betaFeedback

WorksComputer Science

A physics-informed GAN Framework based on Model-free Data-Driven Computational Mechanics

Abstract

Model-free data-driven computational mechanics, first proposed by Kirchdoerfer and Ortiz, replace phenomenological models with numerical simulations based on sample data sets in strain-stress space. In this study, we integrate this paradigm within physics-informed generative adversarial networks (GANs). We enhance the conventional physics-informed neural network framework by implementing the principles of data-driven computational mechanics into GANs. Specifically, the generator is informed by physical constraints, while the discriminator utilizes the closest strain-stress data to discern the authenticity of the generator's output. This combined approach presents a new formalism to harness data-driven mechanics and deep learning to simulate and predict mechanical behaviors.

Introduction

The simulation of boundary value problems typically contains two equations: conservation and constitutive laws. While conservation laws are derived from universal principles, constitutive laws are usually obtained by fitting model parameters to given strain-stress data [1]. Nevertheless, material modeling can be ill-posed and adds uncertainties to the solutions, particularly in highly complex systems. The model-free data-driven method, introduced by Kirchdoerfer and Ortiz [2], bypasses the step of material modeling, incorporating experimental data directly into the numerical simulations of boundary-value problems.

The data-driven scheme bypasses the empirical material modeling step by computing the closest point in the material data set consistent with the problem’s compatibility and equilibrium condition. Consequently, it provides an alternative formulation of the classical initial-boundary-value problem based on nearest-neighbor clustering.

The approach has been fine-tuned for diverse applications: from non-linear elasticity [2, 3, 4, 5, 6] to dynamics [7] and finite strain [8]. It’s also been adapted for material data identification [9], non-local mechanics [10], electro-mechanical problems [11], homogenization schemes [12], and model-driven coupling [13].

Ibañez et al. [14, 15] refined the approach using a manifold learning method that maps data into a lower-dimensional space to use the locally linear embeddings. Eggersmann et al. [16] presented a second-order data-driven approach that uses tensor voting [17] to obtain point-wise tangent space, enabling the search for additional states close to the original data. For inelastic boundary value problems, Eggersmann et al. [18] include local histories in the data set to investigate materials with memory. Karapiperis et al. [19] have also suggested a variation of the scheme, considering multiscale modeling. In addition, we recently developed a paradigm incorporating the tangent space into the distance-minimizing data-driven formulation and classifies the underlying data structure into subsets according to various material behaviors [20]. The framework features a parametrization of the material history and an optimal sampling of the mechanical system’s state space.

The paradigm’s dependence on the nearest-neighbor clustering of data points proposes research areas in machine-learning methods, particularly Artificial Neural Networks (ANNs), that are known to approximate any continuous function for appropriate network parameters [21, 22]. The flexibility and quality of neural networks led to success in a wide range of applications, e.g., image recognition [23], language processing [24], or generative modeling [25, 26]. An extension to neural networks is physics-informed deep learning, successfully used in solving physical-related problems such as fluid mechanics [27, 28], aerodynamics [29, 30], shell structures [31] or material science [32, 33]. Physics-Informed Neural Networks (PINNs) can be trained to fulfill training data and learn optimal solutions for allocated physics-governing equations by specifying appropriate loss functions [34, 35]. The physics-based loss competes against a data-based loss, which is needed to provide fundamental knowledge of the system. Thus, partial differential equations act as additional constraints during network training, resulting in a multi-objective optimization problem. Optimizing data and physics give physics-informed neural networks flexibility in solving forward and inverse problems [28, 29, 30, 36, 37, 38]. The trade-off between the individual losses can be influenced using hyper-parameter [39, 40, 41]. For example, adaptive activation functions [42, 43], or manually weighted losses [44], can improve the quality of the neural network for specific problems. Another approach to overcome the local convergence issue due to global approximation is the usage of adaptive training strategies and domain decomposition [45]. This investigation combines the model-free data-driven approach with Generative Adversarial networks (GANs). In machine learning, GANs have emerged as a powerful tool consisting of two neural networks – the generator, which creates data, and the discriminator, which evaluates the authenticity of the generated data. Through their adversarial game, GANs are adept at generating high-fidelity data, often indistinguishable from actual data [26]. An extension is the integration of physics-informed neural networks with the GAN structure. For instance, the pursuit of robust uncertainty quantification within the framework of PINNs has led to recent methodologies. The PIG-GAN framework [46] harnesses the capabilities of a physics-informed generator to address adversarial uncertainty. On the other hand, the PID-GAN approach [47] uses a physics-informed discriminator, carving out a distinct avenue to achieve reliable uncertainty quantification while maintaining fidelity to the governing physics. Another stride in this direction is the DeqGAN, which offers a unique perspective on PINNs by learning the loss function via generative adversarial networks. This methodology provides a robust avenue for solving the challenges traditionally associated with defining appropriate loss functions for PINNs [48]. In our approach, the generator is a physics-informed neural network, and the discriminator employs the closest strain-stress data to evaluate the authenticity of the generator’s results. This synergized methodology matches model-free data-driven computational mechanics and deep learning principles, aspiring to more accurately simulate and predict intricate mechanical behaviors.

Section 2 provides a general setting by introducing the definitions and derivation of the distance-minimizing data-driven computing method based on [4]. § 3 introduces the framework of artificial neural networks and generative adversarial networks. In addition, we propose using a physics-informed GAN to solve the distance-minimizing data-driven problem. § 4 exhibits the performance of the proposed method using a numerical example involving a non-linear elastic in-plane boundary value problem. Finally, § 5 summarizes the results and suggests future research subjects.

Model-free Data-driven setting

The following will summarize the classical data-driven computational mechanics method for the reader’s convenience based on the definitions and formulations in [4]. We consider an elastic body Ω⊂Rd\Omega\subset\mathbb{R}^{d} whose internal states are defined by displacement field u:Ω→Rd\mathbf{u} : \Omega\to\mathbb{R}^{d} and the compatibility and equilibrium conditions

ε(x)−∇symu(x)=0,in Ω,∇⋅σ(x)−f(x)=0,in Ω,(1)\begin{aligned} \boldsymbol{\varepsilon}(\mathbf{x})-\nabla^{\mathrm{sym}}\mathbf{u}(\mathbf{x})&=0,\quad&&\text{in }\Omega,\\ \nabla\cdot\boldsymbol{\sigma}(\mathbf{x})-\mathbf{f}(\mathbf{x})&=0,\quad&&\text{in }\Omega, \tag*{(1)} \end{aligned}

and boundary conditions

u(x)=g(x),on ΓD,σ(x)⋅n(x)=t(x),on ΓN,(2)\begin{aligned} \mathbf{u}(\mathbf{x})&=\mathbf{g}(\mathbf{x}),\quad&&\text{on }\Gamma_{D},\\ \boldsymbol{\sigma}(\mathbf{x})\cdot\mathbf{n}(\mathbf{x})&=\mathbf{t}(\mathbf{x}),\quad&&\text{on }\Gamma_{N}, \tag*{(2)} \end{aligned}

where ε:Ω→Rsymd×d\boldsymbol{\varepsilon}:\Omega\to\mathbb{R}_{\mathrm{sym}}^{d\times d} is the strain field and σ:Ω→Rsymd×d\boldsymbol{\sigma}:\Omega\to\mathbb{R}_{\mathrm{sym}}^{d\times d} is the stress field. The boundary Γ\Gamma of the domain Ω\Omega is defined by the Dirichlet (ΓD\Gamma_{D}) and Neumann (ΓN\Gamma_{N}) with Γ=ΓD∪ΓN\Gamma=\Gamma_{D}\cup\Gamma_{N} and ΓD∩ΓN=∅\Gamma_{D}\cap\Gamma_{N}=\varnothing. In addition, f:Ω→Rd\mathbf{f}:\Omega\to\mathbb{R}^{d} is the body force, and g,t,n:Γ→Rd\mathbf{g},\mathbf{t},\mathbf{n}:\Gamma\to\mathbb{R}^{d} define the boundary displacement, applied traction and outer normal, respectively.

We define Zloc⊂Rsymd×d×Rsymd×d\mathcal{Z}_{\mathrm{loc}}\subset\mathbb{R}_{\mathrm{sym}}^{d\times d}\times\mathbb{R}_{\mathrm{sym}}^{d\times d} as the local phase space consisting of pairs z(x)=(ε(x),σ(x))\mathbf{z}(\mathbf{x})=(\boldsymbol{\varepsilon}(\mathbf{x}),\boldsymbol{\sigma}(\mathbf{x})) describing the local state of the system at material point x\mathbf{x}. The global phase space Z\mathcal{Z} is defined as the collection of the state functions, i.e.

