Physics-Informed Neural Networks for Solving Contact Problems in Three Dimensions

Tarik Sahin \orcidID 0000-0002-4134-3726, Daniel Wolff \orcidID 0000-0001-5767-7803 and Alexander Popp \orcidID 0000-0002-8820-466X · 2024

Read Physics-Informed Neural Networks for Solving Contact Problems in Three Dimensions by Tarik Sahin \orcidID 0000-0002-4134-3726, Daniel Wolff \orcidID 0000-0001-5767-7803 and Alexander Popp \orcidID 0000-0002-8820-466X (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2412-09022.txt.

Introduction

Tarik Sahin and co-authors ask if neural nets trained with physics rules can solve touch problems in three dimensions. Touch, or contact, tasks ask how bodies press, stick, or slide where they meet. The group works in 2024 with small shape changes in springy solids. Edge rules are built into the net as hard limits, so the net cannot break them.

Soft rules in the loss handle the one-way nature of contact, where push is fine but pull across a gap is not. A smooth math trick helps state those rules in a form nets can learn. Two test cases check the method: a flat patch test and the classic ball-press case. The study maps how well the nets cope with true 3D touch.

Catalog record

CREATORTarik Sahin \orcidID 0000-0002-4134-3726, Daniel Wolff \orcidID 0000-0001-5767-7803 and Alexander Popp \orcidID 0000-0002-8820-466X
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE25 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2412-09022

