Rapid Kirigami Simulation using the Bar & Hinge Approach
Raj Pradip Khawale, Elaheh Mehdizadeh, John Brigham, Evgueni T. Filipov
cond-mat.other, cs.CE
2026-08-14
A bar-hinge-and-3NTS reduced model of kirigami matches Abaqus open area within 3.8–6.8% and Mylar tests within 4% on Poisson’s ratio; a 10×10 sheet drops from 81,645 s to 166 s.
Kirigami cuts a thin sheet so it can stretch, buckle, and morph. Anyone sweeping cut patterns usually lands on shell finite elements. Shells resolve stress, but thin-sheet large deformation, buckling, and locking make them slow and brittle to converge. Closed-form models are fast and stuck to one cell family. GAN and deep-learning surrogates need a training set and tend to stay inside that set.
The gap is a reduced-order model that accepts an arbitrary cut layout, tracks global shape and reaction force, and is cheap enough for parametric sweeps.
University of Michigan and University of Pittsburgh take the origami bar-and-hinge idea and put it on a kirigami mesh. The sheet is a network of three spring types.
Nodes translate only: \(3n \times 3n\) in 3D, \(2n \times 2n\) in 2D. Constitutive response is linear elastic; geometry is nonlinear, solved with displacement-controlled Newton–Raphson and adaptive steps. Thin sheets buckle from small defects, so a modal z-perturbation of about \(10^{-2}\) times a characteristic length is added up front. Meshes come from MATLAB PDE Toolbox, sized by a max edge length \(L{\max}\), with automatic refinement near cuts.
Benchmarks ran on an Apple M2 with 64 GB of RAM. Four periodic cellular sheets were stretched 20% across the cuts (the densest 10×10 case only to strain 0.17):
| Cells | Bar-hinge nodes / s | Abaqus nodes / s | Open-area gap |
| 2×5 | 7542 / 89.4 | 67325 / 1144 | 4% |
| 5×10 | 11581 / 102.1 | 69453 / 1835 | 4.5% |
| 5×15 | 14068 / 185.9 | 71929 / 4352 | 3.8% |
| 10×10 | 13704 / 165.9 | 75899 / 81645 | 6.8% |
Simple geometries are about 13–23× faster. The 10×10 sheet took Abaqus 22.7 hours and the reduced model 166 s, roughly 492×. The paper’s headline is “at least tenfold,” which folds in the easy cases. The reduced model runs a bit stiff, more so on the harder sheets.
On the lab side, a single-cut disk is checked against the stretching-buckling experiment of Chaudhary et al. Rainbow and honeycomb Mylar sheets were laser-cut and pulled on a Mark-10 to strain 0.5. Poisson’s ratio, peak out-of-plane displacement, and open area differ by under 4%, 4%, and 5%. Force–strain tracks well at small strain and again runs stiff later. Non-periodic and intersecting cuts at 12% stretch produce plausible global shapes; no extra error table is given.
Four workloads show where the speed goes. Seventy-five Latin-hypercube cut designs for permeability took about five hours on an M2 Max; permeability scales roughly with open area cubed. A three-sided crawler with triangular, circular, and trapezoidal cuts shows anisotropic friction, strongest for triangles. A square–rhombus nonreciprocal lattice (\(\theta=\pi/16\), Mylar \(E=3.5\,\mathrm{GPa}\), 6.75 mm sides) already splits \(u{21}-u{12}\) for \(F<10\,\mathrm{N}\). A polarizable metamaterial sweep costs about 20 s per twist angle.
If the question is global shape and reaction force, not crack-tip stress, this spring network turns “sweep a few hundred cut patterns” from an FEA farm into a laptop job. The MATLAB code is public.
It is not a neural surrogate, so a new cut does not need a new training set. It is also not a new continuum theory: origami bars and hinges, wired onto a kirigami triangulation. Against a GAN full-field predictor it is slower, but it emits stiffness matrices and internal forces and can do eigenmodes and buckling. \(Fs\) and \(FB\) are global constants. Change the material, thickness, or mesh density and they need a re-check.
The authors are clear: local stress concentrations and edge effects are out of reach, so fracture and yield are not. Shear stiffness is systematically high because a bar network approximates shear through axial stretch, and one global \(Fs\) cannot cancel that. Mesh convergence still exists. The code is MATLAB scripts, not a packaged tool. Material nonlinearity, plasticity, and rate effects are absent.
The FEA comparison is not matched on degrees of freedom. Abaqus uses 5–9× more nodes, so part of the speedup is a coarser mesh. The 10×10 jump toward 500× is largely shell-element convergence pain, not asymptotic complexity. The model is slightly stiff almost everywhere, worse at large strain, which is the unsafe side for peak-load estimates. Crawler friction uses the fit \(\muf=\mu0+Kf\cdot h\cdot\rho\cdot Sf\), not contact. \(Fs\) and \(FB\) were tuned on the calibration set; hold-out range is not reported.