Z={z:z∈Zloc}.(3)\mathcal{Z}=\{\mathbf{z}:\mathbf{z}\in\mathcal{Z}_{\mathrm{loc}}\}. \tag*{(3)}

The data-driven distance-minimization problem, introduced by [2], reads

arg min⁡z∈C, z^∈D d(z,z^),(4)\underset{\mathbf{z}\in\mathcal{C},\,\widehat{\mathbf{z}}\in\mathcal{D}}{\operatorname{arg\,min}}\,d(\mathbf{z},\widehat{\mathbf{z}}), \tag*{(4)}

where C⊂Z\mathcal{C}\subset\mathcal{Z} denotes the constraint set defined by

C:={z∈Z:(1) and (2)};(5)\mathcal{C}:=\{\mathbf{z}\in\mathcal{Z}:(1)\text{ and }(2)\}; \tag*{(5)}

containing all states fulfilling compatibility and equilibrium. The set D⊂Z\mathcal{D}\subset\mathcal{Z} consists of a finite number of experimental measurements achieved from small-scale simulations and is defined by

D={z∈Z:z(x)∈Dloc}withDloc={(εi,σi)}i=1n.(6)\mathcal{D}=\{\mathbf{z}\in\mathcal{Z}:\mathbf{z}(\mathbf{x})\in\mathcal{D}_{\mathrm{loc}}\}\quad\text{with}\quad\mathcal{D}_{\mathrm{loc}}=\{(\boldsymbol{\varepsilon}_{i},\boldsymbol{\sigma}_{i})\}_{i=1}^{n}. \tag*{(6)}

where ne∈Nn_e \in\mathbb{N} is the number of local data points associated with the material point. The distance d:Z×Z→Rd : \mathcal{Z} \times\mathcal{Z} \to\mathbb{R} is defined by

d(z,z^):=∥z−z^∥,(7)d(\mathbf{z},\hat{\mathbf{z}}) := \lVert\mathbf{z} - \hat{\mathbf{z}} \rVert, \tag*{(7)}

metricized by the norm

∥z∥2:=∫Ω(12Cε⋅ε+12C−1σ⋅σ) dx,(8)\lVert\mathbf{z} \rVert^{2} := \int_{\Omega} \left( \frac{1}{2}\mathbf{C}\boldsymbol{\varepsilon} \cdot\boldsymbol{\varepsilon} + \frac{1}{2}\mathbf{C}^{-1}\boldsymbol{\sigma} \cdot\boldsymbol{\sigma} \right) \,\mathrm{d}x, \tag*{(8)}

where C∈Rsymd×d\mathbf{C} \in\mathbb{R}_{\mathrm{sym}}^{d \times d} is a symmetric positive definite matrix typically being of the type of elastic stiffness. Thus, the data-driven paradigm aims to find the closest point z\mathbf{z} in the constraint set C\mathcal{C} to z^\hat{\mathbf{z}} in the material data set D\mathcal{D}.

Challenges such as data availability, noise, inconsistency, and high dimensionality frequently arise in the data-driven paradigm. Traditional analytical and computational methods may need to be adjusted when addressing these issues. Consequently, the incorporation of machine learning, particularly methods like generative adversarial networks coupled with physics-informed generators, is considered. This integration is aimed at effectively handling the complexities of data-driven datasets, ensuring the outcomes remain consistent with domain-specific knowledge. The following sections will present a detailed discussion on the principles of artificial neural networks and physics-informed neural networks, illustrating the approach of physics-informed generative adversarial networks to solve the data-driven boundary value problem (4).

Generative adversarial networks with physics-informed generators for model-free data-driven problems

This section delves into the application of Generative Adversarial Networks (GANs) equipped with Physics-Informed Generators for addressing the model-free data-driven problem. A GAN involves a competitive dynamic between two neural networks, forming a zero-sum game: one network’s success implies the other’s setback. To harness GANs for resolving the data-driven boundary value problem depicted in (4), Section 3.1 initiates with a concise overview of Artificial Neural Networks (ANNs) and explains physics-informed neural networks (PINNs). Section 3.2 lays out the foundational principles of GANs, and in Section 3.3, we pivot to the novel approach of leveraging GANs augmented with PINNs to solve the data-driven boundary value problem.

Physics-informed neural networks

Based on the universal function approximation theorem [49], an artificial neural network is a parametrized, non-linear function composition that can approximate arbitrary Borel measurable functions. This section introduces the basic concept based on the definitions and formulations in [45]. For this purpose, we introduce a densely connected feed-forward neural network, denoted by the map N:Rdx×[0,T]→Rdy\mathcal{N} : \mathbb{R}^{d_x} \times[0,T] \to\mathbb{R}^{d_y}, which is defined by a composition of nL∈Nn_L \in\mathbb{N} non-linear functions:

N:Rdx×[0,T]→Rdy(9)\mathcal{N} : \mathbb{R}^{d_x} \times[0,T] \to\mathbb{R}^{d_y} \tag*{(9)}
(x,t)↦N(x,t)=y(ℓ)∘…∘y(0)=y,(10)(\mathbf{x},t) \mapsto\mathcal{N}(\mathbf{x},t) = \mathbf{y}^{(\ell)} \circ\ldots\circ\mathbf{y}^{(0)} = \mathbf{y}, \tag*{(10)}

for ℓ=1,…,nL\ell= 1,\ldots,n_L, where x\mathbf{x} denotes the spatial part of the input vector of dimension dx∈Nd_x \in\mathbb{N} at time t∈[0,T]t \in[0,T] with T>0T > 0 and y\mathbf{y} denotes the output vector of dimension dy∈Nd_y \in\mathbb{N}. In this context, y(0)\mathbf{y}^{(0)} and y(nL)\mathbf{y}^{(n_L)} are called the input and output layer, such that

y(0)=(x,t),y(nL)=N(x,t).(11)\mathbf{y}^{(0)} = (\mathbf{x},t), \qquad\mathbf{y}^{(n_L)} = \mathcal{N}(\mathbf{x},t). \tag*{(11)}

The functions y(ℓ)\mathbf{y}^{(\ell)} are called hidden layers and define a ℓ\ell-fold composition, mapping the input (x,t)(\mathbf{x},t) to the output y\mathbf{y} by

y(ℓ)={yη(ℓ), η=1,…,ηu},with yη(ℓ)=act⁡(ℓ)(Wη(ℓ)y(ℓ−1)+bη(ℓ)).(12)\mathbf{y}^{(\ell)} = \{\mathbf{y}^{(\ell)}_\eta,\,\eta= 1,\ldots,\eta_u\}, \quad\text{with } \mathbf{y}^{(\ell)}_\eta= \operatorname{act}^{(\ell)}\left(\mathbf{W}^{(\ell)}_\eta\mathbf{y}^{(\ell-1)} + \mathbf{b}^{(\ell)}_\eta\right). \tag*{(12)}

We call yη(ℓ)\mathbf{y}^{(\ell)}_\eta the η\etath neural unit of the ℓ\ellth layer y(ℓ)\mathbf{y}^{(\ell)}, where ηu∈N\eta_u \in\mathbb{N} is the total number of neural units per layer. Wη(ℓ)\mathbf{W}^{(\ell)}_\eta and bη(ℓ)\mathbf{b}^{(\ell)}_\eta denote the weight matrix and bias vector of the η\etath neural unit in the ℓ\ellth layer y(ℓ)\mathbf{y}^{(\ell)}. Furthermore act⁡(ℓ)(⋅):R→R\operatorname{act}^{(\ell)}(\cdot) : \mathbb{R} \to\mathbb{R} is a non-linear activation function. All weights and biases of all layers y(ℓ)\mathbf{y}^{(\ell)} are assembled in

θ={(Wη(ℓ),bη(ℓ)); ℓ=1,…,nL, η=1,…,ηu},(13)\theta= \left\{ \left(\mathbf{W}^{(\ell)}_\eta,\mathbf{b}^{(\ell)}_\eta\right);\,\ell= 1,\ldots,n_L,\,\eta= 1,\ldots,\eta_u \right\}, \tag*{(13)}

including all parameters of the neural network N(x,t)\mathcal{N}(\mathbf{x},t). As a result, the notation N(x,t;θ)\mathcal{N}(\mathbf{x},t;\theta) highlights the dependence of a neural network’s output on the input and the current realization of the weights and biases. Figure 1 illustrates the network’s topology, a combination of layers, neural units, and activation functions.