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Physics-Informed Neural Networks for Solving Contact Problems in Three Dimensions Tarik Sahin \orcidID 0000-0002-4134-3726, Daniel Wolff \orcidID 0000-0001-5767-7803 and Alexander Popp \orcidID 0000-0002-8820-466X Abstract This paper explores the application of physics-informed neural networks (PINNs) to tackle forward problems in 3D contact mechanics, focusing on small deformation elasticity. We utilize a mixed-variable formulation, enhanced with output transformations, to enforce Dirichlet and Neumann boundary conditions as hard constraints. The inherent inequality constraints in contact mechanics, particularly the Karush-Kuhn-Tucker (KKT) conditions, are addressed as soft constraints by integrating them into the network’s loss function. To enforce the KKT conditions, we leverage the nonlinear complementarity problem (NCP) approach, specifically using the Fischer-Burmeister function, which is known for its advantageous properties in optimization. We investigate two benchmark examples of PINNs in 3D contact mechanics: a single contact patch test and the Hertzian contact problem. 1 Introduction Machine learning methods often demand large datasets, which are challenging to obtain in engineering due to costly experiments, the unavailability of certain measurements (e.g., inaccessible components or the absence of suitable sensors), or long simulation runtimes. When real-world data is available, it is frequently noisy or mislabeled. These limitations can undermine the performance of classical data-driven models, as they lack a physics-based mechanism to validate predictions. Physics-informed neural networks (PINNs) have been introduced to address these challenges [ 1 ] . Using automatic differentiation, PINNs integrate physical knowledge (e.g., partial differential equations (PDEs)) into the neural network’s loss function, introducing physics-based regularization that mitigates the black-box nature of purely data-driven approaches. For forward problems, PINNs serve as mesh-free PDE solvers, requiring only sampling points from the (arbitrary, potentially geometrically intricate) domain [ 2 ] . They also overcome the curse of dimensionality when approximating high-dimensional functions [ 3 ] and excel in solving inverse problems by seamlessly incorporating measurement data [ 4 ] . PINNs can be also used for thermo-mechanical coupling problems using a mixed formulation [ 5 ] . In this work, we extend PINNs to 3D contact mechanics, building on previous 2D examples [ 6 ] . The employed PINNs are developed based on a mixed-variable formulation inspired by the Hellinger-Reissner principle, where both displacement and stress fields are network outputs. Contact mechanics involves Karush-Kuhn-Tucker (KKT) type constraints, commonly referred to as Hertz-Signorini-Moreau conditions, which we incorporate as soft constraints using the Fischer-Burmeister nonlinear complementarity problem (NCP) function. This approach efficiently handles inequality constraints in contact mechanics. We validate the method with two examples assuming small deformation elasticity for simplicity:: (1) a single contact patch test where all points are in contact and (2) the Hertzian contact problem, where the contact area is detected as part of the solution. 2 Problem Formulation: Contact Mechanics We consider a 3D contact problem between an elastic body and a fixed rigid obstacle, as illustrated in Fig. 1 . In the reference configuration, the elastic body is denoted by \Omega_{0} , while in the current configuration, it is represented by \Omega_{t} . The rigid obstacle remains fixed in its configuration, \Omega_{r} . Fig. 1 c illustrates a scenario where the two bodies are in contact. The surface of the elastic body can be divided into three regions: the Dirichlet boundary \partial\Omega_{u} , where displacements are specified; the Neumann boundary \partial\Omega_{\sigma} , where tractions are applied; and the potential contact boundary \partial\Omega_{c} , where contact constraints may apply. The true contact surface is a subset of \partial\Omega_{c} and is determined as part of the solution procedure. Figure 1: Contact problem between an elastic body and a rigid obstacle. (a) Reference configuration, (b) current configuration, (c) boundary conditions illustration, including the gap g_{n} , tangential tractions t_{\tau}^{\xi} , t_{\tau}^{\eta} and contact pressure p_{n} . We define the target boundary value problem (BVP) as follows: \displaystyle\bm{\nabla\cdot}\bm{\sigma}+\mathbf{\hat{b}} \displaystyle=\mathbf{0}\hskip 8.53581pt\text{in }\Omega_{t}, (1) \displaystyle\bm{u} \displaystyle=\bm{\hat{u}}\hskip 7.39772pt\text{on }\partial\Omega_{u}, (2) \displaystyle\bm{\sigma}\cdot\bm{n} \displaystyle=\bm{\hat{t}}\quad\text{on }\partial\Omega_{\sigma}. (3) In this formulation, \bm{\sigma} is the Cauchy stress tensor, and \bm{u} is the displacement vector, which is the primary unknown. The vector \mathbf{\hat{b}} represents body forces acting on the domain, and \bm{n} is the outward unit normal vector on the boundary. Displacements prescribed on the Dirichlet boundary \partial\Omega_{u} are denoted by \bm{\hat{u}} , while tractions prescribed on the Neumann boundary \partial\Omega_{\sigma} are represented by \bm{\hat{t}} . In addition to these equilibrium equations, we also account for the kinematic and constitutive equations that describe the behavior of the elastic material. For simplicity, linear kinematics and linear-elastic material behavior will be assumed in the numerical examples section, given by: \displaystyle\bm{\varepsilon} \displaystyle=\frac{1}{2}(\bm{\nabla}\bm{u}+\bm{\nabla}\bm{u}^{T}), (4) \displaystyle\bm{\sigma} \displaystyle=\mathbb{C}:\bm{\varepsilon}. (5) Here, \bm{\varepsilon} is the infinitesimal strain tensor, and \mathbb{C} is the fourth-order elasticity tensor. For the case of linear isotropic elasticity, the constitutive equation can be simplified using Hooke’s law: \bm{\sigma}=\lambda\tr(\bm{\varepsilon})\mathbf{I}+2\mu\bm{\varepsilon}, (6) where \lambda and \mu are the Lamé parameters, \tr(\cdot) denotes the trace operator (sum of diagonal components), and \mathbf{I} is the identity tensor. The displacement vector \bm{u} of the elastic body is determined by describing the motion of a point in the reference configuration \bm{X}\in\Omega_{0} to a point in the current configuration \bm{x}\in\Omega_{t} , as shown in Figs. 1 a and 1 b. This relation is expressed as follows: \bm{u}=\bm{x}-\bm{X}. (7) The gap function g_{n} , which measures the separation between the elastic and rigid bodies in the current configuration, is defined as: g_{n}=-\bm{n}\cdot(\bm{x}-\bm{\hat{x}}). (8) Here, \bm{\hat{x}} denotes the closest point projection of \bm{x} onto the surface of \Omega_{r} (see Fig. 1 b). Since all contact constraints are enforced in the current configuration, the contact pressure p_{n} , and tangential traction components t_{\tau}^{\xi} and t_{\tau}^{\eta} and are obtained by decomposing the contact traction vector \bm{t}_{c} via traction vector decomposition (TVD) as follows: \bm{t}_{c}=p_{n}\bm{n}+t_{\tau}^{\xi}\bm{\tau}^{\xi}+t_{\tau}^{\eta}\bm{\tau}^{\eta},\quad p_{n}=\bm{t}_{c}\cdot\bm{n},\quad t_{\tau}^{\xi}=\bm{t}_{c}\cdot\bm{\tau}^{\xi},\quad t_{\tau}^{\eta}=\bm{t}_{c}\cdot\bm{\tau}^{\eta}, (9) where \bm{t}_{c}=\bm{\sigma}\cdot\bm{n}\quad\quad\text{on }\partial\Omega_{c}. (10) For a frictionless contact problem, the classical Karush-Kuhn-Tucker (KKT) type constraints (Eq. 11 ), also known as the Hertz-Signorini-Moreau (HSM) conditions, are employed along with the frictionless sliding conditions (Eq. 12 ): \displaystyle g_{n}\geqslant 0,\quad p_{n}\leqslant 0 \displaystyle,\quad p_{n}g_{n}=0, (11) \displaystyle t_{\tau}^{\xi}=0,\quad \displaystyle t_{\tau}^{\eta}=0. (12) For further details on more advanced contact constitutive laws, including friction, the reader is referred to [ 7 , 8 ] . In this study, we adopt a mixed-variable approach, where both the displacement field \bm{u} and the stress tensor \bm{\sigma} are treated as primary variables for solving contact problems. An additional key requirement is the stress-to-stress coupling between the primary variable \bm{\sigma} and the secondary variable \bm{\sigma}^{u} , defined by the relation \bm{\sigma}=\bm{\sigma}^{u}. (13) The term \bm{\sigma}^{u} is computed by inserting the kinematic equation (Eq. 4 ) into the constitutive law (Eq. 5 ). This condition ensures that the two master fields remain consistent and compatible throughout the solution process. This method has been comprehensively discussed in the work by Sahin et al. [ 6 ] . 3 Extending Physics-Informed Neural Networks for 3D Contact Mechanics 3.1 Loss Formulation in PINNs In quasi-static solid mechanics using PINNs, a fully-connected neural network (FNN) \mathcal{N} with parameters \bm{\theta} and spatial coordinates \bm{x} to the displacement \bm{u} and stress tensor \bm{\sigma} : \bm{\tilde{u}}=\mathcal{N}_{\bm{u}}(\bm{x};\bm{\theta})^{\prime},\quad\bm{\tilde{\sigma}}=\mathcal{N}_{\bm{\sigma}}(\bm{x};\bm{\theta})^{\prime}. (14) Here, (.)^{\prime} denotes a user-defined output transformation and (\tilde{.}) indicates a predicted quantity. It is important to note that, due to the conservation of angular momentum, \bm{\sigma}=\bm{\sigma}^{T} , only 6 stress components are predicted instead of 9. The total loss \pazocal{L}_{E}

Read the complete text