Learning Parameterized Nonlinear Elasticity on Curved Surfaces

Yankang Liu, Ke Zhang, Maziar Raissi, Roya Zandi · 2026

Read Learning Parameterized Nonlinear Elasticity on Curved Surfaces by Yankang Liu, Ke Zhang, Maziar Raissi, Roya Zandi (2026) 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-2604-12170.txt.

Introduction

Liu and three colleagues learn nonlinear elasticity on curved surfaces with a physics-informed neural network that builds the physics into its loss. One trained model then spans a whole family of equilibria across shapes and materials, unlike solvers restarted per parameter choice. On a spheroidal single-disclination test it matches known solutions, even for choices left out of training.

Catalog record

CREATORYankang Liu, Ke Zhang, Maziar Raissi, Roya Zandi
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE24014
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.12170

Browse the full library · About the Tower of Babel

Full text

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

Learning Parameterized Nonlinear Elasticity on Curved Surfaces Yankang Liu, Ke Zhang, Maziar Raissi, Roya Zandi Abstract We learn parameterized nonlinear elasticity on curved surfaces using a physics-informed neural network that enforces governing equations and boundary conditions directly through the loss function, enabling a single trained model to represent a continuous family of elastic equilibria across geometric and material parameters. Nonlinear elasticity on curved manifolds underlies the mechanics of crystalline shells, elastic membranes, and viral capsids, where curvature and topological defects determine equilibrium structure and stability. Traditional exact and finite element solvers rely on symmetry reduction and must be reinitialized for each parameter choice, limiting scalability when symmetry is broken or parameters vary. We validate the proposed learning-based solver on a benchmark problem from curved elasticity, namely the one-dimensional single disclination on a spheroidal surface with known exact and numerical solutions. The network accurately reproduces these solutions, including parameter combinations excluded from training, demonstrating generalization across geometry and material regimes. This study establishes a scalable framework for learning nonlinear elastic systems on curved manifolds and lays the groundwork for extensions to fully two-dimensional and multi-defect configurations relevant to protein shells and other curved elastic networks. 1 Introduction Figure 1: Protein lattices on curved viral capsids. Icosahedral (HBV) and conical (HIV) capsids exhibit curvature-induced lattice defects that govern global shape and mechanical equilibrium. Red circles indicate pentameric defects required by curvature. Nonlinear elasticity on curved surfaces governs the mechanics of crystalline protein shells, elastic membranes, and viral capsids 3 ; 4 . During viral assembly, proteins form curved lattices whose equilibrium structures can be understood as free energy minimizing elastic networks constrained by geometry. Curvature induces geometric frustration that requires topological defects, known as disclinations, in the lattice structure. The number and arrangement of these defects determine global shape and mechanical stability. Representative examples include icosahedral hepatitis B virus (HBV) capsids and the conical capsid of HIV, both of which exhibit ordered protein lattices on curved manifolds with defect-mediated geometry, see Fig. 1 . Computing these equilibria requires solving nonlinear elasticity equations on curved manifolds. Exact and finite element methods are effective in highly symmetric configurations, such as a centered disclination on a spheroidal surface 1 , where symmetry reduction yields a one-dimensional boundary value problem. However, when symmetry is broken, parameters vary, or multiple interacting defects are present, the system becomes fully two dimensional and classical solvers face scaling and stability limitations 5 . Moreover, traditional approaches must typically be reinitialized for each new parameter choice, making exploration of continuous geometric and material regimes computationally expensive. From a machine learning perspective, this setting corresponds to learning solutions of a geometry-dependent nonlinear partial differential equation with parameter-dependent metrics. While physics-informed neural networks (PINNs) have demonstrated promise across a range of differential equations 8 , most benchmarks focus on flat domains or relatively simple operators. It remains unclear whether physics-constrained learning can reliably handle nonlinear elasticity on curved manifolds and generalize across continuous parameter spaces. In this work, we learn parameterized nonlinear elasticity on curved surfaces using a physics-informed neural network that enforces governing equations and boundary conditions directly through the loss function. The network is conditioned on geometric and material parameters, enabling a single trained model to represent a continuous family of elastic equilibria rather than solving isolated instances. We validate the approach on a benchmark problem consisting of a single centered disclination on a spheroidal surface with known exact and numerical solutions. Our goal is not to revisit the benchmark physics, but to establish a scalable learning-based solver that generalizes across parameter regimes and provides a foundation for extending physics-informed learning to fully two-dimensional and multi-defect configurations relevant to protein shells and other curved elastic networks. 2 Problem Setup 2.1 Nonlinear Elasticity on a Curved Surface We consider nonlinear elasticity of a crystalline shell constrained to a curved surface. The deformation is described by a mapping between a reference configuration \bar{x} and the actual surface x , inducing distinct metrics on the two manifolds. Mechanical equilibrium follows from minimizing the elastic energy under geometric compatibility, leading to the covariant force balance equation 2 ; 7 ; 4 \nabla_{\alpha}\sigma^{\alpha\beta}+(\bar{\Gamma}^{\beta}_{\gamma\nu}-\Gamma^{\beta}_{\gamma\nu})\sigma^{\gamma\nu}=0, (1) where \sigma^{\alpha\beta} is the in-plane stress tensor and \Gamma , \bar{\Gamma} denote the Christoffel symbols associated with the actual and reference geometries. This equation defines a nonlinear partial differential system whose coefficients depend explicitly on surface curvature and material parameters, resulting in a geometry-dependent nonlinear operator. We solve these equations subject to the boundary conditions n_{\rho}\sigma^{\rho\lambda}\bar{g}_{\lambda\nu}+\tau n_{\nu}/r_{A}=0 and \bar{\mathbf{r}}(\mathbf{r}=(0,0))=(0,0) . See Appendix A for details. In general the system is two dimensional. In this work, we focus on a symmetry-reduced benchmark consisting of a centered disclination on a spheroidal surface 1 . Under rotational symmetry, the problem reduces to a one-dimensional boundary value equation for the radial deformation \bar{r}(r) , while retaining curvature-dependent nonlinear structure. Thus, although reduced in dimension, the governing equation remains geometrically nontrivial. The equilibria depend on geometric and material parameters (\beta,q,\nu_{p}) , where \beta controls spheroidal curvature (see Appendix B ), q denotes defect charge, and \nu_{p} is the Poisson ratio. Varying these parameters generates a continuous family of elastic equilibria forming a parameterized solution manifold. Explicit forms of the reduced equations and boundary conditions are provided in Appendix A . 2.2 Physics-Informed Learning of Parameterized Equilibria We approximate the radial deformation \bar{r}(r)\big|_{(\beta,q,\nu_{p})} using a deep fully connected neural network u(r;\beta,q,\nu_{p}) with hyperbolic tangent activations. The spatial coordinate is r\in[0,r_{b}] , and (\beta,q,\nu_{p}) denote geometric and material parameters. Automatic differentiation is used to compute derivatives with respect to the spatial coordinate and network parameters, while the parameters (\beta,q,\nu_{p}) will not be differentiated. By explicitly conditioning on these parameters, a single network represents a continuous family of elastic equilibria rather than isolated parameter instances. Let F[\cdot] denote the nonlinear differential operator defined by the reduced equilibrium equation (equation 17 ). The physical residual at a point \tilde{r}\equiv(r,\beta,q,\nu_{p}) is denoted as R(\tilde{r})\equiv F[u(\tilde{r})] with the analogous definition at the boundary. Given N sampled points, the loss functions are \mathcal{L}_{\text{phys}}\equiv\sum_{i=1}^{N}|R(\tilde{r}_{i})|^{2}/N and \mathcal{L}_{\text{bc}}\equiv\sum_{i=1}^{N}|R(\tilde{r}_{bc,i})|^{2}/N . The boundary condition at the origin is nothing but a regularization, which is enforced via \mathcal{L}_{\text{0}}\equiv\sum_{i=1}^{N}|u(\tilde{r}_{0,i})|^{2}/N . A small subset of ground truth data is incorporated through \mathcal{L}_{\text{data}}\equiv\sum_{i=1}^{N}|u(\tilde{r}_{i})-\bar{r}(r_{i})|^{2}/N . The total loss is the superposition of these contributions. 3 Experiments We evaluate whether a single parameter-conditioned PINN can learn solutions across a continuous parameter space and remain stable under variations in collocation densities and network architectures. Ground-truth solutions are generated using finite element method for \beta\in[1,3] , q\in[0,1] , \nu_{p}\in[0.1,0.9] (step sizes are all 0.1), and 101 spatial points per configuration, yielding approximately 2\times 10^{5} solution samples. Configurations with \beta\in\{1,2,3\} , q\in\{0,1\} , or \nu_{p}\in\{0.3,0.8\} are excluded from training to evaluate interpolation across parameter space. Samples with r<0.1 are removed to avoid numerical instability near the origin. The size of the training set after the exclusion is around 0.57N where N is the number of the sampled points from the whole data set. Figure 2 compares neural network predictions with finite element solutions for parameter combinations excluded from training with N=4000 . The model employs four hidden layers with 200 neurons per layers. Despite not being observed during optimization, the predicted deformation profiles exactly match the true solutions,

Read the complete text