Two branches of mathematics turn out to share the same equations. The equations that say whether a curved surface with given internal distances can sit in ordinary three-dimensional space, called the Gauss–Codazzi equations, have the same form as the equations of steady gas flow. Where the surface curves like a sphere (positive Gaussian curvature), the problem behaves like gentle subsonic flow. Where it curves like a saddle (negative curvature), it behaves like supersonic flow, and supersonic flow forms shocks. This correspondence has been worked out carefully on paper. Nobody has asked the physical question it raises. A thin elastic sheet whose internal distances are programmed to be saddle-like cannot lie flat, so it buckles and often forms sharp ridges. Are those ridges the geometric twins of gas shocks? Are they placed where a shock-capturing gas solver would put them? And does the sheet's thickness act like viscosity, choosing one physical shape out of the infinitely many shapes the mathematics allows?
We propose to answer this by experiment. We will fabricate sheets with programmed internal geometry at several thicknesses, 3D-scan the shapes they settle into, and compare the scans against two competing predictions. The first comes from solving the Gauss–Codazzi equations as a compressible-flow problem with vanishing artificial viscosity. The second comes from minimizing the sheet's elastic energy. Both selection rules appear in the theory literature, and nobody has said which one physical sheets obey. Settling that question is the contribution. A cheap gas-dynamics solver for crease prediction would be a useful by-product if the first hypothesis wins.
Core Claim
The surface–fluid correspondence is surveyed by Li. Its explicit dictionary between curvature and density, velocity and pressure comes from Slemrod. Together they set up the problem as follows. Write the surface's second fundamental form, the quantities describing how it bends in space, as $(L, M, N)$, and its metric, the internal distances, as $g$. The Gauss equation then reads
$$ LN - M^2 = K \det g, $$
where $K$ is the Gaussian curvature fixed by $g$ alone. The two Codazzi equations play the role of momentum balance, and this Gauss equation plays the role of the equation of state. When $K<0$ the system is hyperbolic, which means it can carry jumps in $(L, M, N)$ across curves; these are creases or ridges.
Acharya, Chen, Li, Slemrod and Wang motivate the selection question. They prove that this kind of system has infinitely many rough "wrinkled" solutions, the twins of Nash–Kuiper isometric embeddings. They also show that these solutions are time-reversible, so entropy-rate criteria cannot pick among them. They conclude that a physical choice must come from either a viscosity limit or energy minimization.
Our claim is that, for real sheets, bending stiffness supplies the viscosity. A sheet of thickness $t$ resists bending with stiffness $B \propto t^3$ and resists stretching with stiffness $Y \propto t$. The length scale over which bending smooths a crease is therefore $\sqrt{B/Y} \propto t$. This length plays the role of the artificial viscosity $\varepsilon$ in a shock-capturing scheme. Viscous shock width scales linearly with $\varepsilon$, so we predict a ridge width $w \propto t$. We also predict that, as $t \to 0$, ridge positions converge to the shock locations of the vanishing-viscosity gas solution, not to the ridges of an energy-minimizing isometric embedding where the two differ. Both statements can be falsified with a ruler, a scanner and sheets of three or four thicknesses.
Research Program
The computational side builds a solver for the Gauss–Codazzi system written in gas-dynamic variables. It uses a finite-volume, shock-capturing discretization with an explicit artificial-viscosity parameter $\varepsilon$, and it handles domains where $K$ changes sign, which correspond to mixed subsonic–supersonic flow. The open problem flagged in Li's survey is that analysis breaks down on such domains. Numerically, however, transonic flow solvers routinely handle mixed type, which is precisely the toolkit the correspondence lets us import. We will make the solver differentiable, in the style of Aquarium and Lettuce. Both are flow solvers that expose gradients, and Aquarium has been checked against soft-robot hardware. Differentiability lets us fit the as-fabricated metric from measured dot spacings instead of assuming the design metric. That matters because manufacturing deviations are the usual source of mismatch between simulation and experiment, as Yan et al. found for wall-thickness variation in bellows buckling.
The closest computational work is Flouris et al.'s curved-space lattice Boltzmann method. It also handles moving walls through a changing metric, encoding wall motion as Christoffel-symbol forces, but it uses the metric only as a coordinate trick for fluid next to a deforming wall. It never invokes the Gauss–Codazzi duality, is limited to deformations of about 10% of domain size, and was validated only against synthetic numerical benchmarks. We do not simulate a fluid at all. We use fluid numerics to solve a geometry problem and test the result against physical objects.
The experimental side programs the metric of thin sheets in two independent ways, so that conclusions do not rest on one fabrication route. The first route uses heat-shrinkable polymer film with printed shrink patterns. The second uses 3D-printed elastomer bilayers whose layers swell by different amounts. We prescribe three families of target metric: purely saddle-like discs ($K<0$ everywhere), sphere-like caps ($K>0$, the control case where no ridges are predicted), and annuli whose curvature changes sign along a circle (mixed type). Each family is made at three to four thicknesses spanning roughly a factor of ten. A grid of fiducial dots photographed before and after activation gives the realized metric. Structured-light scanning gives the 3D shape, and from the scan we extract ridges as curves of maximal principal curvature.
Baselines
The naive reference point is the smooth embedding obtained by solving the same Gauss–Codazzi system with large, fixed artificial viscosity. This predicts gentle rounded shapes with no localized ridges. If the scanned sheets match it as well as any sharper model does, the shock picture adds nothing.
The strongest existing method is standard nonlinear thin-shell finite-element simulation of the sheet with the measured metric imposed as a growth or shrink strain. We will run it in Abaqus, the structural solver Prévereaud et al. coupled to CFD and validated against wind-tunnel data. That study is the clearest precedent in the file for this kind of engineering pipeline matching experiments; its axial-force coefficients fell within 0.6% to 9.5% of supersonic tunnel measurements. That figure is for aerodynamic loads, not sheet shape. For ridge location on metric-programmed sheets we found no reported result on Lacuna, so the baseline numbers will come from our own runs. For fairness, the shell model receives the same measured metric, the same boundary conditions and the actual thickness. Its mesh is refined until ridge positions stop changing, and we report wall-clock time beside accuracy. Matching it would mean ridge-location error within its experimental scatter. Beating it would mean either lower error at the thinnest sheets, where the shell mesh must resolve features of width proportional to $t$, or equal error at a cost one to two orders of magnitude lower.
The scientific competitor is energy minimization. We compute an isometric or near-isometric embedding by directly minimizing a discrete elastic energy of stretching plus bending as $t \to 0$. This is the other selection principle named by Acharya et al.. The test is decisive only on metrics where the two principles place ridges differently. We will identify those metrics computationally before fabrication and over-sample them.
The key ablation isolates the entropy-type selection. We run the gas solver with a non-dissipative central scheme and with several forms of artificial viscosity: isotropic diffusion, and a bending-like fourth-order term that mimics a plate. If ridge predictions depend strongly on the form of the dissipation, then "viscosity selects the sheet" is not one prediction but several, and the experiment tells us which form physics uses.
Why not just…?
Why not just run nonlinear shell finite elements and skip the analogy? This is the real threat. If shell simulation matches the scans at every thickness at acceptable cost, then the gas-dynamics view is an interpretation, not a predictive tool. The proposal survives that outcome only on its scientific half. The thickness series still tells us whether ridge width scales as $t$, and whether thin-sheet ridges converge to shock locations or to energy-minimizing ones. Shell simulation cannot reach that limit cheaply, because its mesh must shrink with $t$. We will report the cost-versus-accuracy comparison plainly, including the case where finite elements win.
Why not just minimize elastic energy directly? Energy minimization is a legitimate selection rule and may be the correct one. That is why it is a co-hypothesis and not a strawman. Its weaknesses are practical: nonconvex minimization lands in metastable states that depend on initialization. The experiment settles the question on the metrics where the two rules disagree. If energy minimization wins there, we will report that the viscosity reading of Acharya et al. does not describe elastic sheets.
Why not just train a neural surrogate on simulated sheets? On CFDBench, standard neural operators reach errors as high as 300% on unseen geometries and parameters. Sharp, thickness-dependent features in new curvature layouts are exactly that out-of-distribution setting. A surrogate also cannot answer the selection question, because it learns whichever selection its training simulator used.
Proposed Experiments and Decisive Checks
The first check is the positive-curvature control. Caps must show no localized ridges, and the gas solver, in its elliptic regime, must predict none. A failure here means the fabrication or the mapping is broken. The second check is scaling. Across thicknesses, measured ridge width plotted against $t$ should have a log–log slope near one. A slope near one-half or one-third would point to a different regularization mechanism and falsify the bending-as-viscosity claim. The third check is selection. On the disagreement metrics, we measure the mean distance between scanned ridges and predicted ridges, normalized by sheet diameter, for the vanishing-viscosity solver, energy minimization and shell finite elements, and test whether the gap between hypotheses exceeds the scatter across repeat samples. The fourth check targets the mixed-type annuli, the regime where analysis is open. Here we test whether the boundary between ridged and smooth regions in the scans coincides with the measured $K=0$ curve, and whether ridges terminate there as shocks end at a sonic line. Throughout, we report RMS surface distance between scan and prediction alongside the ridge metrics, so that a model cannot win on ridges while getting the overall shape wrong.
References
- Some Recent Developments on Isometric Immersions via Compensated Compactness and Gauge Transforms
- From the Nash--Kuiper Theorem of Isometric Embeddings to the Euler Equations for Steady Fluid Motions: Analogues, Examples, and Extensions
- Fluids, Elasticity, Geometry, and the Existence of Wrinkled Solutions
- Aquarium: A Fully Differentiable Fluid-Structure Interaction Solver for Robotics Applications
- Lettuce: PyTorch-Based Lattice Boltzmann Framework
- Mechanism of Mechanical Analysis on Torsional Buckling of U-Shaped Bellows in FLNG Cryogenic Hoses
- Fluid structure interaction with curved space lattice Boltzmann
- Numerical Aero-shape Characterization of an Inflatable Heat Shield Re-entry Vehicle in the EFESTO-2 Project
- CFDBench: A Comprehensive Benchmark for Machine Learning Methods in Fluid Dynamics