Figure 1. Schematic representation of the neural network architecture N(x,t;θ)\mathcal{N}(\boldsymbol{x},t;\boldsymbol{\theta}), starting with multi-input parameters x\boldsymbol{x} and tt at input layer y(0)\boldsymbol{y}^{(0)}, progressing through sequential hidden layer y(1),…,y(nL−1)\boldsymbol{y}^{(1)},\ldots,\boldsymbol{y}^{(n_L-1)}. The architecture concludes with an output layer y(nL)\boldsymbol{y}^{(n_L)}, producing the final output y\boldsymbol{y}.

The main idea of solving boundary value problems with an artificial neural network is the reformulation to an optimization problem [35, 50, 51], where the residual of the differential equations is to be minimized. To solve the differential equation (1) and (2), a suitable topology for the artificial neural network and, consequently, the physics-informed neural networks described in § 3.1 has to be chosen. Since (1) is stationary, we can reduce the artificial neural network to N(x;θ)\mathcal{N}(\boldsymbol{x};\boldsymbol{\theta}). Thus, we can define neural networks as an ansatz for the displacement and stress field i.e.

u(x,t)≈Nu(x;θu),(14)\boldsymbol{u}(\boldsymbol{x},t) \approx\mathcal{N}_{u}(\boldsymbol{x};\boldsymbol{\theta}_{u}), \tag*{(14)}
σ(x,t)≈Nσ(x;θσ),(15)\boldsymbol{\sigma}(\boldsymbol{x},t) \approx\mathcal{N}_{\sigma}(\boldsymbol{x};\boldsymbol{\theta}_{\sigma}), \tag*{(15)}

with trainable network parameters θ:={θu,θσ}\boldsymbol{\theta} := \{\boldsymbol{\theta}_{u},\boldsymbol{\theta}_{\sigma}\}. Notably, there is no separate network for the strain tensor. The strain tensor is deduced using the kinematics and differentiation applied to the displacement network, i.e. ε=∇symNu(x;θu)\boldsymbol{\varepsilon} = \nabla^{\mathrm{sym}}\mathcal{N}_{u}(\boldsymbol{x};\boldsymbol{\theta}_{u}). The architecture of this artificial neural network is visualized in Figure 2.

Figure 2. Schematic representation of the neural network’s topology, illustrating the progression from input through multiple hidden layers, resulting in displacement components uiu_i and stress components σij\sigma_{ij}.

Using the neural network ansatz we can rewrite the physics (1) and (2) as

RΩ=∇⋅Nσ(x,θσ)−f(x),in Ω,RΓD=Nu(x,θu)−g(x),on ΓD,RΓN=Nσ(x,θσ)⋅n(x)−t(x),on ΓN,(16)\begin{aligned} R_{\Omega} &= \nabla\cdot\mathcal{N}_{\sigma}(\mathbf{x},\boldsymbol{\theta}_{\sigma})-\mathbf{f}(\mathbf{x}), &&\text{in }\Omega,\\ R_{\Gamma_D} &= \mathcal{N}_{u}(\mathbf{x},\boldsymbol{\theta}_{u})-\mathbf{g}(\mathbf{x}), &&\text{on }\Gamma_D,\\ R_{\Gamma_N} &= \mathcal{N}_{\sigma}(\mathbf{x},\boldsymbol{\theta}_{\sigma})\cdot\mathbf{n}(\mathbf{x})-\mathbf{t}(\mathbf{x}), &&\text{on }\Gamma_N, \tag*{(16)} \end{aligned}

where RPDER_{\mathrm{PDE}} penalizes the residual of the equilibrium equation, and the equations RΓDR_{\Gamma_D} and RΓNR_{\Gamma_N} describe the discrepancy of the Dirichlet and Neumann boundary conditions. Notice that if Nu\mathcal{N}_{u} and Nσ\mathcal{N}_{\sigma} is a solution to the original boundary value problem, they minimize the differential equation-based residuals.

The parameters θ\boldsymbol{\theta} of the networks can be found by incorporating the physics-induced residuals into the training process of a neural network as components of the loss function. For this, we use a collocation method discretizing the domain Ω\Omega and the boundary Γ:=ΓD∪ΓN\Gamma:= \Gamma_D \cup\Gamma_N into sets of sample points SΩS_{\Omega} and SΓS_{\Gamma} with cardinalities ∣SΩ∣\lvert S_{\Omega}\rvert and ∣SΓ∣\lvert S_{\Gamma}\rvert. Then, an optimization problem to find the optimal parameters θ⋆\boldsymbol{\theta}^{\star}, also called training, is defined as

θ⋆=arg⁡min⁡θ LC(17)\boldsymbol{\theta}^{\star}=\underset{\boldsymbol{\theta}}{\arg\min}\,L_C \tag*{(17)}

with LC:=LΩ(x,θ)+LΓ(x,θ)L_C := L_\Omega(\mathbf{x},\boldsymbol{\theta}) + L_\Gamma(\mathbf{x},\boldsymbol{\theta}) given by the local losses

LΩ=1∣SΩ∣∑x∈SΩ∥RΩ(x;θ)∥22(18)L_\Omega= \frac{1}{\lvert S_\Omega\rvert}\sum_{\mathbf{x}\in S_\Omega}\lVert R_\Omega(\mathbf{x};\boldsymbol{\theta})\rVert_2^2 \tag*{(18)}
LΓ=1∣SΓ∣(∑x∈SΓD∥RΓD(x;θ)∥22+∑x∈SΓN∥RΓN(x;θ)∥22).(19)L_\Gamma= \frac{1}{\lvert S_\Gamma\rvert}\left(\sum_{\mathbf{x}\in S_{\Gamma_D}}\lVert R_{\Gamma_D}(\mathbf{x};\boldsymbol{\theta})\rVert_2^2 + \sum_{\mathbf{x}\in S_{\Gamma_N}}\lVert R_{\Gamma_N}(\mathbf{x};\boldsymbol{\theta})\rVert_2^2\right). \tag*{(19)}

The expressions penalize the residual of the governing equations and the discrepancy of the Dirichlet and Neumann boundary conditions, respectively. Notice that in the three-dimensional setting, one defines the neural networks as tuples, i.e.

Nu(x;θu)={Nui(x;θui)∣l=1,2},(20)\mathcal{N}_u(\mathbf{x};\boldsymbol{\theta}_u)=\{\mathcal{N}_{u_i}(\mathbf{x};\boldsymbol{\theta}_{u_i})\mid l=1,2\}, \tag*{(20)}
Nσ(x;θσ)={Nσij(x;θσij)∣i,j=1,2,3 and ij=ji},(21)\mathcal{N}_\sigma(\mathbf{x};\boldsymbol{\theta}_\sigma)=\{\mathcal{N}_{\sigma_{ij}}(\mathbf{x};\boldsymbol{\theta}_{\sigma_{ij}})\mid i,j=1,2,3\ \text{and}\ ij=ji\}, \tag*{(21)}

including the three components uiu_i of displacement u\mathbf{u} and six stress components σij\sigma_{ij}, where ij=jiij=ji ensures the symmetry of the stress tensor σ\boldsymbol{\sigma}. Figure 3 illustrates the complete network’s structure.

Figure 3. Schematic representation of the physics-informed neural network’s topology, illustrating the progression from input through multiple hidden layers, resulting in displacement components uiu_i and stress components σij\sigma_{ij}. These outputs undergo automatic differentiation to compute a physics-based loss function LCL_C. The loss is minimized using an optimizer to refine the network parameters θ\boldsymbol{\theta}.

While the PINN framework provides a straightforward method to solve physical-enhanced problems, it has challenges. Notably, there have been instances where the optimization yields solutions with unexpected or non-physical behaviors even when carefully tailored to encapsulate the physics [31]. Additionally, the current PINN formulation must minimize the difference between the network’s outputs and the available strain-stress data D\mathcal{D} due to the nature of the data-driven distance minimization problem (4). If we integrate the distance as an additional loss into the global loss, the whole problem becomes a nested optimization, leading to training challenges. The neural network could optimize in an undesired direction during each training epoch. If the approximated strain-stress point is not accurate, the corresponding data point might be suboptimal concerning the optimization algorithm, further complicating the learning process. To address these challenges, we consider the integration of PINNs with generative adversarial networks. GANs are proficient at generating outputs with the same properties as actual data, providing a potential approach to generating realistic strain-stress solutions. Their flexibility ensures adaptability across diverse data types suited for various physical conditions. Moreover, the inherent capability of GANs to discern and capitalize on intricate patterns may lead to a more robust representation of underlying physics. Additionally, with conditional GANs, generating outputs based on specific conditions becomes feasible, allowing for more targeted solutions. The combined PINN-GAN approach seeks to ensure physical consistency and alignment with observed data, leveraging the strengths of both methodologies. For clarity, we will provide a brief overview of GAN theory in the following.

Intermezzo to generative adversarial networks

Introduced by Goodfellow et al. [25], generative adversarial networks illustrate a novel approach to generating data using neural architectures. These networks comprise two distinctive neural entities: the generator (GG) and the discriminator (DD). The underlying goal of a GAN is to generate data instances that emulate the properties of actual data. The generation is achieved by setting the two networks against each other in a competitive game, often described as a dual-player minimax game.

Taking reference from the definitions provided in (9), we define the real data space as Dreal⊂Rdy\mathbb{D}_{\mathrm{real}} \subset\mathbb{R}^{d_y}, where dyd_y is the dimension of the space, i.e., dy=dim⁡(Dreal)d_y = \dim(\mathbb{D}_{\mathrm{real}}). The main objective of GANs is to produce synthetic data denoted as ysyn\mathbf{y}_{\mathrm{syn}}, residing in the same space as our real data yreal\mathbf{y}_{\mathrm{real}}. The generator can be defined as a function G:Rdx→RdyG : \mathbb{R}^{d_x} \to\mathbb{R}^{d_y}, which transforms a random noise vector x\mathbf{x} into synthetic data ysyn\mathbf{y}_{\mathrm{syn}}. In contrast, the discriminator operates as a function D:Rdy→RD : \mathbb{R}^{d_y} \to\mathbb{R}, that provides a measure of authenticity for a given data sample. Mathematically, these networks can be illustrated as:

G:Rdx→Rdy,D:Rdy→[0,1]x↦NG(x;θG),y↦ND(y;θD).(22)\begin{aligned} G &: \mathbb{R}^{d_x} \to\mathbb{R}^{d_y}, & \qquad D &: \mathbb{R}^{d_y} \to[0,1] \\ \mathbf{x} &\mapsto\mathcal{N}_G(\mathbf{x};\theta_G), & \qquad\mathbf{y} &\mapsto\mathcal{N}_D(\mathbf{y};\theta_D). \tag*{(22)} \end{aligned}

Here, NG(x;θG)\mathcal{N}_G(\mathbf{x};\boldsymbol{\theta}_G) and ND(y;θD)\mathcal{N}_D(\mathbf{y};\boldsymbol{\theta}_D) describe the neural networks with their corresponding trainable parameters θG\boldsymbol{\theta}_G and θD\boldsymbol{\theta}_D. The adversarial game between the generator and the discriminator during training can be encapsulated in the following objective

L(G,D)=Ey∼pdata[ln⁡D(y)]+Ex∼px[ln⁡(1−D(G(x)))](23)L(G,D)=\mathbb{E}_{\mathbf{y}\sim p_{\mathrm{data}}}[\ln D(\mathbf{y})]+\mathbb{E}_{\mathbf{x}\sim p_{\mathbf{x}}}[\ln(1-D(G(\mathbf{x})))] \tag*{(23)}

leading to the optimization:

min⁡Gmax⁡DL(G,D)(24)\min_G\max_D L(G,D) \tag*{(24)}

where EE represents a random variable’s expectation or expected value. It provides a weighted average of a function concerning its probability distribution. Specifically,

Ey∼pdata[ln⁡D(y)](25)\mathbb{E}_{\mathbf{y}\sim p_{\mathrm{data}}}[\ln D(\mathbf{y})] \tag*{(25)}

represents the average logarithmic score assigned by the discriminator to actual data samples drawn from the distribution pdatap_{\mathrm{data}}. On the other hand, the expression

Ex∼px[ln⁡(1−D(G(x)))](26)\mathbb{E}_{\mathbf{x}\sim p_{\mathbf{x}}}[\ln(1-D(G(\mathbf{x})))] \tag*{(26)}

reflects the average logarithmic score the discriminator accords to the synthetic or generated data, which is created from a random noise vector x\mathbf{x} following the noise distribution pxp_{\mathbf{x}}.

The competition between the two networks is straightforward: the generator GG aims to produce data that the discriminator DD cannot distinguish from accurate data. In contrast, the discriminator tries to better distinguish real data from fake data produced by GG. The probability distributions pdatap_{\mathrm{data}} and pxp_{\mathbf{x}} depict the actual data and noise distributions, respectively. The terms in the objective function essentially capture the average confidence levels of the discriminator in judging the authenticity of both original and fake data samples. The procedure of the GAN’s interplay between the generator and the discriminator is illustrated in Figure 4.

Figure 4. Schematic representation of a generative adversarial network (GAN) showcasing the interaction between the generator producing data from random input and the discriminator evaluating the authenticity of both real and generated data.

Traditional GANs deploy a sigmoid activation function for the discriminator’s final layer, ensuring its outputs fall within [0,1][0,1]. The GANs can suffer from issues like mode collapse (where the generator generates limited varieties of samples), vanishing gradients, and general training instability. To address some of these challenges, the Wasserstein GAN (WGAN) [52] changes the objective function to leverage the Wasserstein distance [53]. The WGAN objective can be described as:

LWGAN(G,D)=Ey∼pdata[D(y)]−Ex∼px[D(G(x))],(27)L_{WGAN}(G,D) = \mathbb{E}_{y\sim p_{\mathrm{data}}}[D(y)] - \mathbb{E}_{x\sim p_x}[D(G(x))], \tag*{(27)}

leading to the following optimization:

min⁡Gmax⁡DLWGAN(G,D).(28)\min_G \max_D L_{WGAN}(G,D). \tag*{(28)}

WGANs are known to provide more stable and consistent training dynamics [52]. Building on the WGAN, the Wasserstein GAN with Gradient Penalty (WGAN-GP) introduced a regularization term to ensure that the discriminator’s gradients remain bounded [54]. This gradient penalty aims to enforce the Lipschitz continuity condition, which addresses the vanishing gradient problem. The gradient penalty is defined as:

GP=E[(∥∇y~D(y~)∥2−1)2],(29)\mathrm{GP} = \mathbb{E}\left[\left(\lVert\nabla_{\tilde{y}}D(\tilde{y})\rVert_2 - 1\right)^2\right], \tag*{(29)}

where y~=δyreal+(1−δ)ysyn\tilde{y} = \delta y_{\mathrm{real}} + (1-\delta)y_{\mathrm{syn}} and δ\delta is sampled from a uniform distribution in [0,1][0,1]. The optimization for WGAN-GP thus becomes

min⁡Gmax⁡DLWGAN(G,D)+ω⋅GP,(30)\min_G \max_D L_{WGAN}(G,D) + \omega\cdot\mathrm{GP}, \tag*{(30)}

where ω∈R+\omega\in\mathbb{R}_{+} is a hyperparameter determining the weight of the gradient penalty in the overall objective [55].

Physics-informed GANs for data-driven mechanics problems

In the classical data-driven computational mechanics paradigm § 2, the objective is to find the closest point z\mathbf{z} in the constraint set C\mathcal{C} to z^\hat{\mathbf{z}} in the material dataset D\mathcal{D}, as formalized in equation (4). This context motivates our modified GAN approach for data-driven mechanics problems. To utilize GANs for solving differential equations in a data-driven mechanics setting, we propose a novel approach wherein the generator in the GAN architecture is identified as a physics-informed neural network (PINN). In this paradigm, while the generator outputs plausible solutions adhering to the underlying physics, the discriminator is trained to distinguish between the generator’s predictions and actual strain-stress data.

In the conventional GAN setup from § 3.2, the generator GG maps the input vector x\mathbf{x} into synthetic data, ysyn\mathbf{y}_{\mathrm{syn}}. Instead of treating x\mathbf{x} as a random noise vector, it represents the collocation points x\mathbf{x} in the domain SΩS_{\Omega}. Thus, the generator is formalized as a mapping G:SΩ→(Nu,Nσ)G:S_{\Omega}\to(\mathcal{N}_{u},\mathcal{N}_{\sigma}), where Nu\mathcal{N}_{u} and Nσ\mathcal{N}_{\sigma} represents the neural network approximation for the displacement and stress field, respectively. Therefore, the generator can be defined as:

G(x,θG):=(Nu(x;θu),Nσ(x;θσ))(31)G(\mathbf{x},\boldsymbol{\theta}_{G}) := \left(\mathcal{N}_{u}(\mathbf{x};\boldsymbol{\theta}_{u}),\mathcal{N}_{\sigma}(\mathbf{x};\boldsymbol{\theta}_{\sigma})\right) \tag*{(31)}

where θG:=(θu,θσ)\boldsymbol{\theta}_{G}:=(\boldsymbol{\theta}_{u},\boldsymbol{\theta}_{\sigma}) denotes the trainable parameters of the generator network. Building upon the physics-informed aspect, we differentiate Nu\mathcal{N}_{u} and employ the kinematics equation to obtain the strain ε\boldsymbol{\varepsilon}. Given ε=∇symNu\boldsymbol{\varepsilon}=\nabla^{\mathrm{sym}}\mathcal{N}_{u}, the generator’s output evolves from merely the neural network predictions Nu\mathcal{N}_{u} and Nσ\mathcal{N}_{\sigma} to the strain-stress pair z:=(ε,Nσ)\mathbf{z}:=(\boldsymbol{\varepsilon},\mathcal{N}_{\sigma}).

Once we obtain the strain-stress output from the generator, to stay consistent with the data-driven mechanics’ paradigm, we compute the strain-stress data points z^∈D\hat{\mathbf{z}}\in\mathcal{D} closest to the output z\mathbf{z}, which corresponds to:

z^=arg min⁡z^∈D d(z,z^),(32)\hat{\mathbf{z}}=\underset{\hat{\mathbf{z}}\in\mathcal{D}}{\operatorname{arg\,min}}\,d(\mathbf{z},\hat{\mathbf{z}}), \tag*{(32)}

with distance (7). We then use z\mathbf{z} and z^\hat{\mathbf{z}} as synthetic and real data for the discriminator’s training. For the discriminator D(y,θD)D(\mathbf{y},\boldsymbol{\theta}_{D}), we establish the mapping D:R2d→[0,1]D:\mathbb{R}^{2d}\to[0,1], aligning with the conventional GAN framework. To accommodate strain-stress pairs as inputs for the discriminator, we convert a pair into a 2d2d-vector y\mathbf{y} by applying Voigt-Notation to both the strain and stress, then merging them into a single vector. Given strain-stress data z^∈D\hat{z} \in D, it assesses the data’s authenticity, furnishing scores to guide the generator’s training. With the generator now representing a PINN, the adversarial loss in equation (23) has to integrate the physics-informed loss LCL_C, derived from the residuals of the governing differential equations:

L(G,D)=Ez^∼pD[ln⁡D(z^)]+Ex∼pSΩ[ln⁡(1−D(G(x)))+LC].(33)L(G,D)=\mathbb{E}_{\hat{z}\sim p_D}[\ln D(\hat{z})]+\mathbb{E}_{x\sim p_{S_\Omega}}[\ln(1-D(G(x)))+L_C]. \tag*{(33)}

The collaborative training between the discriminator and the physics-informed generator ensures that the latter learns to craft data that confounds the discriminator and aligns closely with intrinsic physics. Figure 5 illustrated the physics-enhanced GAN approach for the data-driven mechanics problem. Regarding Wasserstein GANs and their gradient penalty variants, their objectives concerning the physics-informed generator must be modified. For instance, with the Wasserstein GAN objective, the loss function becomes:

Figure 5. Schematic representation of a physics-informed generative adversarial network (GAN) incorporating collocation points and strain-stress data for physical point generation and discrimination.

LWGAN(G,D)=Ez^∼pD[D(z^)]−Ex∼pSΩ[D(G(x))+LC],(34)L_{\mathrm{WGAN}}(G,D)=\mathbb{E}_{\hat{z}\sim p_D}[D(\hat{z})]-\mathbb{E}_{x\sim p_{S_\Omega}}[D(G(x))+L_C], \tag*{(34)}

Moreover, for the WGAN-GP, the combined objective is:

L(G,D)=LWGAN+ω⋅GP.(35)L(G,D)=L_{\mathrm{WGAN}}+\omega\cdot\mathrm{GP}. \tag*{(35)}

By incorporating GANs with physics-informed principles, the models produce data that adheres to the statistics of observed datasets and the underlying differential equations. This integration addresses the nested optimization issue commonly found in the PINN-based data-driven mechanics. With the capability of GANs to generate outputs mirroring accurate data, the solutions are both statistically relevant and in line with physical principles. Using GANs simplifies the optimization process, making the training more stable and less prone to errors from inaccurate strain-stress approximations. However, it is worth noting that the loss values obtained while training a traditional GAN are often unreliable. In many studies, qualitative and quantitative evaluation methods are employed to assess the performance of the GAN. Qualitative evaluations, while offering a quick visual validation, can be subjective. Typically, they involve human observers who evaluate the realism of a generated sample. The overall presumption has been that if the generated sample appears realistic, the GAN’s training is deemed successful, regardless of potential fluctuations in loss values. Nevertheless, such evaluations can be biased and do not always represent the complete performance spectrum of the GAN. For instance, the generated samples might still appear high quality even in scenarios where mode collapse occurs. Considering these challenges, especially in the context of our work where the goal is not generating images but accurately representing strain-stress states, we decided on WGAN + GP. Unlike traditional GANs, the loss of WGANs has a convergence point. Ideally, this point is reached when the generator is so adept at producing samples that no Lipschitz continuous discriminator can differentiate between real and generated samples. This characteristic of WGAN provides a more stable and consistent evaluation metric, ensuring that the generated strain-stress states are physically accurate. The effectiveness of this method will be showcased in a two-dimensional numerical example.

Numerical benchmark of a non-Linear elastic plate with hole

This section illustrates the application of GANs to the data-driven computing paradigm [2] in a typical benchmark, considering stress analysis of non-linear elastic material. We discuss the problem setup and test environments and give a proper definition of the geometry and boundary conditions and the material parameters for data generation. We limit the simulation to noiseless synthetic data sets, which consist of strain-stress points created numerically using a material model rather than obtained by actual experimental measurements. However, experimental data is often noisy and contains outliers. This issue can be addressed with noise reduction algorithms such as tensor voting [56], Kalman filtering [57], and deep learning-based techniques.

In this benchmark, we investigate a 2d2d in-plain plate with a hole subject to a distributed force. The geometry, boundary conditions, and displacements are chosen according to a similar test presented in [16] and illustrated in Figure 6.

Figure 6. Illustration of a square plate subjected to external forces, alongside its top-right quadrant representing the symmetry section with specified boundary conditions and force distribution t(y)t(y) applied.

Geometry:

The system is defined by Ω=[−ℓ2,ℓ2]2∖Br(0)\Omega= \left[-\frac{\ell}{2},\frac{\ell}{2}\right]^2 \setminus B_r(0), where BrB_r refers to the open ball of radius r=ℓ4r = \frac{\ell}{4} centered at the origin (0,0)(0,0). The side lengths of the plate are equal to ℓ=2m\ell= 2\mathrm{m}. Due to the symmetry of geometry, only one-quarter of the system is simulated, cf. Figure 6.

Displacements are fixed at the quarter plate’s left surface x=0x = 0 in xx-direction and at the bottom surface y=0y = 0 in yy-direction. The corresponding conditions read as follows:

{ux=0,if x=0;uy=0,if y=0;(36)\begin{cases} u_x = 0, & \text{if } x = 0;\\ u_y = 0, & \text{if } y = 0; \tag*{(36)} \end{cases}

where σx\sigma_x is the stress and uxu_x and uyu_y are the displacements in xx and yy-directions, respectively. In addition, we define boundary conditions for the stress, especially for x=ℓ2x = \frac{\ell}{2} the plate is subjected to a distributed force t(y)=200cos⁡(πy2)t(y) = 200\cos\left(\frac{\pi y}{2}\right) in xx-direction. The boundary conditions for the stress components read

{σxx=t(y),if x=1;σyy=0,if y=1;σxy=0,if (x,y)∈∂Ω.(37)\begin{cases} \sigma_{xx} = t(y), & \text{if } x = 1;\\ \sigma_{yy} = 0, & \text{if } y = 1;\\ \sigma_{xy} = 0, & \text{if } (x,y) \in\partial\Omega. \tag*{(37)} \end{cases}

Notice that numerical methods based on the weak form of a boundary value problem innately satisfy shear-free boundary conditions on free boundaries. However, our PINN approach utilizes the strong form of the boundary value problem, so it is crucial to impose the zero stress boundary conditions directly [45].

To train the network, we utilize 1282128^2 quasi-random points produced using the Sobol sequence [58]. For testing, we generate 2562256^2 domain points using a uniform random distribution.

Material parameters: The boundary value problem considers the non-linear elastic material behavior of [16] defined by

σ=λg(tr⁡(ε))I+με+Cε,(38)\boldsymbol{\sigma} = \lambda g(\operatorname{tr}(\boldsymbol{\varepsilon}))\boldsymbol{I} + \mu\boldsymbol{\varepsilon} + \boldsymbol{C}\boldsymbol{\varepsilon}, \tag*{(38)}

with g(x)=((∣x∣+a)p−ap)sgn⁡(x)g(x) = ((|x| + a)^p - a^p)\operatorname{sgn}(x) and a,p∈Ra,p \in\mathbb{R}. The applied material parameters are Young’s modulus EE, Poisson’s ratio ν\nu, and orthotropic elasticity tensor for plane strain given by

C=(C112ν(λˉ+G⊥)02ν(λˉ+G⊥)λˉ+2G⊥000G∥),(39)\boldsymbol{C} = \begin{pmatrix} C_{11} & 2\nu(\bar{\lambda} + G_{\perp}) & 0 \\ 2\nu(\bar{\lambda} + G_{\perp}) & \bar{\lambda} + 2G_{\perp} & 0 \\ 0 & 0 & G_{\parallel} \end{pmatrix}, \tag*{(39)}

where λ=Eν(1+ν)(1−2ν)\lambda= \frac{E\nu}{(1+\nu)(1-2\nu)} and μ=E2(1+ν)\mu= \frac{E}{2(1+\nu)} are the well known Lamé constants and C11=4.6875EC_{11} = 4.6875E, G⊥=0.3EG_{\perp} = 0.3E, G∥=0.2EG_{\parallel} = 0.2E and λˉ=2ν2+115−20ν2E\bar{\lambda} = \frac{2\nu^2+1}{15-20\nu^2}E are additional material parameters. The exact parameter values used for the reference solution and synthetic data are given in Table 1.

EE [MPa]ν\nu [−]aa [−]pp [−]
1×1041 \times10^40.30.30.0010.0010.0050.005

Table 1. Material parameters

Synthetic data: In order to simulate actual experimental measurements, we generate data artificially using the non-linear material model (38) based on the given material parameters. We investigate normal data distributions of 1003100^3 strain-stress data points with a fixed random seed. The data is created by a zero-mean normal distribution with a standard deviation of 0.0050.005 in all strain dimensions.

WGAN parameter:

For the adversarial network, the model consists of the generator and the discriminator setup. The generator

G(x,θG)={Nu(x,θu),Nσ(x,θσ)}(40)G(\mathbf{x},\theta_G)=\{\mathcal{N}_u(\mathbf{x},\theta_u),\mathcal{N}_{\sigma}(\mathbf{x},\theta_{\sigma})\} \tag*{(40)}

with x=(x,y)\mathbf{x}=(x,y), θG=(θu,θσ)\theta_G=(\theta_u,\theta_{\sigma}) and

Nu(x,θu)={Nui(x,θui)∣i=x,y},(41)\mathcal{N}_u(\mathbf{x},\theta_u)=\{\mathcal{N}_{u_i}(\mathbf{x},\theta_{u_i})\mid i=x,y\}, \tag*{(41)}
Nσ(x,θσ)={Nσij(x,θσij)∣i,j=x,y},(42)\mathcal{N}_{\sigma}(\mathbf{x},\theta_{\sigma})=\{\mathcal{N}_{\sigma_{ij}}(\mathbf{x},\theta_{\sigma_{ij}})\mid i,j=x,y\}, \tag*{(42)}

is constructed with a series of fully connected layers. The architecture utilizes 4 hidden layers, each with 64 neurons. The activation function used across these layers is the Swish function, defined as

Hardswish⁡(x)={0if x≤−3,xif x≥3,x2+3x6otherwise.(43)\operatorname{Hardswish}(x)= \begin{cases} 0 & \text{if } x\leq-3,\\ x & \text{if } x\geq3,\\ \frac{x^2+3x}{6} & \text{otherwise}. \end{cases} \tag*{(43)}

In addition, to optimize the network training, we hard enforce the boundary conditions from (36) and (37), such that the output of the generator is given by

Nux(x;θux)=x⋅N^ux(x;θux),Nuy(x;θuy)=y⋅N^ux(x;θuy),Nσxx(x;θσxx)=(1−x)⋅N^σxx(x;θσxx),Nσyy(x;θσyy)=(1−y)⋅N^σyy(x;θσyy),Nσxy(x;θσxy)=xy(x2+y2−0.25)⋅N^σxy(x;θσxy),(44)\begin{aligned} \mathcal{N}_{u_x}(\mathbf{x};\theta_{u_x})&=x\cdot\widehat{\mathcal{N}}_{u_x}(\mathbf{x};\theta_{u_x}),\\ \mathcal{N}_{u_y}(\mathbf{x};\theta_{u_y})&=y\cdot\widehat{\mathcal{N}}_{u_x}(\mathbf{x};\theta_{u_y}),\\ \mathcal{N}_{\sigma_{xx}}(\mathbf{x};\theta_{\sigma_{xx}})&=(1-x)\cdot\widehat{\mathcal{N}}_{\sigma_{xx}}(\mathbf{x};\theta_{\sigma_{xx}}),\\ \mathcal{N}_{\sigma_{yy}}(\mathbf{x};\theta_{\sigma_{yy}})&=(1-y)\cdot\widehat{\mathcal{N}}_{\sigma_{yy}}(\mathbf{x};\theta_{\sigma_{yy}}),\\ \mathcal{N}_{\sigma_{xy}}(\mathbf{x};\theta_{\sigma_{xy}})&=xy(x^2+y^2-0.25)\cdot\widehat{\mathcal{N}}_{\sigma_{xy}}(\mathbf{x};\theta_{\sigma_{xy}}), \tag*{(44)} \end{aligned}

with x=(x,y)\mathbf{x}=(x,y) and θ=(θui,θσij)\theta=(\theta_{u_i},\theta_{\sigma_{ij}}) being the tuple of all trainable network parameters regarding the displacement and stress component. In order to obtain the strains and optimize the loss function, the spatial derivatives are obtained by automatic differentiation.

On the other hand, the discriminator D(y,θD)D(\mathbf{y},\theta_D) comprises a network architecture of 3 hidden layers, each with 16 neurons, which uses the LeakyReLU activation function defined as

LeakyReLU⁡(x)={xif x≥0,αxif x<0,(45)\operatorname{LeakyReLU}(x)= \begin{cases} x & \text{if } x\geq0,\\ \alpha x & \text{if } x<0, \end{cases} \tag*{(45)}

with a slope of α=0.2\alpha= 0.2 for negative values. Regarding optimization, both the generator and the discriminator use the ADAM optimizer with a learning rate of 0.02. The beta values for the moment estimates are set as (0.5, 0.999). A learning rate scheduler is employed with a maximum learning rate of 0.02. It is set to adjust the rate over a total of 200 steps for both the Generator and Discriminator.

Result:

The WGAN+GP frameworks were utilized in our numerical evaluations to investigate their effectiveness in computing non-linear elastic materials through a data-driven approach. Figure 7 depicts the distribution of strain-stress achieved after 200 training epochs. Due to the utilization of batch processing during training, the number of training steps exceeded this epoch count. The findings provide a profound understanding of the training quality and effectiveness.

Visualization of displacement and stress distribution after 200 training epochs

Figure 7. Visualization of displacement and stress distribution after 200 training epochs offering insights into the material’s behavior under the applied loads and conditions (36) and (37). From top-left to bottom-right: uxu_x showcases a gradient, indicating a maximum displacement at (x,y)=(1,0)(x, y) = (1, 0); uyu_y reveals a displacement trend with negative values highlighted in x=0x = 0; σx\sigma_x shows a maximum stress magnitude at y=0y = 0; σy\sigma_y displays a similar gradient; and σxy\sigma_{xy} captures a pronounced shear stress distribution inside the plate.

We investigated the loss values during training for a clearer perspective on model behavior. The Wasserstein-enhanced architectures showcased robustness and consistency during training. Figure 8 displays the losses for the discriminator and generator of the model across the epochs. Notably, shallow loss values for either the generator or the discriminator can be counterproductive. It generally indicates that one network is dominating the other, leading to a stagnation in the training process. Ideally, there should be a balance where both networks challenge each other, encouraging continuous improvement. Despite their effectiveness, traditional adversarial networks present non-interpretable loss values, making it challenging to discern training quality. The WGAN+GP approach offers direct insights into the quality of data generation, making it more user-friendly in solution analysis. In addition, we plot the minimum distance between the generated states and the data set D\mathcal{D} in Figure 9. Given the data-driven approach, the learning process trains with a finite set of data points. Consequently, the losses do not converge to zero but to a positive lower bound. In our case, both losses decrease over training, representing this convergence. In data-driven mechanics, the approach displayed a commendable ability to simulate stress-strain distributions. WGAN, with improved loss interpretability and smoother training, stands out as the preferable choice for intricate computational mechanics tasks.

Figure 8. Comparative visualization of the discriminator and generator loss metrics over training iterations for a WGAN+GP model, showcasing the dynamic interplay and convergence patterns. The shaded area shows the maximum range of loss for individual training batches.

Figure 9. Visualization of the distance metrics over 200 epochs. The distance of the generated z\mathbf{z} to the dataset D\mathcal{D} illustrates how closely the model-generated outputs match the dataset over training iterations.

Conclusion

The model-free data-driven method, developed by Kirchdoerfer and Ortiz, uses experimental data directly in simulations, bypassing the entire material modeling step. The paradigm uses nearest-neighbor clustering to reformulate boundary value problems. The approach has been diversified for many applications. Challenges such as data availability, noise, inconsistency, and high dimensionality frequently arise in the data-driven paradigm. Traditional analytical and computational methods may need to be adjusted when addressing these issues.

Consequently, the incorporation of machine learning methods is considered, especially physics-informed neural networks. In solving boundary value problems with ANNs, the idea is to transform it into an optimization problem. The residual of the differential equations is minimized, and the neural network approximates the displacement and stress field. However, there are challenges with PINNs. There have been instances where the optimization yields solutions with unexpected or non-physical behaviors even when carefully tailored to encapsulate the physics. If we integrate the distance as an additional loss into the global loss, the whole problem becomes a nested optimization, leading to training challenges. In addition, approximated strain-stress fields can correspond to suboptimal data points influencing the direction and rate of the convergence.

To address these challenges, we consider the integration of PINNs with generative adversarial networks. GANs are proficient at generating outputs with the same properties as actual data, providing a potential approach to generating realistic strain-stress solutions. Their flexibility ensures adaptability across diverse data types suited for various physical conditions. Moreover, the inherent capability of GANs to distinguish and capitalize on intricate patterns may lead to a more robust representation of underlying physics. The combined PINN-GAN approach seeks to ensure physical consistency and alignment with observed data, leveraging the strengths of both methodologies.

This research introduced an approach to WGANs + GP tailored for data-driven mechanics problems. The generator is identified as a PINN, ensuring that generated outputs conform to underlying physical principles. Instead of random noise, the generator utilizes collocation points from the domain and maps them to neural network approximations of strain and stress fields. The discriminator is then trained using the generated and the closest actual strain-stress data. By integrating WGANs with physics-informed principles, the model outputs adhere to observed dataset statistics and differential equations. This results in improved optimization, more stable training, and accurate, physically consistent solutions. In this regard, we investigated a non-linear elastic plate with a hole benchmark. The results indicate that our proposed method provides reasonable outcomes. Furthermore, we observed robust and consistent training of the networks and noted the convergence of the data-driven solution as data size increased.

As we advance our research, we aim to delve deeper into other convergence criteria for the GAN or WGAN. We plan to explore metrics such as the Inception Score [59], Frechet Inception Distance [60], and perceptual similarity measures [61] to provide a broader assessment of the generated outputs. These metrics will help to analyze the quality of the generated material states.

Another area of interest is using the discriminator in the GAN framework for material identification. The discriminator’s ability to distinguish between actual and generated outputs can be used to identify different material states. This approach could offer a novelty to classify materials, and we want to explore this further.

In addition, we plan to extend our method to more complex and varied material properties. We also consider integrating advanced machine learning techniques to improve prediction accuracy, especially when dealing with sparse datasets. We are considering hybrid network architectures that combine convolutional and regression layers. The traditional image-based GAN structure inspires this design. By adding these layers, we hope to combine the advantages of image-based GANs with our current data-focused method.

References

  1. [1]S. Timoshenko, History of strength of materials: with a brief account of the history of theory of elasticity and theory of structures. Courier Corporation, 1983.
  2. [2]T. Kirchdoerfer and M. Ortiz, “Data-driven computational mechanics,” Computer Methods in Applied Mechanics and Engineering, vol. 304, pp. 81–101, 2016.DOI
  3. [3]T. Kirchdoerfer and M. Ortiz, “Data driven computing with noisy material data sets,” Computer Methods in Applied Mechanics and Engineering, vol. 326, pp. 622–641, 2017.DOI
  4. [4]S. Conti, S. Müller, and M. Ortiz, “Data-driven problems in elasticity,” Archive for Rational Mechanics and Analysis, vol. 229, p. 79–123, Jan 2018.DOI
  5. [5]L. T. K. Nguyen and M.-A. Keip, “A data-driven approach to nonlinear elasticity,” Computers & Structures, vol. 194, pp. 97–115, 2018.DOI
  6. [6]A. Galetzka, D. Loukrezis, and H. D. Gersem, “Data-driven solvers for strongly nonlinear material response,” ArXiv, vol. abs/2008.08482, 2020.DOI
  7. [7]T. Kirchdoerfer and M. Ortiz, “Data-driven computing in dynamics,” International Journal for Numerical Methods in Engineering, vol. 113, no. 11, pp. 1697–1710, 2018.arxiv.org/abs/1706.04061
  8. [8]A. Platzer, A. Leygue, L. Stainier, and M. Ortiz, “Finite element solver for data-driven finite strain elasticity,” Computer Methods in Applied Mechanics and Engineering, vol. 379, p. 113756, 2021.DOI
  9. [9]L. Stainier, A. Leygue, and M. Ortiz, “Model-free data-driven methods in mechanics: material data identification and solvers,” Computational Mechanics, pp. 1–13, 2019.DOI
  10. [10]K. Karapiperis, M. Ortiz, and J. Andrade, “Data-driven nonlocal mechanics: Discovering the internal length scales of materials,” Computer Methods in Applied Mechanics and Engineering, vol. 386, p. 114039, 2021.DOI
  11. [11]E. Marenić, G. Seychal, and J.-C. Passieux, “Data driven approach in multiphysics framework: Application to coupled electro-mechanical problems,” Computer Methods in Applied Mechanics and Engineering, vol. 395, p. 114959, 2022.DOI
  12. [12]S. Zschocke, F. Leichsenring, W. Graf, and M. Kaliske, “A concept for data-driven computational mechanics in the presence of polymorphic uncertain properties,” Engineering Structures, vol. 267, p. 114672, 2022.DOI
  13. [13]J. Yang, W. Huang, Q. Huang, and H. Hu, “An investigation on the coupling of data-driven computing and model-driven computing,” Computer Methods in Applied Mechanics and Engineering, vol. 393, p. 114798, 2022.DOI
  14. [14]R. Ibañez, D. Borzacchiello, J. V. Aguado, E. Abisset-Chavanne, E. Cueto, P. Ladeveze, and F. Chinesta, “Data-driven non-linear elasticity: constitutive manifold construction and problem discretization,” Computational Mechanics, vol. 60, no. 5, pp. 813–826, 2017.DOI
  15. [15]R. Ibañez, E. Abisset-Chavanne, J. V. Aguado, D. Gonzalez, E. Cueto, and F. Chinesta, “A manifold learning approach to data-driven computational elasticity and inelasticity,” Archives of Computational Methods in Engineering, vol. 25, no. 1, pp. 47–57, 2018.
  16. [16]R. Eggersmann, L. Stainier, M. Ortiz, and S. Reese, “Model-free data-driven computational mechanics enhanced by tensor voting,” Computer Methods in Applied Mechanics and Engineering, vol. 373, p. 113499, 2021.DOI
  17. [17]P. Mordohai and G. Medioni, “Dimensionality estimation, manifold learning and function approximation using tensor voting,” Journal of Machine Learning Research, vol. 11, pp. 411–450, 01 2010.DOI
  18. [18]R. Eggersmann, T. Kirchdoerfer, S. Reese, L. Stainier, and M. Ortiz, “Model-free data-driven inelasticity,” Computer Methods in Applied Mechanics and Engineering, vol. 350, pp. 81–99, 2019.DOI
  19. [19]K. Karapiperis, L. Stainier, M. Ortiz, and J. Andrade, “Data-driven multiscale modeling in mechanics,” Journal of the Mechanics and Physics of Solids, vol. 147, p. 104239, 2021.DOI
  20. [20]K. Ciftci and K. Hackl, “Model-free data-driven simulation of inelastic materials using structured data sets, tangent space information and transition rules,” Computational Mechanics, May 2022.arxiv.org/abs/2101.10730
  21. [21]B. C. Csáji et al., “Approximation with artificial neural networks,” Faculty of Sciences, Etvs Lornd University, Hungary, vol. 24, no. 48, p. 7, 2001.
  22. [22]Z. Lu, H. Pu, F. Wang, Z. Hu, and L. Wang, “The expressive power of neural networks: A view from the width,” Advances in neural information processing systems, vol. 30, 2017.arxiv.org/abs/1709.02540
  23. [23]K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 770–778, 2016.arxiv.org/abs/1512.03385
  24. [24]J. D. Lafferty, A. McCallum, and F. Pereira, “Conditional random fields: Probabilistic models for segmenting and labeling sequence data,” in International Conference on Machine Learning, 2001.
  25. [25]I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial nets,” Advances in neural information processing systems, vol. 27, 2014.
  26. [26]I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y. Bengio, “Generative adversarial networks,” Communications of the ACM, vol. 63, no. 11, pp. 139–144, 2020.
  27. [27]M. Raissi, A. Yazdani, and G. E. Karniadakis, “Hidden fluid mechanics: Learning velocity and pressure fields from flow visualizations,” Science, vol. 367, no. 6481, pp. 1026–1030, 2020.DOI
  28. [28]L. Sun, H. Gao, S. Pan, and J.-X. Wang, “Surrogate modeling for fluid flows based on physics-constrained deep learning without simula- tion data,” Computer Methods in Applied Mechanics and Engineering, vol. 361, p. 112732, 2020.DOI
  29. [29]Z. Mao, A. D. Jagtap, and G. E. Karniadakis, “Physics-informed neural networks for high-speed flows,” Computer Methods in Applied Mechanics and Engineering, vol. 360, p. 112789, 2020.DOI
  30. [30]A. Dourado and F. A. Viana, “Physics-informed neural networks for missing physics estimation in cumulative damage models: a case study in corrosion fatigue,” Journal of Computing and Information Science in Engineering, vol. 20, no. 6, 2020.DOI
  31. [31]J.-H. Bastek and D. M. Kochmann, “Physics-informed neural networks for shell structures,” European Journal of Mechanics-A/Solids, vol. 97, p. 104849, 2023.DOI
  32. [32]Q. He, D. Barajas-Solano, G. Tartakovsky, and A. M. Tartakovsky, “Physics-informed neural networks for multiphysics data assimilation with application to subsurface transport,” Advances in Water Resources, vol. 141, p. 103610, 2020.DOI
  33. [33]M. Yin, X. Zheng, J. D. Humphrey, and G. E. Karniadakis, “Non-invasive inference of thrombus material properties with physics-informed neural networks,” Computer Methods in Applied Mechanics and Engineering, vol. 375, p. 113603, 2021.DOI
  34. [34]I. E. Lagaris, A. Likas, and D. I. Fotiadis, “Artificial neural networks for solving ordinary and partial differential equations,” IEEE transactions on neural networks, vol. 9, no. 5, pp. 987–1000, 1998.DOI
  35. [35]M. Raissi, P. Perdikaris, and G. E. Karniadakis, “Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations,” Journal of Computational physics, vol. 378, pp. 686–707, 2019.DOI
  36. [36]E. Haghighat, M. Raissi, A. Moure, H. Gomez, and R. Juanes, “A physics-informed deep learning framework for inversion and surrogate modeling in solid mechanics,” Computer Methods in Applied Mechanics and Engineering, vol. 379, p. 113741, 2021.DOI
  37. [37]L. Yang, X. Meng, and G. E. Karniadakis, “B-pinns: Bayesian physics-informed neural networks for forward and inverse pde problems with noisy data,” Journal of Computational Physics, vol. 425, p. 109913, 2021.DOI
  38. [38]T. Kadeethum, T. M. Jørgensen, and H. M. Nick, “Physics-informed neural networks for solving inverse problems of nonlinear biot’s equations: Batch training,” in 54th US Rock Mechanics/Geomechanics Symposium, OnePetro, 2020.arxiv.org/abs/2005.09638
  39. [39]J. Platt and A. Barr, “Constrained differential optimization,” in Neural Information Processing Systems, 1987.
  40. [40]C.-L. Hwang and A. S. M. Masud, Multiple objective decision making—methods and applications: a state-of-the-art survey, vol. 164. Springer Science & Business Media, 2012.
  41. [41]S. Wang, Y. Teng, and P. Perdikaris, “Understanding and mitigating gradient flow pathologies in physics-informed neural networks,” SIAM Journal on Scientific Computing, vol. 43, no. 5, pp. A3055–A3081, 2021.arxiv.org/abs/2001.04536
  42. [42]A. D. Jagtap, K. Kawaguchi, and G. E. Karniadakis, “Adaptive activation functions accelerate convergence in deep and physics-informed neural networks,” Journal of Computational Physics, vol. 404, p. 109136, 2020.DOI
  43. [43]A. D. Jagtap, K. Kawaguchi, and G. Em Karniadakis, “Locally adaptive activation functions with slope recovery for deep and physics-informed neural networks,” Proceedings of the Royal Society A, vol. 476, no. 2239, p. 20200334, 2020.DOI
  44. [44]X. Jin, S. Cai, H. Li, and G. E. Karniadakis, “Nsfnets (navier-stokes flow nets): Physics-informed neural networks for the incompressible navier-stokes equations,” Journal of Computational Physics, vol. 426, p. 109951, 2021.DOI
  45. [45]A. Henkes, H. Wessels, and R. Mahnken, “Physics informed neural networks for continuum micromechanics,” Computer Methods in Applied Mechanics and Engineering, vol. 393, p. 114790, 2022.DOI
  46. [46]Y. Yang and P. Perdikaris, “Adversarial uncertainty quantification in physics-informed neural networks,” Journal of Computational Physics, vol. 394, pp. 136–152, 2019.DOI
  47. [47]A. Daw, M. Maruf, and A. Karpatne, “Pid-gan: A gan framework based on a physics-informed discriminator for uncertainty quantification with physics,” in Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp. 237–247, 2021.DOI
  48. [48]B. Bullwinkel, D. Randle, P. Protopapas, and D. Sondak, “DEQGAN: Learning the loss function for PINNs with generative adversarial networks,” 2022.arxiv.org/abs/2209.07081
  49. [49]K. Hornik, M. Stinchcombe, and H. White, “Multilayer feedforward networks are universal approximators,” Neural networks, vol. 2, no. 5, pp. 359–366, 1989.
  50. [50]J. Berg and K. Nyström, “A unified deep artificial neural network approach to partial differential equations in complex geometries,” Neurocomputing, vol. 317, pp. 28–41, 2018.DOI
  51. [51]S. Cai, Z. Mao, Z. Wang, M. Yin, and G. E. Karniadakis, “Physics-informed neural networks (pinns) for fluid mechanics: A review,” Acta Mechanica Sinica, pp. 1–12, 2022.arxiv.org/abs/2105.09506
  52. [52]M. Arjovsky, S. Chintala, and L. Bottou, “Wasserstein generative adversarial networks,” in International conference on machine learning, pp. 214–223, PMLR, 2017.
  53. [53]L. V. Kantorovich, “Mathematical methods of organizing and planning production,” Management science, vol. 6, no. 4, pp. 366–422, 1960.DOI
  54. [54]L. Mescheder, A. Geiger, and S. Nowozin, “Which training methods for gans do actually converge?,” in International conference on machine learning, pp. 3481–3490, PMLR, 2018.arxiv.org/abs/1801.04406
  55. [55]I. Gulrajani, F. Ahmed, M. Arjovsky, V. Dumoulin, and A. C. Courville, “Improved training of wasserstein gans,” Advances in neural information processing systems, vol. 30, 2017.arxiv.org/abs/1704.00028
  56. [56]C. Kim, H. Son, and C. Kim, “Fully automated registration of 3d data to a 3d cad model for project progress monitoring,” Automation in Construction, vol. 35, pp. 587–594, 2013.DOI
  57. [57]R. E. Kalman, “A new approach to linear filtering and prediction problems,” Journal of Basic Engineering, vol. 82, no. 1, pp. 35–45, 1960.
  58. [58]I. M. Sobol’, “On the distribution of points in a cube and the approximate evaluation of integrals,” Zhurnal Vychislitel’noi Matematiki i Matematicheskoi Fiziki, vol. 7, no. 4, pp. 784–802, 1967.DOI
  59. [59]T. Salimans, I. Goodfellow, W. Zaremba, V. Cheung, A. Radford, and X. Chen, “Improved techniques for training gans,” Advances in neural information processing systems, vol. 29, 2016.arxiv.org/abs/1606.03498
  60. [60]M. Heusel, H. Ramsauer, T. Unterthiner, B. Nessler, and S. Hochreiter, “Gans trained by a two time-scale update rule converge to a local nash equilibrium,” Advances in neural information processing systems, vol. 30, 2017.arxiv.org/abs/1706.08500
  61. [61]R. Zhang, P. Isola, A. A. Efros, E. Schechtman, and O. Wang, “The unreasonable effectiveness of deep features as a perceptual metric,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 586–595, 2018.arxiv.org/abs/1801.03924

Paper details

vlnc:W_p9u5tzf2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 7

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents