquad\gamma|_{t=T}=0,\qquad\zeta|_{t=T}=0. \tag*{(29)}

where K0:={φ∈H1(Ω), φ=0 on ∂uΩ}\mathcal{K}_{0}:=\{\varphi\in H^{1}(\Omega),\ \varphi=0\text{ on }\partial_{u}\Omega\} is the space of kinematically admissible displacement variations. This is the weak form of (18).

Numerical method in elasto-viscoplasticity

We discuss the finite element discretization and numerical method to solve the forward and adjoint problems for both the quasistatic and dynamic setting. We consider a spatial discretization with standard p=1p = 1 Lagrange polynomial shape functions for the displacements

u=∑i=1nuuiNi(x),(30)u = \sum_{i=1}^{n_u} u_i N_i(x), \tag*{(30)}

where Ni:Ω↦RnN_i : \Omega\mapsto\mathbb{R}^n are the standard vector valued shape functions with compact support. The fields qq and εp\varepsilon^p are discretized at quadrature points as

q(xg)=qg,εp(xg)=εgp,(31)q(x_g) = q_g,\qquad\varepsilon^p(x_g) = \varepsilon^p_g, \tag*{(31)}

for some Gauss point xgx_g, g=1,…,ngg = 1,\ldots,n_g.

Forward Problem

Quasistatic

The governing relations in this setting are

0=∫Ω[Cεe⋅∇δu−b⋅δu] dΩ−∫∂fΩf⋅δu dS∀δu∈K0,0∈σM−∂Wp∂q−∂ψ(q˙)on Ω,ε˙p=q˙Mon Ω,q∣t=0=0,εp∣t=0=0(32)\begin{aligned} 0 &= \int_{\Omega} [\mathbb{C}\varepsilon^e \cdot\nabla\delta u - b \cdot\delta u]\,\mathrm{d}\Omega- \int_{\partial_f\Omega} f \cdot\delta u\,\mathrm{d}S && \forall\delta u \in\mathcal{K}_0,\\ 0 &\in\sigma_M - \frac{\partial W^p}{\partial q} - \partial\psi(\dot q) && \text{on }\Omega,\\ \dot{\varepsilon}^p &= \dot q M && \text{on }\Omega,\\ q|_{t=0} &= 0,\quad\varepsilon^p|_{t=0} = 0 \tag*{(32)} \end{aligned}

where K0:={φ∈H1(Ω), φ=0 on ∂uΩ}\mathcal{K}_0 := \{\varphi\in H^1(\Omega),\ \varphi= 0\text{ on }\partial_u\Omega\} is the space of kinematically admissible displacement variations.

We consider a fully implicit approach, with a backwards Euler approximation of the temporal derivatives of the plastic variables. We examine this from the nn to n+1n+1 time step. That is, we look to solve for {un+1,qn+1,εp,n+1}\{u^{n+1},q^{n+1},\varepsilon^{p,n+1}\} given {un,qn,εp,n}\{u^n,q^n,\varepsilon^{p,n}\} assuming the discretizations in (30) and (31). Thus, we look to solve

0=Ri:=∫Ω[C(ε(un+1)−εp,n+1)⋅∇Ni−b⋅Ni] dΩ−∫∂fΩf⋅Ni dSi=1,…,nu,0∈[σM(∇un+1,εp,n+1)−∂Wp∂q−∂ψ(qn+1−qnΔt)]xgg=1,…,ng,[εp,n+1−εp,nΔt]xg=[qn+1−qnΔtM(∇un+1,εp,n+1)]xgg=1,…,ng.(33)\begin{aligned} 0 = R_i := {}& \int_{\Omega} \left[\mathbb{C}\left(\varepsilon(u^{n+1})-\varepsilon^{p,n+1}\right)\cdot\nabla N_i-b\cdot N_i\right]\,\mathrm{d}\Omega-\int_{\partial_f\Omega}f\cdot N_i\,\mathrm{d}S && i=1,\ldots,n_u,\\ 0 &\in\left[\sigma_M(\nabla u^{n+1},\varepsilon^{p,n+1})-\frac{\partial W^p}{\partial q}-\partial\psi\left(\frac{q^{n+1}-q^n}{\Delta t}\right)\right]_{x_g} && g=1,\ldots,n_g,\\ \left[\frac{\varepsilon^{p,n+1}-\varepsilon^{p,n}}{\Delta t}\right]_{x_g} &= \left[\frac{q^{n+1}-q^n}{\Delta t}M(\nabla u^{n+1},\varepsilon^{p,n+1})\right]_{x_g} && g=1,\ldots,n_g. \tag*{(33)} \end{aligned}

We solve this through a nested Newton-Raphson approach. From the unidirectional nature of MM, we have M(∇un+1,εp,n+1)=M(∇un+1,εp,n)M(\nabla u^{n+1},\varepsilon^{p,n+1})=M(\nabla u^{n+1},\varepsilon^{p,n}). Then, from the last line (33), we may explicitly write εp,n+1=εp,n+1(∇un+1,qn+1,εp,n,qn)\varepsilon^{p,n+1}=\varepsilon^{p,n+1}(\nabla u^{n+1},q^{n+1},\varepsilon^{p,n},q^n). We may then reduce the plastic updates to a single scalar yield equation from the second line of (33),

0∈[σM(∇un+1,εp,n+1(∇un+1,qn+1,εp,n,qn))−∂Wp∂q(qn+1)−∂ψ(qn+1−qnΔt)]xg(34)0 \in\left[\sigma_M\left(\nabla u^{n+1},\varepsilon^{p,n+1}(\nabla u^{n+1},q^{n+1},\varepsilon^{p,n},q^n)\right)-\frac{\partial W^p}{\partial q}(q^{n+1})-\partial\psi\left(\frac{q^{n+1}-q^n}{\Delta t}\right)\right]_{x_g} \tag*{(34)}

for g=1,…,ngg = 1,\ldots,n_g. Then, given ∇un+1\nabla u^{n+1}, εp,n\varepsilon^{p,n}, and qnq^n, the above equation is a scalar relation for qn+1q^{n+1} at each quadrature point. We solve this through a standard Newton-Raphson method. Then, from the solution of the above, we may write qn+1=qn+1(∇un+1,εp,n,qn)q^{n+1}=q^{n+1}(\nabla u^{n+1},\varepsilon^{p,n},q^n) and therefore εp,n+1=εp,n+1(∇un+1,εp,n,qn)\varepsilon^{p,n+1}=\varepsilon^{p,n+1}(\nabla u^{n+1},\varepsilon^{p,n},q^n). This allows us to reduce the entire system to

0=Ri:=∫Ω[C(ε(un+1)−εp,n+1(∇un+1,εp,n,qn))⋅∇Ni−b⋅Ni] dΩ−∫∂fΩf⋅Ni dS(35)0=R_i:=\int_{\Omega}\left[\mathbb{C}\left(\varepsilon(u^{n+1})-\varepsilon^{p,n+1}(\nabla u^{n+1},\varepsilon^{p,n},q^n)\right)\cdot\nabla N_i-b\cdot N_i\right]\,d\Omega-\int_{\partial_f\Omega}f\cdot N_i\,dS \tag*{(35)}

for i=1,…,nui=1,\ldots,n_u, where the plastic update relations are accounted for through the dependence of εp,n+1\varepsilon^{p,n+1} on ∇un+1\nabla u^{n+1}, εp,n\varepsilon^{p,n} and qnq^n. We solve this system through a Newton-Raphson method. Thus, it is necessary to compute the stiffness matrix

Kij:=∫Ω[C(∇Nj−∂εp,n+1∂∇u⋅∇Nj)⋅∇Ni] dΩ(36)K_{ij}:=\int_{\Omega}\left[\mathbb{C}\left(\nabla N_j-\frac{\partial\varepsilon^{p,n+1}}{\partial\nabla u}\cdot\nabla N_j\right)\cdot\nabla N_i\right]\,d\Omega \tag*{(36)}

where

∂εp,n+1∂∇u=M(∇un+1,εp,n)⊗∂qn+1∂∇u+(qn+1−qn)∂M∂∇u.(37)\frac{\partial\varepsilon^{p,n+1}}{\partial\nabla u}=M(\nabla u^{n+1},\varepsilon^{p,n})\otimes\frac{\partial q^{n+1}}{\partial\nabla u}+(q^{n+1}-q^n)\frac{\partial M}{\partial\nabla u}. \tag*{(37)}

Here, the derivative ∂qn+1∂∇u\frac{\partial q^{n+1}}{\partial\nabla u} is found through an implicit differentiation of (34) giving

∂qn+1∂∇u=2μ∂2Wp∂q2+1Δt∂2ψ∂q˙2+3μM(∇un+1,εp,n).(38)\frac{\partial q^{n+1}}{\partial\nabla u}=\frac{2\mu}{\frac{\partial^2W_p}{\partial q^2}+\frac{1}{\Delta t}\frac{\partial^2\psi}{\partial\dot{q}^2}+3\mu}M(\nabla u^{n+1},\varepsilon^{p,n}). \tag*{(38)}

Then, starting from un+1,1=unu^{n+1,1}=u^n and p=1p=1, we conduct Newton Raphson iterations over pp in the form

Kij(un+1,p)Δuj=−Ri(un+1,p),i=1,…,nu,j=1,…,nuujn+1,p+1=Δuj+ujn+1,p,j=1,…,nu.(39)\begin{aligned} K_{ij}(u^{n+1,p})\Delta u_j&=-R_i(u^{n+1,p}),\qquad i=1,\ldots,n_u,\quad j=1,\ldots,n_u\\ u_j^{n+1,p+1}&=\Delta u_j+u_j^{n+1,p},\qquad j=1,\ldots,n_u. \tag*{(39)} \end{aligned}

until ∣R(un+1,p+1)∣<tol\lvert R(u^{n+1,p+1})\rvert<tol, and set un+1=un+1,p+1u^{n+1}=u^{n+1,p+1} and qn+1=qn+1(∇un+1,εp,n,qn)q^{n+1}=q^{n+1}(\nabla u^{n+1},\varepsilon^{p,n},q^n) and εp,n+1=εp,n+1(∇un+1,εp,n,qn)\varepsilon^{p,n+1}=\varepsilon^{p,n+1}(\nabla u^{n+1},\varepsilon^{p,n},q^n).

Dynamic

We specialize to the dynamic compression test where a disc (possibly with a hole) is compressed axially. See § 3.2 and Appendix § 8. The governing equations are (51).

We use an explicit central difference scheme to update the displacement field. The plasticity updates qq and εp\varepsilon^p are then updated implicitly with a backwards Euler update. For the nn to n+1n+1 time-step the displacement updates are

uˉ¨in=Mij−1Fjn(uˉn,εp,n,qn,tn),uˉ˙in+1/2=uˉ˙in−1/2+Δtnuˉ¨inuˉin+1=uˉin+Δtn+1/2uˉ˙in+1/2.(40)\begin{aligned} \ddot{\bar{u}}_i^n&=M_{ij}^{-1}F_j^n(\bar{u}^n,\varepsilon^{p,n},q^n,t^n),\\ \dot{\bar{u}}_i^{n+1/2}&=\dot{\bar{u}}_i^{n-1/2}+\Delta t^n\ddot{\bar{u}}_i^n\\ \bar{u}_i^{n+1}&=\bar{u}_i^n+\Delta t^{n+1/2}\dot{\bar{u}}_i^{n+1/2}. \tag*{(40)} \end{aligned}

where

Mij=∫Ωρ(x)Ni⋅Ni dΩ,Fjn=∫Ω[−σ(εn,εp,n,αn,η)⋅∇Nj+b⋅Nj] dΩ.(41)M_{ij}=\int_{\Omega}\rho(x)N_i\cdot N_i\,d\Omega,\qquad F_j^n=\int_{\Omega}\left[-\sigma(\varepsilon^n,\varepsilon^{p,n},\alpha^n,\eta)\cdot\nabla N_j+b\cdot N_j\right]\,d\Omega. \tag*{(41)}

In standard fashion, the integrals above are approximated with Gauss quadrature. We then update the plasticity variables through an implicit backwards Euler discretization. For this, we employ a predictor-corrector scheme [41] to solve point-wise at each quadrature point,

0∈σˉM(εn+1∣xg,εgp,(n+1),η(xg))−∂Wp∂q−∂g∗(qgn+1−qgnΔt,η(xg)),0\in\bar{\sigma}_M(\varepsilon^{n+1}|_{x_g},\varepsilon_g^{p,(n+1)},\eta(x_g))-\frac{\partial W^p}{\partial q}-\partial g^*\left(\frac{q_g^{n+1}-q_g^n}{\Delta t},\eta(x_g)\right),
εgp,(n+1)=εgp,n+ΔqM(εgn+1,εgp,(n+1)).(42)\varepsilon_g^{p,(n+1)}=\varepsilon_g^{p,n}+\Delta qM(\varepsilon_g^{n+1},\varepsilon_g^{p,(n+1)}). \tag*{(42)}

Adjoint Problem

Quasistatic

The governing set of quasistatic adjoint relations are

0=∫Ω[∂o∂u⋅δPu+(∇v⋅∂2We∂ε∂ε+q˙γ∂σˉM∂ε−q˙ζ⋅∂M∂ε)⋅∇δPu] dΩ∀δPu∈U,0 = \int_{\Omega} \left[ \frac{\partial o}{\partial u}\cdot\delta_P u + \left(\nabla v\cdot\frac{\partial^2 W^{e}}{\partial\varepsilon\partial\varepsilon}+\dot{q}\gamma\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon}-\dot{q}\zeta\cdot\frac{\partial M}{\partial\varepsilon}\right)\cdot\nabla\delta_P u \right]\,d\Omega\quad\forall\delta_P u\in\mathcal{U},
ddt[γ(σˉM−∂Wp∂q−∂g˙∗∂q˙)−γ˙q∂2g∗∂q˙2−ζ⋅M]=∂o∂q−γ˙q∂2Wp∂q2on Ω,\frac{d}{dt}\left[\gamma\left(\bar{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial\dot{g}^{*}}{\partial\dot{q}}\right)-\dot{\gamma}q\frac{\partial^{2}g^{*}}{\partial\dot{q}^{2}}-\zeta\cdot M\right]=\frac{\partial o}{\partial q}-\dot{\gamma}q\frac{\partial^{2}W^{p}}{\partial q^{2}}\quad\text{on }\Omega,
dζdt=∂o∂εp+∇v⋅∂2We∂ε∂εp+q˙γ∂σˉM∂εp−q˙ζ⋅∂M∂εpon Ω,\frac{d\zeta}{dt}=\frac{\partial o}{\partial\varepsilon^{p}}+\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon^{p}}+\dot{q}\gamma\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon^{p}}-\dot{q}\zeta\cdot\frac{\partial M}{\partial\varepsilon^{p}}\quad\text{on }\Omega,
γ∣t=T=0,ζ∣t=T=0.(43)\left.\gamma\right|_{t=T}=0,\qquad\left.\zeta\right|_{t=T}=0. \tag*{(43)}

As the boundary conditions are found at the final time t=Tt=T, we solve this system backwards in time. We consider a fully implicit approach, with a backward Euler approximation of the temporal derivatives of the plastic variables. We examine this from the n+1n+1 to nn time step. That is, we look to solve for {vn,γn,ζn}\{v^{n},\gamma^{n},\zeta^{n}\} given {vn+1,γn+1,ζn+1}\{v^{n+1},\gamma^{n+1},\zeta^{n+1}\} assuming the discretizations in (30) and (31) given the complete set of forward-problem solution variables. Thus, we look to solve

0=Riadj=∫Ω[∂o∂u∣tn⋅Ni+(∇vn⋅∂2We∂ε∂ε∣tn+γn(q˙∂σˉM∂ε)tn−ζn⋅(q˙∂M∂ε)tn)⋅∇Ni] dΩ0=R_{i}^{adj}=\int_{\Omega}\left[\left.\frac{\partial o}{\partial u}\right|_{t_n}\cdot N_i+\left(\nabla v^{n}\cdot\left.\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon}\right|_{t_n}+\gamma^{n}\left(\dot{q}\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon}\right)_{t_n}-\zeta^{n}\cdot\left(\dot{q}\frac{\partial M}{\partial\varepsilon}\right)_{t_n}\right)\cdot\nabla N_i\right]\,d\Omega
1Δt[γ(σˉM−∂Wp∂q−∂g˙∗∂q˙)−γ˙q∂2g∗∂q˙2−ζ⋅M]tntn+1=∂o∂q∣tn−γn(q˙∂2Wp∂q2)tn\frac{1}{\Delta t}\left[\gamma\left(\bar{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial\dot{g}^{*}}{\partial\dot{q}}\right)-\dot{\gamma}q\frac{\partial^{2}g^{*}}{\partial\dot{q}^{2}}-\zeta\cdot M\right]_{t_n}^{t_{n+1}}=\left.\frac{\partial o}{\partial q}\right|_{t_n}-\gamma^{n}\left(\dot{q}\frac{\partial^{2}W^{p}}{\partial q^{2}}\right)_{t_n}
ζn+1−ζnΔt=∂o∂εp∣tn+∇vn⋅∂2We∂ε∂εp∣tn+γn(q˙∂σˉM∂εp)tn−ζn⋅(q˙∂M∂εp)tn(44)\frac{\zeta^{n+1}-\zeta^{n}}{\Delta t}=\left.\frac{\partial o}{\partial\varepsilon^{p}}\right|_{t_n}+\left.\nabla v^{n}\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon^{p}}\right|_{t_n}+\gamma^{n}\left(\dot{q}\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon^{p}}\right)_{t_n}-\zeta^{n}\cdot\left(\dot{q}\frac{\partial M}{\partial\varepsilon^{p}}\right)_{t_n} \tag*{(44)}

This is a linear set of equations for the {vn,γn,ζn}\{v^{n},\gamma^{n},\zeta^{n}\} and can be solved through direct inversion.

Dynamic

Similar to the forward problem, we apply an explicit central difference scheme to update the adjoint field vv. The other adjoint variables, namely γ\gamma and ζ\zeta are solved locally using then updated implicitly with a backwards Euler update. For the nn to n+1n+1 time-step the displacement updates are

vˉ¨in=Mij−1Hjn(vˉn,εp,n,qn,tn),\ddot{\bar{v}}_{i}^{n}=M_{ij}^{-1}H_{j}^{n}(\bar{v}^{n},\varepsilon^{p,n},q^{n},t^{n}),
vˉ˙in+1/2=vˉ˙in−1/2+Δtnvˉ¨in,\dot{\bar{v}}_{i}^{n+1/2}=\dot{\bar{v}}_{i}^{n-1/2}+\Delta t^{n}\ddot{\bar{v}}_{i}^{n},
vˉin+1=vˉin+Δtn+1/2vˉ˙in+1/2.(45)\bar{v}_{i}^{n+1}=\bar{v}_{i}^{n}+\Delta t^{n+1/2}\dot{\bar{v}}_{i}^{n+1/2}. \tag*{(45)}

where

Mij=∫Ωρ(x)Ni⋅Nj dΩ,M_{ij}=\int_{\Omega}\rho(x)N_i\cdot N_j\,d\Omega,
Hjn=∫Ω[(∇vn⋅∂2We∂ε∂ε+γnq˙n∂σˉM∂ε−q˙nζn⋅∂M∂ε)⋅∇Nj−∂o∂u⋅Nj] dΩ.(46)H_{j}^{n}=\int_{\Omega}\left[\left(\nabla v^{n}\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon}+\gamma^{n}\dot{q}^{n}\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon}-\dot{q}^{n}\zeta^{n}\cdot\frac{\partial M}{\partial\varepsilon}\right)\cdot\nabla N_j-\frac{\partial o}{\partial u}\cdot N_j\right]\,d\Omega. \tag*{(46)}

In standard fashion, the integrals above are approximated with Gauss quadrature. We then update the adjoint variables γ\gamma and ζ\zeta through an implicit backwards Euler discretization. For this, we solve a linear set of equations point-wise at each quadrature point,

γn+1(σˉM−∂Wp∂q−∂gˉ∗∂q˙−q˙∂2gˉ∗∂q˙2)−ζn+1⋅M=γn(σˉM−∂Wp∂q−∂gˉ∗∂q˙−q˙∂2gˉ∗∂q˙2),−ζn⋅M+Δt∂o∂q−Δtγnq˙∂2Wp∂q2,ζn+1=ζn+∂o∂εp+∇v⋅∂2We∂ε∂εp+γnq˙∂σˉM∂εp−q˙ζn⋅∂M∂εp.(47)\begin{aligned} \gamma^{n+1}\left(\bar{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial\bar{g}^{*}}{\partial\dot{q}}-\dot{q}\frac{\partial^{2}\bar{g}^{*}}{\partial\dot{q}^{2}}\right)-\zeta^{n+1}\cdot M&=\gamma^{n}\left(\bar{\sigma}_{M}-\frac{\partial W^{p}}{\partial q}-\frac{\partial\bar{g}^{*}}{\partial\dot{q}}-\dot{q}\frac{\partial^{2}\bar{g}^{*}}{\partial\dot{q}^{2}}\right),\\ -\zeta^{n}\cdot M+\Delta t\frac{\partial o}{\partial q}-\Delta t\gamma^{n}\dot{q}\frac{\partial^{2}W^{p}}{\partial q^{2}},\\ \zeta^{n+1}&=\zeta^{n}+\frac{\partial o}{\partial\varepsilon^{p}}+\nabla v\cdot\frac{\partial^{2}W^{e}}{\partial\varepsilon\partial\varepsilon^{p}}+\gamma^{n}\dot{q}\frac{\partial\bar{\sigma}_{M}}{\partial\varepsilon^{p}}-\dot{q}\zeta^{n}\cdot\frac{\partial M}{\partial\varepsilon^{p}}. \tag*{(47)} \end{aligned}

Dimension Reduction for Annular Specimen

The annulus ring being thin allows us to perform computations on a reduced dimensional space enabling quicker algorithm. Reiterating our approximation of uniform axial strain along the e3e_{3} direction,

u(X1,X2,X3,t)=uˉ(X1,X2,t)+(λ(t)−1)X3e3,(48)u(X_{1},X_{2},X_{3},t)=\bar{u}(X_{1},X_{2},t)+(\lambda(t)-1)X_{3}e_{3}, \tag*{(48)}

where uˉ:=Ωˉ↦R2\bar{u}:=\bar{\Omega}\mapsto\mathbb{R}^{2} is the in-plane displacement. The corresponding deformation gradient is given by

∇u=(∇ˉuˉ00λ),(49)\nabla u=\begin{pmatrix}\bar{\nabla}\bar{u}&0\\0&\lambda\end{pmatrix}, \tag*{(49)}

where ∇ˉ\bar{\nabla} is the gradient computed along the in-plane directions {e1,e2}\{e_{1},e_{2}\}. The governing equations are,

0=∫Ω[ρu¨⋅δu+σ⋅∇δu−b⋅δu] dΩ−∫∂fΩf⋅δu dS∀δu∈K0,0∈σM−∂Wp∂q−∂ψ(q˙)on Ω,ε˙p=q˙Mon Ω.(50)\begin{aligned} 0&=\int_{\Omega}\left[\rho\ddot{u}\cdot\delta u+\sigma\cdot\nabla\delta u-b\cdot\delta u\right]\,d\Omega-\int_{\partial_{f}\Omega}f\cdot\delta u\,dS &&\forall\delta u\in\mathcal{K}_{0},\\ 0&\in\sigma_{M}-\frac{\partial W^{p}}{\partial q}-\partial\psi(\dot{q}) &&\text{on }\Omega,\\ \dot{\varepsilon}^{p}&=\dot{q}M &&\text{on }\Omega. \tag*{(50)} \end{aligned}

Substituting the approximation of the displacement 48 into the above equations, we obtain the simplified form

0=−h2∥Ω∥3ρλ¨−∫Ωσ33+∫Ωf⋅e3 ds∣X3=h,0=∫Ω[ρuˉ¨⋅δuˉ+σ⋅∇ˉδuˉ−b⋅δuˉ] dΩ∀δuˉ∈K0,0∈σM−∂Wp∂q−∂ψ(q˙)on Ω,ε˙p=q˙Mon Ω.(51)\begin{aligned} 0&=-\frac{h^{2}\lVert\Omega\rVert}{3}\rho\ddot{\lambda}-\int_{\Omega}\sigma_{33}+\left.\int_{\Omega}f\cdot e_{3}\,ds\right|_{X_{3}=h},\\ 0&=\int_{\Omega}\left[\rho\ddot{\bar{u}}\cdot\delta\bar{u}+\sigma\cdot\bar{\nabla}\delta\bar{u}-b\cdot\delta\bar{u}\right]\,d\Omega&&\forall\delta\bar{u}\in\mathcal{K}_{0},\\ 0&\in\sigma_{M}-\frac{\partial W^{p}}{\partial q}-\partial\psi(\dot{q}) &&\text{on }\Omega,\\ \dot{\varepsilon}^{p}&=\dot{q}M &&\text{on }\Omega. \tag*{(51)} \end{aligned}

Since uˉ\bar{u} exists in a 2-dimensional space, the computational cost for solving the above equations are cheaper than solving for a 3-dimensional system of equations. The dynamic boundary condition involves specifying λ(t)\lambda(t); therefor, the first equation is not necessary. The net force on the surface X3=hX_{3}=h can be calculated using

f=∫Ωσ33 dΩ∣X3=h(52)f=\left.\int_{\Omega}\sigma_{33}\,d\Omega\right|_{X_{3}=h} \tag*{(52)}

Our experimental data consists of the final in plane displacement uˉexp⁡(X1,X2,T)\bar{u}^{\exp}(X_{1},X_{2},T) and the net axial force fRexp⁡f_{R}^{\exp}, which are compared with the inplane-displacements obtained from equation 51 and the surface forces computed using 52.

References

  1. [1]A. Akerson. Optimal structures for failure resistance under impact. Journal of the Mechanics and Physics of Solids, 172:105172, 3 2023.DOI
  2. [2]A. Akerson, B. Bourdin, and K. Bhattacharya. Optimal design of responsive structures. Structural and Multidisciplinary Optimization, 65, 2022.arxiv.org/abs/2107.08895
  3. [3]A. Akerson and T. Liu. Mechanics, modeling, and shape optimization of electrostatic zipper actuators. Journal of the Mechanics and Physics of Solids, 181:105446, 10 2023.DOI
  4. [4]D. Arndt, W. Bangerth, D. Davydov, T. Heister, L. Heltai, M. Kronbichler, M. Maier, J.-P. Pelteret, B. Turcksin, and D. Wells. The deal.ii finite element library: Design, features, and insights. Computers & Mathematics with Applications, 81:407–422, 2021.DOI
  5. [5]W. Bangerth, R. Hartmann, and G. Kanschat. Deal.II - -A general-purpose object-oriented finite element library. ACM Transactions on Mathematical Software, 33:1–27, 2007.DOI
  6. [6]M. Bendsøe and O. Sigmund. Topology optimization. Theory, methods, and applications. 2nd ed., corrected printing. Springer, Berlin, Heidelberg, 2004.
  7. [7]J. V. Bernier, R. M. Suter, A. D. Rollett, and J. D. Almer. High-energy x-ray diffraction microscopy in materials science. Annual Review of Materials Research, 50:395–436, 2020.DOI
  8. [9]E. Billington and A. Tate. Physics of Deformation and Flow: An Introduction. McGraw Hill, 1981.
  9. [10]M. Bonnet and A. Constantinescu. Inverse problems in elasticity. Inverse Problems, 21:R1–R50, 2005.DOI
  10. [11]A. Buljac, C. Jailin, A. Mendoza, J. Neggers, T. Taillandier-Thomas, A. Bouterf, B. Smaniotto, F. Hild, and S. Roux. Digital Volume Correlation: Review of Progress and Challenges. Experimental Mechanics, 58:661–708, 2018.DOI
  11. [12]R. Carson, M. Obstałecki, M. Miller, and P. Dawson. Characterizing heterogeneous intra-granular deformations in polycrystalline solids using diffraction-based and mechanics-based metrics. Modelling and Simulation in Materials Science and Engineering, 25(5):055008, July 2017.
  12. [13]T. Chu, W. Ranson, and M. A. Sutton. Applications of digital-image-correlation techniques to experimental mechanics. Experimental Mechanics, 25:232–244, 1985.DOI
  13. [15]A. Constantinescu and N. Tardieu. On the identification of elastoviscoplastic constitutive laws from indentation tests. Inverse Problems in Engineering, 9:19–44, 2001.DOI
  14. [17]C. Efstathiou, D. E. Boyce, J. S. Park, U. Lienert, P. R. Dawson, and M. P. Miller. A method for measuring single-crystal elastic moduli using high-energy X-ray diffraction and a crystal-based finite element model. Acta Materialia, 58:5806–5819, 2010.DOI
  15. [18]A. Eghtesad, J. N. Fuhg, and N. Bouklas. NN-EVP: A physics informed neural network-based elasto-viscoplastic framework for predictions of grain size-aware flow response under large deformations, 2023. arXiv:2307.04301.arxiv.org/abs/2307.04301
  16. [19]M. Grediac, F. Pierron, S. Avril, and E. Toussaint. The virtual fields method for extracting constitutive parameters from full-field measurements: A review. Strain, 42:233–253, 2006.
  17. [20]M. Gurtin. Introduction to Continuum Mechanics. Academic Press, 1970.
  18. [21]C. M. Hamel, K. N. Long, and S. L. B. Kramer. Calibrating constitutive models with full-field data via physics informed neural networks, 2022. arXiv:2203.16577.arxiv.org/abs/2203.16577
  19. [23]F. Hild and S. Roux. Digital image correlation: from displacement measurement to identification of elastic properties ? a review. Strain, 42:69–80, 2006.arxiv.org/abs/physics/0511122
  20. [24]M. Hinze, R. Pinnau, M. Ulbrich, and S. Ulbrich. Optimization with PDE Constraints. Springer, 2009.
  21. [25]M. F. Horstemeyer and D. J. Bammann. Historical review of internal state variable theory for inelasticity. International Journal of Plasticity, 26:1310–1334, 2010.DOI
  22. [26]A. D. Kammers and S. Daly. Digital Image Correlation under Scanning Electron Microscopy: Methodology and Validation. Experimental Mechanics, 53:1743–1761, 2013.DOI
  23. [28]H. B. Khaniki, M. H. Ghayesh, R. Chin, and M. Amabili. Hyperelastic structures: A review on the mechanics and biomechanics. International Journal of Non-Linear Mechanics, 148:104275, 2023.DOI
  24. [29]C. Kim and M.-G. Lee. Finite element-based virtual fields method with pseudo-real deformation fields for identifying constitutive parameters. International Journal of Solids and Structures, 233:111204, 2021.DOI
  25. [32]S. L. B. Kramer, Scherzinger W.M. Implementation and Evaluation of the Virtual Fields Method: Determining Constitutive Model Parameters From Full-Field Deformation Data, Sandia Report, SAND2014- 17871, 2014.DOI
  26. [34]J. Lubliner. Plasticity Theory. Dover books on engineering. Dover Publications, 2008.
  27. [35]A. Marek, F. M. Davis, and F. Pierron. Sensitivity-based virtual fields for the non-linear virtual fields method. Computational Mechanics, 60:409–431, Sept. 2017.DOI
  28. [36]F. Mathieu, H. Leclerc, F. Hild, and S. Roux. Estimation of elastoplastic parameters via weighted FEMU and integrated-DIC. Experimental Mechanics, 55:105–119, 2015.DOI
  29. [37]S. K. Melly, L. Liu, Y. Liu, and J. Leng. A review on material models for isotropic hyperelasticity. International Journal of Mechanical System Dynamics, 1:71–88, 2021.DOI
  30. [39]A. A. Oberai, N. H. Gokhale, and G. R. Feijà o. Solution of inverse problems in elasticity imaging using the adjoint method. Inverse Problems, 19:297–313, Apr. 2003.DOI
  31. [40]M. Obstalecki, S. L. Wong, P. R. Dawson, and M. P. Miller. Quantitative analysis of crystal scale deformation heterogeneity during cyclic plasticity using high-energy X-ray diffraction and finite-element simulation. Acta Materialia, 75:259–272, 2014.DOI
  32. [41]M. Ortiz and L. Stainier. The variational formulation of viscoplastic constitutive updates. Computer Methods in Applied Mechanics and Engineering, 171:419–444, 1999.DOI
  33. [43]C. Pedersen. Topology optimization design of crushed 2d-frames for desired energy absorption history. Structural and Multidisciplinary Optimization, 25:368–382, 01 2003.DOI
  34. [44]F. Pierron and M. Grédiac. The Virtual Fields Method: Extracting Constitutive Mechanical Parameters from Full-field Deformation Measurements. Springer New York, New York, NY, 2012.DOI
  35. [46]S. Ravindran, V. Gandhi, A. Joshi, and G. Ravichandran. Three-dimensional full-field velocity measurements in shock compression experiments using stereo digital image correlation. Review of Scientific Instruments, 94(2), 2023.DOI
  36. [50]M. Sutton, J. Yan, V. Tiwari, H. Schreier, and J. Orteu. The effect of out-of-plane motion on 2d and 3d digital image correlation measurements. Optics and Lasers in Engineering, 46:746–757, 2008.DOI
  37. [51]M. A. Sutton, J. J. Orteu, and H. Schreier. Image correlation for shape, motion and deformation measurements: basic concepts, theory and applications. Springer Science & Business Media, 2009.DOI
  38. [52]K. Svanberg. The method of moving asymptotes - a new method for structural optimization. International Journal for Numerical Methods in Engineering., 24:359–373, 1987.DOI
  39. [53]K. Upadhyay, J. N. Fuhg, N. Bouklas, and K. T. Ramesh. Physics-informed Data-driven Discovery of Constitutive Models with Application to Strain-Rate-sensitive Soft Materials, 2023. arXiv:2304.13897.DOI
  40. [54]J. Wang, T. Li, F. Cui, C.-Y. Hui, J. Yeo, and A. T. Zehnder. Metamodeling of constitutive model using Gaussian process machine learning. Journal of the Mechanics and Physics of Solids, 154:104532, 2021.DOI
  41. [55]J. Wang, T. Li, C.-Y. Hui, J. Yeo, and A. Zehnder. Gaussian process to identify hydrogel constitutive model. In Challenges in Mechanics of Time Dependent Materials, Mechanics of Biological Systems and Materials & Micro-and Nanomechanics, Volume 2: Proceedings of the 2021 Annual Conference & Exposition on Experimental and Applied Mechanics, pages 79–83. Springer, 2022.DOI
  42. [56]E. Wielewski, D. E. Boyce, J.-S. Park, M. P. Miller, and P. R. Dawson. A methodology to determine the elastic moduli of crystals by matching experimental and simulated lattice strain pole figures using discrete harmonics. Acta Materialia, 126:469–480, 2017.DOI
  43. [57]T. Wu, B. Rosić, L. De Lorenzis, and H. G. Matthies. Parameter identification for phase-field modeling of fracture: a Bayesian approach with sampling-free update. Computational Mechanics, 67:435–453, 2021.DOI
  44. [59]J. Yang, L. Hazlett, A. Landauer, and C. Franck. Augmented Lagrangian Digital Volume Correlation (ALDVC). Experimental Mechanics, 60:1205–1223, 2020.DOI
  45. [60]H. Zhou, R. A. Lebensonh, P. Reischig, W. Ludwig, and K. Bhattacharya. Imposing equilibrium on experimental 3-D stress fields using Hodge decomposition and FFT-based optimization. Mechanics of Materials, 164:104109, 2022.arxiv.org/abs/2105.01612

Paper details

vlnc:W_5z857yzm
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 1

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents