1. Continuum Mechanics Primer

Kinematics of deformation, deformation gradient, small vs finite strain, stress measures, linear elasticity, Voigt notation, plane stress/strain, compatibility, hyperelasticity

Contents

1. Why Continuum Mechanics Before iFEM 2. Configurations, Motion, and Displacement 3. The Deformation Gradient F 4. Strain Measures: Small Strain vs Green–Lagrange 5. Compatibility: When Is a Strain Field Integrable? 6. Stress Measures: Cauchy, First and Second Piola–Kirchhoff 7. Linear Elasticity and the Hooke Tensor 8. Voigt Notation and the Factor-of-2 Shear Trap 9. Plane Stress and Plane Strain 10. Hyperelasticity: Stress From a Strain-Energy Density 11. Neo-Hookean, Mooney–Rivlin, and Incompressibility 12. Choosing the Right Theory: A Decision Guide Interactive: Strain-Convention Trap Lab Interactive: Uniaxial Stress Explorer Interactive: From Motion to Stress and Back Flashcards

1. Why Continuum Mechanics Before iFEM

Continuum mechanics starts with a deliberate act of forgetting: replace the molecular detail of a body with smooth fields — a displacement $\mathbf{u}(\mathbf{X})$, a strain field, a stress field. The continuum hypothesis is justified whenever the smallest feature you care about is far larger than the microstructure. For a silicone tactile skin the polymer network mesh size is around 10 nm while the skin geometry lives at the millimeter scale — five orders of magnitude of separation. Field theory applies without apology.

This entire study guide is about inverting parts of one map. The forward elasticity map runs: geometry + boundary conditions + constitutive law $\rightarrow$ displacement field everywhere $\rightarrow$ strains $\rightarrow$ stresses $\rightarrow$ reaction and contact forces. iFEM inverts a piece of it: sparse measured strains (strain gauges, fibers) or vision-derived surface displacements $\rightarrow$ best-fit full displacement field $\rightarrow$ interior strain and stress $\rightarrow$ contact load reconstruction. You cannot pose, regularize, or debug the inverse map without owning the forward vocabulary — which is exactly what this module builds.

Core Problem

Every inverse method on this site is an instance of the observation model $$\mathbf{y} = \mathcal{H}(\mathbf{u}, p, f, X_0) + b + \eta,$$ where $\mathbf{y}$ is what you measure and $\mathcal{H}$ maps the displacement field $\mathbf{u}$, material parameters $p$, loads $f$, and reference configuration $X_0$ to the ideal measurements those would produce; the bias $b$ and the noise $\eta$ are added on top of $\mathcal{H}$, not passed into it. This module supplies the vocabulary inside $\mathcal{H}$: what $\mathbf{u}$, strain, and stress are, and which relations between them are exact versus modeled.

Three ingredient types make up the forward map, and it matters enormously which is which:

Ingredient	What it is	Status
Kinematics	The geometry of deformation: configurations, $\mathbf{F}$, strain measures	Material-independent, exact
Balance / stress	Internal force bookkeeping and equilibrium	Material-independent, exact
Constitutive law	The material-specific bridge from strain to stress	The only place modeling error about the material enters

Why it matters

When an inverse reconstruction disagrees with reality, this three-way split tells you where to look. Kinematics and balance cannot be wrong — they are mathematics. Every modeling sin about the material lives in the constitutive law, and every modeling sin about the deformation regime lives in pretending small-strain formulas hold at large deformation.

Roadmap. Kinematics first: configurations, the deformation gradient $\mathbf{F}$, and strain measures (§2–4). Then compatibility (§5) — the integrability condition that turns out to be the hidden regularizer of iFEM shape sensing. Then stress measures (§6), linear elasticity and its engineering machinery — the Hooke tensor, Voigt notation, plane problems (§7–9) — and finally finite-strain material models: strain energy, neo-Hookean, Mooney–Rivlin, incompressibility (§10–11), closing with a decision guide (§12).

The site spans two regimes, and this primer covers both ends. Classic iFEM (Tessler & Spangler 2005) is built on linear, small-strain, first-order shear-deformation shell theory, which places it in regime (1) of §12 — stiff structures at small strain and small rotation. Vision-based tactile skins are built from Ecoflex/DragonSkin-class silicones, which are soft and near-incompressible over a wide range of stretch (Liao, Hossain & Yao 2020 characterize Ecoflex across Shore hardnesses in uniaxial tension) and are deliberately made to deform visibly under contact. Visible motion is not by itself the test: a thin skin can displace and rotate a great deal while straining little — that is regime (2) of §12, and §4 spells out why. Classify a particular device from strains and rotations you have measured or computed, not from how much it appears to move. What does put a tactile skin in finite-strain territory is the ratio: a fingertip indenting a soft pad by an appreciable fraction of its own thickness produces strains well past the few percent where linear theory holds. How large they get is device-specific: treat it as something to measure for your own skin rather than a number to look up. Crucially, this module tells you when each applies.

2. Configurations, Motion, and Displacement

A deforming body occupies two configurations we care about. The reference (material, undeformed) configuration $\Omega_0$ carries material points labeled $\mathbf{X}$ — think of $\mathbf{X}$ as a particle's name. The current (spatial, deformed) configuration $\Omega$ contains the positions $\mathbf{x}$ those particles occupy now. The motion is a smooth, invertible map, and displacement is the difference:

Motion and Displacement

$$\mathbf{x} = \boldsymbol{\varphi}(\mathbf{X}, t), \qquad \mathbf{u}(\mathbf{X}, t) = \boldsymbol{\varphi}(\mathbf{X}, t) - \mathbf{X}$$ Convention used site-wide (Holzapfel; Bonet & Wood): uppercase for referential quantities ($\mathbf{X}$, $\mathbf{N}$, $dA$, $\operatorname{Grad} = \nabla_{\mathbf{X}}$), lowercase for spatial ($\mathbf{x}$, $\mathbf{n}$, $da$, $\operatorname{grad} = \nabla_{\mathbf{x}}$).

The two-configuration picture: the motion $\boldsymbol{\varphi}$ carries the material point $\mathbf{X}$ to its current position $\mathbf{x}$; the displacement $\mathbf{u}$ connects them. Infinitesimal line, volume, and area elements transform through $\mathbf{F}$, $J = \det\mathbf{F}$, and Nanson's formula respectively (§3).

Material vs Spatial Description

A field can be written as a function of $\mathbf{X}$ — the material (Lagrangian) description, where you follow particles — or as a function of $\mathbf{x}$ — the spatial (Eulerian) description, where you watch a fixed location in space. Solid mechanics and FEM are overwhelmingly Lagrangian: the reference configuration is known (it is your undeformed mesh), and boundary conditions attach to material particles. Fluids, having no preferred stress-free configuration, are usually described in Eulerian form — though Lagrangian and ALE descriptions are standard in free-surface and fluid–structure problems.

Key insight

In FEM the displacement field $\mathbf{u}$ is the primary unknown, and that is why iFEM reconstructs $\mathbf{u}$ first and reads everything else off afterwards. Be precise about what "everything else" costs. Kinematics — $\mathbf{F}$, strain, $J$ — follows from $\mathbf{u}$ alone. Stress follows from $\mathbf{u}$ alone only for a compressible displacement-based constitutive law; under exact incompressibility the reaction pressure $p$ is a separate unknown fixed by equilibrium and boundary conditions rather than by $\mathbf{u}$ (§11), so the primal field is really the pair $(\mathbf{u}, p)$. Contact loads cost more still: equilibrium plus the contact conditions on the surface. Recover $\mathbf{u}$ and the kinematics come free; the derived stress and load fields come with a constitutive law attached, and sometimes with extra unknowns and extra equations attached too.

Two Examples to Carry Through the Module

Homogeneous uniaxial stretch. $x_1 = \lambda X_1$, $x_2 = X_2$, $x_3 = X_3$, so $\mathbf{u} = ((\lambda - 1)X_1,\, 0,\, 0)$. Simple, homogeneous, and every strain measure has a clean closed form for it.

Rigid-body motion. $\mathbf{x} = \mathbf{R}\mathbf{X} + \mathbf{c}$ with $\mathbf{R}$ a rotation. The body moves but does not deform: zero strain. This is the acid test every strain measure must pass — and, as §4 shows, the small-strain tensor fails it.

Invertibility. The motion $\boldsymbol{\varphi}$ must be one-to-one with $\det(\partial\boldsymbol{\varphi}/\partial\mathbf{X}) > 0$: no interpenetration of matter, no local inversion. This is not a technicality — it is a physical constraint that every reconstructed displacement field should respect too. An inverse solver that returns a field with negative local volume has produced nonsense, however small its data residual.

3. The Deformation Gradient F

Differentiate the motion once and you get the fundamental kinematic object of finite deformation:

Deformation Gradient

$$\mathbf{F} = \frac{\partial \mathbf{x}}{\partial \mathbf{X}} = \mathbf{I} + \nabla_{\mathbf{X}} \mathbf{u}, \qquad d\mathbf{x} = \mathbf{F}\, d\mathbf{X}$$ $\mathbf{F}$ is a linear map sending each infinitesimal material line element $d\mathbf{X}$ to its deformed image $d\mathbf{x}$. It is a two-point tensor — one leg in each configuration. All strain measures are built from $\mathbf{F}$.

Volumes and areas transform through $\mathbf{F}$ as well:

J = \det \mathbf{F} > 0, \qquad dv = J\, dV, \qquad \mathbf{n}\, da = J \mathbf{F}^{-\mathsf{T}} \mathbf{N}\, dA

$J > 0$ encodes local orientation preservation — the pointwise half of the §2 requirement; global injectivity of $\boldsymbol{\varphi}$ is a separate condition, since a motion can be everywhere orientation-preserving and still fold the body onto itself. $J = 1$ means locally volume-preserving (isochoric) deformation. The area relation is Nanson's formula — file it away, because it is exactly what connects stress measures across configurations in §6.

Polar Decomposition: Every Deformation is Stretch Plus Rotation

Polar Decomposition

$$\mathbf{F} = \mathbf{R}\mathbf{U} = \mathbf{v}\mathbf{R}, \qquad \mathbf{C} = \mathbf{F}^{\mathsf{T}}\mathbf{F} = \mathbf{U}^2$$ Every local deformation is a pure stretch $\mathbf{U}$ (symmetric positive definite, living in the reference frame) followed by a rigid rotation $\mathbf{R}$ — or a rotation first, then the spatial stretch $\mathbf{v}$. $\mathbf{U}$ is computed from the right Cauchy–Green tensor $\mathbf{C} = \mathbf{F}^{\mathsf{T}}\mathbf{F}$; the eigenvalues of $\mathbf{U}$ are the principal stretches $\lambda_1, \lambda_2, \lambda_3$, its eigenvectors the principal directions.

The polar-decomposition commuting square, drawn for $\lambda_1 = 1.5$, $\lambda_2 = 0.75$, $\theta = 25°$. Upper path: stretch $\mathbf{U}$ first (square $\to$ rectangle, embedded circle $\to$ axis-aligned ellipse with semi-axes $\lambda_1, \lambda_2$), then rotate by $\mathbf{R}$. Lower path: rotate first, then apply the spatial stretch $\mathbf{v}$. Both arrive at the same state: $\mathbf{F} = \mathbf{R}\mathbf{U} = \mathbf{v}\mathbf{R}$. Since $\mathbf{C} = \mathbf{F}^{\mathsf{T}}\mathbf{F} = \mathbf{U}^2$, the tensor $\mathbf{C}$ is rotation-free.

Worked Examples with Numbers

(a) Uniaxial stretch. $\mathbf{F} = \operatorname{diag}(\lambda, 1, 1)$, so $J = \lambda$. Here $\mathbf{R} = \mathbf{I}$: the deformation is pure stretch.

(b) Simple shear. $x_1 = X_1 + \gamma X_2$ gives, in the $1$–$2$ plane,

\mathbf{F} = \begin{bmatrix} 1 & \gamma \\ 0 & 1 \end{bmatrix}, \qquad J = 1, \qquad \mathbf{C} = \begin{bmatrix} 1 & \gamma \\ \gamma & 1 + \gamma^2 \end{bmatrix}

Volume is exactly preserved ($J = 1$), yet $\mathbf{C} \neq \mathbf{I}$: the material is genuinely strained. Moreover $\mathbf{C}$ has unequal principal values, and the polar decomposition of this $\mathbf{F}$ contains a nontrivial rotation — simple shear is not "pure" shear; it is a pure shear plus an embedded rotation.

(c) Pure rotation. $\mathbf{F} = \mathbf{R}$, so $\mathbf{C} = \mathbf{R}^{\mathsf{T}}\mathbf{R} = \mathbf{I}$ and $\mathbf{U} = \mathbf{I}$: zero strain, as required.

Key insight

$\mathbf{F}$ contains rotation. A strain measure must filter the rotation out — that is the entire plot of the next section, and the geometric picture to hold on to is the embedded unit circle mapping to an ellipse with semi-axes $\lambda_1, \lambda_2$ along rotated principal directions. The interactive lab below animates exactly this.

4. Strain Measures: Small Strain vs Green–Lagrange

What should a strain measure do? Two things: vanish for rigid motion, and reduce to intuitive elongation for small stretch. The natural finite-strain candidate subtracts the identity from the rotation-free tensor $\mathbf{C}$:

Green–Lagrange Strain

$$\mathbf{E} = \tfrac{1}{2}(\mathbf{C} - \mathbf{I}) = \tfrac{1}{2}\left(\nabla_{\mathbf{X}}\mathbf{u} + \nabla_{\mathbf{X}}\mathbf{u}^{\mathsf{T}} + \nabla_{\mathbf{X}}\mathbf{u}^{\mathsf{T}}\,\nabla_{\mathbf{X}}\mathbf{u}\right)$$ For rigid motion $\mathbf{F} = \mathbf{R}$: $\mathbf{C} = \mathbf{R}^{\mathsf{T}}\mathbf{R} = \mathbf{I}$, so $\mathbf{E} = \mathbf{0}$ exactly — the rigid-motion test is passed.

Drop the quadratic term and you get the small-strain (infinitesimal) tensor — the linearization of $\mathbf{E}$ about the undeformed state:

\boldsymbol{\varepsilon} = \tfrac{1}{2}\left(\nabla \mathbf{u} + \nabla \mathbf{u}^{\mathsf{T}}\right)

Key insight — the failure mode study material glosses over

$\boldsymbol{\varepsilon}$ is valid only when strains and rotations are both small — small-strain kinematics are not objective under finite rigid rotation. For a pure rotation by angle $\theta$: $\varepsilon_{11} = \varepsilon_{22} = \cos\theta - 1 \approx -\theta^2/2$. A rigid $10°$ rotation produces a spurious $-1.52\%$ compressive strain — larger than real working strains in stiff structures. Thin flexible structures (shells, skins) can rotate a lot while straining little; this is why geometrically nonlinear analysis exists even for linear materials. "Small strain" does not mean "small displacement."

One Deformation, Many Numbers

For uniaxial stretch $\lambda$, the common scalar measures are

E_{11} = \tfrac{1}{2}(\lambda^2 - 1), \qquad \varepsilon_{\text{eng}} = \lambda - 1, \qquad \varepsilon_{\ln} = \ln \lambda

Stretch $\lambda$	Engineering $\lambda - 1$	Green–Lagrange $\tfrac{1}{2}(\lambda^2-1)$	Logarithmic $\ln\lambda$
1.1	0.100	0.105	0.0953
1.5	0.500	0.625	0.405
2.0	1.000	1.500	0.693

At $\lambda = 1.1$ each measure differs from engineering strain by about 5%, so the full spread — logarithmic 0.0953 to Green–Lagrange 0.105 — is about 10%. At $\lambda = 2$ they are wildly different numbers for the same physical deformation. Note in particular that at 50% stretch, $E_{11} = 0.625$, not $0.5$: presenting Green–Lagrange components as percent elongation is wrong. A "strain" value at large deformation is meaningless without naming the measure.

Four strain measures against stretch $\lambda$: engineering $\lambda - 1$ (straight line), Green–Lagrange $\tfrac{1}{2}(\lambda^2 - 1)$ (steepest parabola), logarithmic $\ln\lambda$ (the only curve antisymmetric in tension/compression), and Euler–Almansi $\tfrac{1}{2}(1 - \lambda^{-2})$ (flattest in tension, steepest in compression). All four share a common tangent at $\lambda = 1$ (inset): agreeing to first order in $\nabla\mathbf{u}$ is what "small strain" means. Guide lines at $\lambda = 1.1$ and $\lambda = 1.5$ mark the values from the table.

Spatial Counterparts

Just as $\mathbf{C} = \mathbf{F}^{\mathsf{T}}\mathbf{F}$ lives in the reference configuration, the left Cauchy–Green tensor $\mathbf{b} = \mathbf{F}\mathbf{F}^{\mathsf{T}}$ lives in the current one, and its associated strain is the Euler–Almansi strain:

\mathbf{e} = \tfrac{1}{2}\left(\mathbf{I} - \mathbf{b}^{-1}\right), \qquad \mathbf{b} = \mathbf{F}\mathbf{F}^{\mathsf{T}}

$\mathbf{e}$ is the spatial counterpart of $\mathbf{E}$, and it linearizes to the same $\boldsymbol{\varepsilon}$. That is the general pattern: all these measures agree to first order in $\nabla\mathbf{u}$ — that agreement is what "small strain" means. Full derivations: Holzapfel ch. 2; Bonet & Wood ch. 4.

Interactive Tool — Strain-Convention Trap Lab

Deform a material patch with the sliders (the deformation is composed as $\mathbf{F} = \mathbf{R}(\theta)\,\mathbf{S}$: stretch/shear $\mathbf{S}$ first, rotation last) and compare what the Green–Lagrange tensor $\mathbf{E}$ and the small-strain tensor $\boldsymbol{\varepsilon}$ report. The patch is embedded as plane strain, $F_{33} = 1$, so the in-plane determinant shown below really is the volume ratio $J$ and not merely an area ratio. The ledger on the right keeps the inverse-problem bookkeeping honest: what a strain rosette measures, what is unknown, what is assumed, what is observable, and how large the convention error is. Try the presets.

Unrotated gradient $S_{11}$ 1.00

Unrotated gradient $S_{22}$ 1.00

Unrotated gradient $S_{12}$ 0.00

Rotation $\theta$ 0°

5. Compatibility: When Is a Strain Field Integrable?

Count components: six strain components derive from only three displacement components. So an arbitrary symmetric tensor field is generally not a strain field — most candidate "strain fields" correspond to no displacement field at all. The integrability requirements are the Saint-Venant compatibility conditions:

Saint-Venant Compatibility

$$\operatorname{inc}(\boldsymbol{\varepsilon}) = \nabla \times \boldsymbol{\varepsilon} \times \nabla = \mathbf{0}, \qquad \operatorname{inc}(\boldsymbol{\varepsilon})_{ij} = \epsilon_{ikm}\,\epsilon_{jln}\,\partial_k \partial_l\, \varepsilon_{mn}$$ In compact form: the incompatibility tensor vanishes — one curl on each index of $\boldsymbol{\varepsilon}$, equivalently $\operatorname{Curl}[(\operatorname{Curl}\boldsymbol{\varepsilon})^{\mathsf{T}}]$, with $\epsilon_{ikm}$ the permutation symbol and not a strain component. Curling twice on the same index is a different operator, $\partial_i\partial_k\varepsilon_{kj} - \nabla^2\varepsilon_{ij}$, which does not vanish for a compatible field; only the form above reduces to the 2D equation below. Since $\operatorname{inc}(\boldsymbol{\varepsilon})$ is symmetric, componentwise these are 6 distinct second-order PDEs in 3D, and all six have to be imposed. Their components do obey the differential identity $\operatorname{Div}\operatorname{inc}\boldsymbol{\varepsilon} = \mathbf{0}$ — the Bianchi-type identity routinely quoted as cutting the count to 3 — but that is a relation among derivatives of the six equations and does not license discarding three of them; counterexamples to the six-to-three reduction are given by Georgiyevskii & Pobedrya (2004). In 2D the single surviving component is $\operatorname{inc}(\boldsymbol{\varepsilon})_{33}$: $$\frac{\partial^2 \varepsilon_{11}}{\partial x_2^2} + \frac{\partial^2 \varepsilon_{22}}{\partial x_1^2} = 2\,\frac{\partial^2 \varepsilon_{12}}{\partial x_1 \partial x_2}$$ Verify it in one line: substitute $\varepsilon_{11} = u_{1,1}$, $\varepsilon_{22} = u_{2,2}$, $\varepsilon_{12} = \tfrac{1}{2}(u_{1,2} + u_{2,1})$ for any smooth $\mathbf{u}$ and both sides equal $u_{1,122} + u_{2,112}$.

Sufficiency comes with a caveat. On a simply-connected domain, compatibility guarantees a single-valued displacement field, reconstructable by the Cesàro path integral and unique up to rigid motion. On multiply-connected domains — a skin patch with a hole, a ring — the conditions are necessary but not sufficient: path integrals around the hole can be multivalued (dislocations exploit exactly this). Classical treatment: Timoshenko & Goodier.

Why This Section Sits Here: the Bridge to Shape Sensing

Measured strains are pointwise, sparse, and noisy. Compatibility constrains the derivatives of a continuous strain field, so a handful of point readings violate nothing by themselves — some smooth displacement field generically interpolates any finite set of them. Two things do fail. A dense strain field naively interpolated from sparse noisy readings is essentially never compatible; and within a chosen finite-element trial space, no displacement field generally reproduces every reading exactly. The reason is not that the fit is overdetermined — an overdetermined system can be perfectly consistent, and a sparse iFEM system can even be underdetermined and admit many exact fits. Exact reproduction fails precisely when the measurement vector lies outside the range of the discrete strain-observation operator, which noisy data generically does. Direct spatial integration of such an interpolated field is therefore path-dependent and accumulates noise along the integration path.

Key insight — compatibility is iFEM's hidden regularizer

iFEM sidesteps incompatibility structurally. It parametrizes the displacement field with finite-element shape functions and minimizes a least-squares functional between the strains of that trial field and the measured strains (Tessler & Spangler 2005). Any strain field derived from an interpolated displacement is compatible by construction — so iFEM is a projection of noisy data onto the compatible manifold. Compatibility is the hidden regularizer that removes the incompatible directions from the solution space — well-posedness needs more on top: pose anchoring, adequate sensor rank, and often explicit regularization (Modules 3–4). This is the single most important conceptual bridge from this primer to the iFEM modules: Module 4 builds the least-squares functional on exactly this observation.

6. Stress Measures: Cauchy, First and Second Piola–Kirchhoff

Cut a body along any internal surface and the two sides exert forces on each other. Cauchy's theorem says the traction — force per unit current area — on a surface with normal $\mathbf{n}$ is linear in the normal:

Cauchy Stress and Equilibrium

$$\mathbf{t} = \boldsymbol{\sigma}\,\mathbf{n}, \qquad \operatorname{div}\boldsymbol{\sigma} + \mathbf{f}_v = \mathbf{0}$$ This defines the Cauchy ("true") stress $\boldsymbol{\sigma}$, with $\mathbf{f}_v$ the body force per unit current volume (e.g. $\rho\,\mathbf{g}$). Many texts call it $\mathbf{b}$; this page does not, because $\mathbf{b} = \mathbf{F}\mathbf{F}^{\mathsf{T}}$ is the left Cauchy–Green tensor everywhere else here (§4, §10–11). Balance of angular momentum makes $\boldsymbol{\sigma}$ symmetric; balance of linear momentum gives the equilibrium PDE — the equation that FEM weak forms discretize (Module 2).

At finite deformation, "force per area" is ambiguous — which area, deformed or undeformed? Each answer defines a different stress measure, related through $\mathbf{F}$ and Nanson's formula:

\mathbf{P} = J\,\boldsymbol{\sigma}\,\mathbf{F}^{-\mathsf{T}}, \qquad \mathbf{S} = \mathbf{F}^{-1}\mathbf{P} = J\,\mathbf{F}^{-1}\boldsymbol{\sigma}\,\mathbf{F}^{-\mathsf{T}}, \qquad \boldsymbol{\sigma} = J^{-1}\,\mathbf{F}\,\mathbf{S}\,\mathbf{F}^{\mathsf{T}}

First Piola–Kirchhoff stress $\mathbf{P}$: the actual force referred to reference area. It is a two-point tensor and generally unsymmetric. Its uniaxial component is exactly the engineering/nominal stress that a load cell plus the initial cross-section gives you.
Second Piola–Kirchhoff stress $\mathbf{S}$: fully referential and symmetric, but its components have no direct physical meaning — the force vector has been pulled back by $\mathbf{F}^{-1}$.

Three bookkeeping conventions for the same physics. Cauchy $\boldsymbol{\sigma}$ divides the force by the current area; first Piola–Kirchhoff $\mathbf{P}$ divides the same force by the reference area; second Piola–Kirchhoff $\mathbf{S}$ additionally pulls the force vector back by $\mathbf{F}^{-1}$, gaining symmetry but losing physical interpretability.

Why S Exists at All: Work Conjugacy

If $\mathbf{S}$ has no physical reading, why keep it? Because of work conjugacy. The internal power density per unit reference volume can be written two equivalent ways, and it relates to the spatial density through $J$; integrating once gives the power of the whole body:

\mathbf{P} : \dot{\mathbf{F}} = \mathbf{S} : \dot{\mathbf{E}} = J\,\boldsymbol{\sigma} : \mathbf{d}, \qquad \text{so} \qquad \dot{W}_{\text{int}} = \int_{\Omega_0} \mathbf{P} : \dot{\mathbf{F}}\; dV = \int_{\Omega} \boldsymbol{\sigma} : \mathbf{d}\; dv

with $\mathbf{d}$ the rate of deformation. The pair $(\mathbf{S}, \mathbf{E})$ makes constitutive theory clean — both are objective material tensors — so material laws are naturally written $\mathbf{S}(\mathbf{E})$ or $\mathbf{S} = 2\,\partial\Psi/\partial\mathbf{C}$ (§10 adds the constraint-reaction term this last form needs when incompressibility is enforced exactly), then pushed forward to $\boldsymbol{\sigma}$ for output.

Key insight — pair stresses and strains from the same configuration

The conjugate pairs are $(\mathbf{P}, \dot{\mathbf{F}})$ and $(\mathbf{S}, \dot{\mathbf{E}})$ per reference volume, and $(\boldsymbol{\sigma}, \mathbf{d})$ per current volume. Never pair Cauchy stress with Green–Lagrange strain, and never pair second Piola–Kirchhoff stress with a spatial strain: the products are not work densities and a constitutive law written across mismatched configurations is dimensionally plausible nonsense. Likewise, never read $\mathbf{S}$ components as physical tractions — push forward to $\boldsymbol{\sigma}$ before comparing with measured contact pressures.

Concrete Numbers

Stretch an incompressible rod to $\lambda = 2$. Incompressibility forces the current area to $A = A_0/\lambda$, so

\sigma = \lambda\, P \quad \text{(incompressible uniaxial: at } \lambda = 2,\ \sigma = 2P\text{)}

True stress is twice nominal at $\lambda = 2$. A hyperelastic fit is meaningless if you don't know which stress the data sheet reports. Division of labor in practice: experiments report $\mathbf{P}$; total-Lagrangian FE codes compute $\mathbf{S}$ internally; contact pressures and material failure arguments live in $\boldsymbol{\sigma}$. At small strain all three coincide to first order ($J \approx 1$, $\mathbf{F} \approx \mathbf{I}$). Full treatment: Holzapfel ch. 3; Bonet & Wood ch. 5; Ogden ch. 3.

7. Linear Elasticity and the Hooke Tensor

Now the constitutive layer, starting at small strain. The most general linear relation between two symmetric second-order tensors is a fourth-order tensor:

Generalized Hooke's Law

$$\sigma_{ij} = C_{ijkl}\,\varepsilon_{kl}, \qquad C_{ijkl} = C_{jikl} = C_{ijlk} = C_{klij}$$ Raw count: 81 components. The minor symmetries ($\boldsymbol{\sigma}$ and $\boldsymbol{\varepsilon}$ symmetric) cut this to 36; the major symmetry $C_{ijkl} = C_{klij}$ — equivalent to the existence of a strain energy $\tfrac{1}{2}\boldsymbol{\varepsilon}:\mathbb{C}:\boldsymbol{\varepsilon}$ — cuts it to 21: the fully anisotropic (triclinic) material.

Isotropy collapses 21 constants to two, the Lamé parameters:

\boldsymbol{\sigma} = \lambda\,\mathrm{tr}(\boldsymbol{\varepsilon})\,\mathbf{I} + 2\mu\,\boldsymbol{\varepsilon}

The engineering constants and the full conversion set:

G = \frac{E}{2(1+\nu)}, \qquad K = \frac{E}{3(1-2\nu)}, \qquad \lambda = \frac{E\nu}{(1+\nu)(1-2\nu)}

with inverses $E = \mu(3\lambda + 2\mu)/(\lambda + \mu)$ and $\nu = \lambda/(2(\lambda + \mu))$, and $G \equiv \mu$. Physical readings: $E$ is the slope of the uniaxial stress–strain curve; $\nu$ the lateral contraction ratio; $G$ the shear stiffness; $K$ the pressure per relative volume change. A useful energy split for isotropy — volumetric and deviatoric responses decouple:

\Psi = \tfrac{1}{2}K\,(\mathrm{tr}\,\boldsymbol{\varepsilon})^2 + \mu\,\operatorname{dev}(\boldsymbol{\varepsilon}) : \operatorname{dev}(\boldsymbol{\varepsilon})

Key insight — the incompressible limit

Isotropic stability admits $-1 < \nu < \tfrac{1}{2}$. As $\nu \to \tfrac{1}{2}$, both $K/G \to \infty$ and $\lambda \to \infty$: volumetric stiffness diverges while shear stiffness stays finite. Setting $\nu = 0.5$ exactly in a standard displacement formulation is a bug — $\lambda$ and the plane-strain stiffness contain $1/(1-2\nu)$ and become singular — and even near-incompressibility makes standard displacement FEM lock volumetrically. The remedies (mixed $u$–$p$ formulations, selective reduced integration) appear in later modules; the mechanical root of the problem appears again in §10.

Orders of Magnitude

Material	Young's modulus $E$	Poisson ratio $\nu$
Steel	$\approx 200$ GPa	$\approx 0.3$
Aluminum	$\approx 70$ GPa	$\approx 0.3$
PDMS	$\approx$ 1–3 MPa	near 0.5
Ecoflex-class platinum-cure silicones	initial stiffness in the tens-of-kPa range (Liao, Hossain & Yao 2020)	$\approx 0.499$ — a numerical stand-in for near-incompressibility, not a measured transverse-contraction value

Six decades softer than steel — which is why tactile skins deform measurably under fingertip forces, and why the same fingertip force that a steel panel would not notice drives an Ecoflex pad deep into the finite-strain regime.

Validity envelope. Linear elasticity assumes small strain and small rotation and linear material response. For elastomers it is only the tangent at the origin of a strongly nonlinear curve — a useful tangent (§11 computes it: $E = 3\mu$), but a tangent.

8. Voigt Notation and the Factor-of-2 Shear Trap

FEM implementations flatten symmetric tensors to vectors. Standard Voigt ordering stacks stress and strain as

\boldsymbol{\sigma} = [\sigma_{11}, \sigma_{22}, \sigma_{33}, \sigma_{23}, \sigma_{13}, \sigma_{12}]^{\mathsf{T}}, \qquad \boldsymbol{\varepsilon} = [\varepsilon_{11}, \varepsilon_{22}, \varepsilon_{33}, \gamma_{23}, \gamma_{13}, \gamma_{12}]^{\mathsf{T}}

where $\gamma_{ij} = 2\varepsilon_{ij}$ is the engineering shear strain. The 2 is not a convention accident:

\gamma_{ij} = 2\,\varepsilon_{ij} \; (i \neq j), \qquad \boldsymbol{\sigma}^{\mathsf{T}}\boldsymbol{\varepsilon}_{\text{Voigt}} = \boldsymbol{\sigma} : \boldsymbol{\varepsilon}

It is exactly what makes the vector dot product reproduce the tensor work density, because each off-diagonal pair $(ij, ji)$ contributes twice to $\boldsymbol{\sigma}:\boldsymbol{\varepsilon}$. With that convention, the full isotropic stiffness matrix reads:

\begin{bmatrix}\sigma_{11}\\\sigma_{22}\\\sigma_{33}\\\sigma_{23}\\\sigma_{13}\\\sigma_{12}\end{bmatrix} = \frac{E}{(1{+}\nu)(1{-}2\nu)} \begin{bmatrix} 1{-}\nu & \nu & \nu & 0 & 0 & 0\\ \nu & 1{-}\nu & \nu & 0 & 0 & 0\\ \nu & \nu & 1{-}\nu & 0 & 0 & 0\\ 0&0&0& \tfrac{1-2\nu}{2} &0&0\\ 0&0&0&0& \tfrac{1-2\nu}{2} &0\\ 0&0&0&0&0& \tfrac{1-2\nu}{2} \end{bmatrix} \begin{bmatrix}\varepsilon_{11}\\\varepsilon_{22}\\\varepsilon_{33}\\\gamma_{23}\\\gamma_{13}\\\gamma_{12}\end{bmatrix}

Check the shear block: $\frac{E}{(1+\nu)(1-2\nu)} \cdot \frac{1-2\nu}{2} = \frac{E}{2(1+\nu)} = G$. The shear entries equal $G$ only because the strain vector carries $\gamma$, not $\varepsilon_{12}$. In the compliance form $\boldsymbol{\varepsilon} = \mathbf{S}\boldsymbol{\sigma}$ the shear diagonal is $1/G$ — the factor 2 migrates to the other side.

Why it matters — two silent bugs

Trap 1: the factor of 2. An FE B-matrix produces $\gamma$; if a material routine or a strain-gauge rosette formula expects tensor shear (or vice versa), you silently halve or double all shear response. No error message, no crash — just wrong stiffness in shear. This is a classic bug in home-grown iFEM/FEM code, and it resurfaces in Module 6 where rosette transforms must use the same engineering-shear convention as the FE strain vector.

Trap 2: the ordering. Voigt component ordering is not universal: classical texts use $(11, 22, 33, 23, 13, 12)$ while e.g. Abaqus uses $(11, 22, 33, 12, 13, 23)$. Porting a stiffness matrix between conventions scrambles the shear terms without any error message. Always check the ordering before porting.

Mandel notation: the norm-preserving alternative

Mandel notation scales the off-diagonal components by $\sqrt{2}$ in both the stress and strain vectors. The payoff: the vector 2-norm equals the tensor norm and eigenvalues are preserved, which makes spectral algorithms (eigen-decompositions of stiffness, projections) clean. The cost: components no longer read as engineering quantities. Voigt remains the FEM default; Mandel is worth knowing when you meet norm-based algorithms.

9. Plane Stress and Plane Strain

Two classical 2D reductions of 3D elasticity (Timoshenko & Goodier ch. 2). They are different assumptions with different reduced stiffness matrices — swapping them is not a rounding error.

Plane stress assumes $\sigma_{33} = \sigma_{13} = \sigma_{23} = 0$: thin sheets loaded in-plane with free faces. The sheet is free to thin, so $\varepsilon_{33}$ is left unconstrained (and is generally nonzero):

\text{Plane stress:}\quad \begin{bmatrix}\sigma_{11}\\\sigma_{22}\\\sigma_{12}\end{bmatrix} = \frac{E}{1-\nu^2} \begin{bmatrix}1 & \nu & 0\\ \nu & 1 & 0\\ 0 & 0 & \tfrac{1-\nu}{2}\end{bmatrix} \begin{bmatrix}\varepsilon_{11}\\\varepsilon_{22}\\\gamma_{12}\end{bmatrix}

Plane strain assumes $\varepsilon_{33} = \varepsilon_{13} = \varepsilon_{23} = 0$: long prismatic bodies or layers constrained against out-of-plane motion. The constraint must then carry a stress $\sigma_{33}$, generally nonzero:

\text{Plane strain:}\quad \begin{bmatrix}\sigma_{11}\\\sigma_{22}\\\sigma_{12}\end{bmatrix} = \frac{E}{(1{+}\nu)(1{-}2\nu)} \begin{bmatrix}1{-}\nu & \nu & 0\\ \nu & 1{-}\nu & 0\\ 0 & 0 & \tfrac{1-2\nu}{2}\end{bmatrix} \begin{bmatrix}\varepsilon_{11}\\\varepsilon_{22}\\\gamma_{12}\end{bmatrix}

The out-of-plane bookkeeping in each case:

\varepsilon_{33} = -\frac{\nu}{1-\nu}\left(\varepsilon_{11}+\varepsilon_{22}\right) \;\text{(plane stress)}, \qquad \sigma_{33} = \nu\left(\sigma_{11}+\sigma_{22}\right) \;\text{(plane strain)}

Note the $1 - 2\nu$ in the plane-strain prefactor: plane strain stiffens dramatically as $\nu \to \tfrac{1}{2}$ and blows up in the incompressible limit, while plane stress stays finite. A substitution rule worth memorizing: any plane-stress solution converts to plane strain via $E \to E/(1-\nu^2)$, $\nu \to \nu/(1-\nu)$.

$\nu$	Plane-stress prefactor $\frac{E}{1-\nu^2}$	Plane-strain $(1,1)$ entry $\frac{E(1-\nu)}{(1+\nu)(1-2\nu)}$
0.3	$1.10\,E$	$1.35\,E$
0.49	$1.32\,E$	$17.1\,E$

At $\nu = 0.3$ the choice changes stiffness by ~20%; at $\nu = 0.49$ by a factor of 13. For rubber, the choice is not a detail.

Left and middle: the two classical reductions and their assumptions. Right: a thin elastomer layer bonded to a rigid substrate under indentation fits neither — it is closer to confined compression, and the near-incompressible material cannot escape sideways.

Why it matters — which applies to a tactile skin?

Honestly, often neither. A thin elastomer layer bonded to a rigid substrate and indented from above is closer to confined compression: free-surface (plane-stress-like) assumptions underestimate stiffness badly, because the near-incompressible material cannot escape sideways. Full 3D or axisymmetric models are usually needed, and the apparent stiffness becomes geometry-dominated — set by the aspect ratio of the contact versus the layer thickness, not by $E$ alone. This is stated explicitly here because sensor-calibration shortcuts frequently assume plane stress by default.

Closing the loop to iFEM: classic iFEM plate formulations are built on plane-stress-based laminate kinematics (first-order shear deformation theory) — which is part of their small-strain validity envelope, revisited in §12 and Module 4.

10. Hyperelasticity: Stress From a Strain-Energy Density

Beyond a few percent strain, the linear law of §7 stops describing elastomers. The finite-strain constitutive framework starts from a scalar potential.

Definition — Hyperelastic Material

A material is hyperelastic if a scalar strain-energy density $\Psi(\mathbf{F})$ exists with $$\mathbf{P} = \frac{\partial \Psi}{\partial \mathbf{F}}, \qquad \mathbf{S} = 2\,\frac{\partial \Psi}{\partial \mathbf{C}}, \qquad \boldsymbol{\sigma} = \frac{2}{J}\,\mathbf{F}\,\frac{\partial \Psi}{\partial \mathbf{C}}\,\mathbf{F}^{\mathsf{T}}$$ These are the unconstrained (compressible) forms. If the material carries the exact constraint $J = 1$, its multiplier $p$ contributes a reaction term to each: $$\mathbf{P} = \frac{\partial \Psi}{\partial \mathbf{F}} - pJ\mathbf{F}^{-\mathsf{T}}, \qquad \mathbf{S} = 2\,\frac{\partial \Psi}{\partial \mathbf{C}} - pJ\mathbf{C}^{-1}, \qquad \boldsymbol{\sigma} = \frac{2}{J}\,\mathbf{F}\,\frac{\partial \Psi}{\partial \mathbf{C}}\,\mathbf{F}^{\mathsf{T}} - p\,\mathbf{I}$$ Feeding the constrained neo-Hookean energy of §11 into the unconstrained recipe returns $\boldsymbol{\sigma} = \mu\mathbf{b}$ and silently drops the $-p\mathbf{I}$ — the sign convention above is the one that reproduces §11's $\boldsymbol{\sigma} = -p\mathbf{I} + \mu\mathbf{b}$. Stress is the gradient of a potential: work is path-independent and no energy is dissipated in a closed cycle. This follows from the Clausius–Planck inequality with zero dissipation (Holzapfel ch. 6).

Two invariance requirements sculpt what $\Psi$ may depend on:

Objectivity (frame indifference): $\Psi(\mathbf{Q}\mathbf{F}) = \Psi(\mathbf{F})$ for all rotations $\mathbf{Q}$ forces $\Psi = \Psi(\mathbf{C})$ — the energy cannot see the rotation part of $\mathbf{F}$. This is the constitutive echo of §4: only the rotation-free part of the deformation may enter.
Material isotropy: $\Psi(\mathbf{F}\mathbf{Q}) = \Psi(\mathbf{F})$ further reduces $\Psi$ to a function of the three principal invariants, or equivalently of the principal stretches $(\lambda_1, \lambda_2, \lambda_3)$ — Ogden's preferred formulation.

I_1 = \mathrm{tr}\,\mathbf{C}, \qquad I_2 = \tfrac{1}{2}\left[(\mathrm{tr}\,\mathbf{C})^2 - \mathrm{tr}(\mathbf{C}^2)\right], \qquad I_3 = \det \mathbf{C} = J^2

The stress recipes above, pushed through the chain rule over the invariants, give the standard representation of $\boldsymbol{\sigma}$ in powers of $\mathbf{b}$.

Near-Incompressibility: the Isochoric–Volumetric Split

Elastomers strongly resist volume change while shearing freely. The standard handling splits the deformation multiplicatively and the energy additively:

\bar{\mathbf{F}} = J^{-1/3}\mathbf{F}, \qquad \bar{I}_1 = J^{-2/3} I_1, \qquad \bar{I}_2 = J^{-4/3} I_2, \qquad \Psi = \Psi_{\text{iso}}(\bar{I}_1, \bar{I}_2) + U(J)

with a volumetric penalty such as $U(J) = \tfrac{\kappa}{2}(J-1)^2$. For elastomers the scale separation is stark: bulk modulus $\kappa \sim 1$ GPa versus shear modulus $\mu \sim 10$ kPa–1 MPa — a ratio of $10^3$–$10^5$. This is the mechanical root of both the $\nu \to \tfrac{1}{2}$ limit of §7 and volumetric locking.

Why it matters — hyperelasticity is an idealization

Hyperelasticity is the conservative, rate-independent idealization. Real skin silicones exhibit hysteresis, rate dependence, stress relaxation, and Mullins softening — documented systematically for Ecoflex by Liao, Hossain & Yao (2020), with visco-hyperelastic extensions for soft-robotics silicones in Kulkarni, Bayre & Khan (2025). A hyperelastic fit represents one chosen backbone curve of that richer behavior — loading, unloading, or equilibrium — and you should state which. Treating the fit as "the" material behavior of a skin silicone is a modeling decision, not a fact.

11. Neo-Hookean, Mooney–Rivlin, and Incompressibility

The simplest useful $\Psi$ is the incompressible neo-Hookean model:

Incompressible Neo-Hookean

$$\Psi_{\text{NH}} = \frac{\mu}{2}\left(I_1 - 3\right) \;\text{ with constraint } J = 1, \qquad \boldsymbol{\sigma} = -p\,\mathbf{I} + \mu\,\mathbf{b}$$ The constraint's reaction is an indeterminate pressure $p$ — a Lagrange multiplier fixed by boundary conditions, not by the material. Statistical-mechanics pedigree: Gaussian polymer chains give $\mu = nkT$.

Worked Example: the Uniaxial Law

Stretch an incompressible neo-Hookean rod: $\lambda_1 = \lambda$, and incompressibility ($J = \lambda_1\lambda_2\lambda_3 = 1$ with lateral symmetry) forces $\lambda_2 = \lambda_3 = \lambda^{-1/2}$. Traction-free lateral faces require $\sigma_{22} = 0$, which fixes the pressure: $p = \mu/\lambda$. Substituting back:

\sigma_{\text{uni}} = \mu\left(\lambda^2 - \frac{1}{\lambda}\right), \qquad P_{\text{uni}} = \frac{\sigma}{\lambda} = \mu\left(\lambda - \frac{1}{\lambda^2}\right), \qquad E = 3\mu

The tangent at $\lambda = 1$ gives the initial Young's modulus $E = 3\mu$ — consistent with $E = 2\mu(1+\nu)$ at $\nu = \tfrac{1}{2}$.

Key insight — E = 2μ(1+ν) vs its incompressible limit E = 3μ

These are not rival conventions: $E = 2\mu(1+\nu)$ is the general isotropic relation and $E = 3\mu$ is its $\nu \to \tfrac{1}{2}$ limit. Mixing them for elastomers costs a direction-dependent amount — applying $E = 2\mu$ where $E = 3\mu$ holds underestimates the modulus by 33%, while back-computing a shear modulus from a tensile modulus as $\mu = E/2$ instead of $E/3$ overestimates $\mu$ by 50%. A real failure mode when comparing a shear-modulus fit against a tensile-test modulus.

Mooney–Rivlin adds the second invariant. Mooney (1940) derived the form by requiring linearity in simple shear; Rivlin (1948, Part IV) developed the general invariant framework it lives in:

\Psi_{\text{MR}} = C_{10}\left(I_1 - 3\right) + C_{01}\left(I_2 - 3\right), \qquad \mu = 2\left(C_{10} + C_{01}\right)

In uniaxial data the two models are hard to distinguish; what the $I_2$ term mainly changes is the balance between uniaxial and equibiaxial response. It does not change the axial response of standard incompressible planar pure shear at all: there $\mathbf{F} = \operatorname{diag}(\lambda, 1, \lambda^{-1})$ gives $I_1 = I_2 = \lambda^2 + 1 + \lambda^{-2}$, so $\Psi_{\text{MR}} = (C_{10}{+}C_{01})(I_1 - 3) = \tfrac{\mu}{2}(I_1 - 3)$ — exactly the matched-$\mu$ neo-Hookean energy. Axial pure-shear data therefore pin only the sum $C_{10}{+}C_{01}$; the split shows up only in the transverse constraint stress. That is why parameters fitted to uniaxial data alone are badly ill-conditioned: $C_{10}$ and $C_{01}$ trade off almost freely under noise, and the resulting pair extrapolates poorly to equibiaxial and general multiaxial modes — a recurring conclusion of Dal, Açıkgöz & Badienia (2021), comparing 44 models over uniaxial, pure-shear and equibiaxial deformations, and Ricker & Wriggers (2023), fitting protocols over nine rubber compounds; the latter states outright that data from a single experiment is insufficient to calibrate a model carrying both $I_1$- and $I_2$-dependence. One further caveat is this page's own reading rather than a result of either paper: for this two-term incompressible model, $C_{10}, C_{01} \ge 0$ is not merely sufficient for polyconvexity but also necessary. Along the volume-preserving rank-one path $\mathbf{F}(t)$ with diagonal $(a, b, c)$, $abc = 1$, and $F_{12} = t$, the coefficient of $t^2$ in $\Psi_{\text{MR}}$ is $C_{10} + C_{01}c^2$; polyconvexity forces convexity along exactly such paths, and demanding non-negativity for every $c > 0$ forces both constants non-negative. So a fitted $C_{01} < 0$ does rule out global polyconvexity — but it is still not a verdict on the fit, because the model can remain locally elliptic over the bounded deformation range you actually intend to use, which has to be checked there before the fit is deployed.

Mooney–Rivlin in detail: parameterization, uniaxial law, and fitting caveats

The uniaxial true stress for incompressible Mooney–Rivlin follows the same route as the neo-Hookean derivation ($\lambda_2 = \lambda_3 = \lambda^{-1/2}$, lateral faces traction-free):

\sigma_{\text{MR}} = 2\left(C_{10} + \frac{C_{01}}{\lambda}\right)\left(\lambda^2 - \frac{1}{\lambda}\right)

with small-strain shear modulus $\mu = 2(C_{10} + C_{01})$: at $\lambda = 1$ the model is indistinguishable from neo-Hookean with the same $\mu$. The $C_{01}/\lambda$ factor is what separates the models away from $\lambda = 1$ — modestly in uniaxial tension (at matched $\mu$ and $C_{01}/(C_{10}{+}C_{01}) = 0.5$, Mooney–Rivlin is already 25% below neo-Hookean at $\lambda = 2$; check it on the explorer below), strongly in equibiaxial and general multiaxial modes — and not at all in the axial response of planar pure shear, where $I_1 = I_2$ collapses Mooney–Rivlin onto matched-$\mu$ neo-Hookean. Practical consequences:

Ill-conditioning, not non-uniqueness: rearranged, the uniaxial law reads $\sigma / [2(\lambda^2 - \lambda^{-1})] = C_{10} + C_{01}/\lambda$ — affine in $1/\lambda$, so exact data at any two distinct non-unit stretches determines both constants. The pair is identifiable in principle; the trouble is that the two regressors are nearly collinear over any realistic $\lambda$ range, so noise opens a long flat valley of near-equivalent fits that extrapolate badly. Multi-mode data (uniaxial + pure shear + equibiaxial, per Dal et al. 2021 and Ricker & Wriggers 2023) is what conditions the problem.
Stability: for this two-term incompressible model $C_{10}, C_{01} \ge 0$ is both sufficient and necessary for global polyconvexity — the $t^2$ coefficient along the isochoric rank-one path of the main text is $C_{10} + C_{01}c^2$, and it must stay non-negative for every $c > 0$. A fitted $C_{01} < 0$ therefore rules global polyconvexity out; it is still a flag to check ellipticity over the deformation range of interest rather than a verdict on its own, since the model may remain elliptic there.
Strain stiffening: neither neo-Hookean nor Mooney–Rivlin contains a limiting chain extensibility, so neither can reproduce the sharp upturn elastomers show at large stretch. The stretch at which that upturn appears is a property of the specific network (crosslink density, filler content, cure state), not a universal threshold — which is one more reason to fit your own compound rather than reuse a published onset. The standard upgrades are the Ogden, Gent, and Arruda–Boyce families — named here, detailed when they are needed in Modules 7 and 9.

This detail is deliberately tucked away: the material-identification module (Module 7) reintroduces these parameterizations just-in-time, with the fitting machinery to go with them.

Compressible implementation form. FE codes rarely enforce $J = 1$ exactly; they use the split of §10 with a large penalty:

\Psi = \frac{\mu}{2}\left(\bar{I}_1 - 3\right) + \frac{\kappa}{2}\left(J - 1\right)^2, \qquad \kappa/\mu \sim 10^3\text{–}10^5

For Ecoflex-class skin silicones, $\mu$ is of order tens of kPa with strong Shore-hardness dependence (Liao, Hossain & Yao 2020) — always re-fit for your batch and cure state rather than trusting datasheet values.

Interactive Tool — Uniaxial Stress Explorer: True vs Nominal, Linear vs Hyperelastic

Move the operating point $\lambda$ and watch true stress $\sigma$ (solid) and nominal stress $P = \sigma/\lambda$ (dashed) diverge. The gray dashed line is the shared linear tangent $3\mu(\lambda - 1)$ at $\lambda = 1$ — all models and both stress measures collapse onto it locally.

Linear elastic $\sigma = E(\lambda{-}1)$, $E = 3\mu$ Incompressible neo-Hookean Incompressible Mooney–Rivlin

Operating point $\lambda$ 1.50

Shear modulus $\mu$ (kPa) 50

MR ratio $C_{01}/(C_{10}{+}C_{01})$ 0.25

12. Choosing the Right Theory: A Decision Guide

The rest of the site points back to this table. Which regime is your problem in?

Regime	Kinematics	Material law	Machinery
(1) Strains $\lesssim 2\%$ and rotations small	$\boldsymbol{\varepsilon}$ (linear)	Hooke (§7)	Voigt (§8), plane reductions available (§9), and superposition holds — for fixed linear loads and boundary conditions. Unilateral contact, a changing contact active set, and other state-dependent constraints break superposition even when the material stays linear. Home turf of classic iFEM: the least-squares functional of Tessler & Spangler (2005) is written in the strain measures of linear first-order shear-deformation plate theory.
(2) Small strains, large rotations (thin shells, cantilevers, flexible skins bending)	Finite: $\mathbf{E}$ or corotational formulations	May stay linear	$\boldsymbol{\varepsilon}$ alone produces the spurious rotation strains of §4 — geometry must go nonlinear even when the material does not.
(3) Elastomeric strains $\gtrsim 10\%$	Full finite-strain	Hyperelasticity + near-incompressibility (§10–11)	Stress-measure bookkeeping of §6 mandatory; expect volumetric locking in naive displacement elements as $\nu \to \tfrac{1}{2}$ — reach for mixed $u$–$p$ or selective reduced integration.

The tactile-skin verdict

A fingertip pressing millimeters into a millimeters-thick Ecoflex pad sits squarely in regime (3). Transplanting small-strain iFEM formulations to soft skins imports a modeling error that no amount of sensor accuracy fixes — extending deformation reconstruction to finite strains is an active research front, not settled machinery.

Forward dependencies of this module. The FEM discretization of Module 2 needs §6's equilibrium in weak form, §7's Hooke law, and §8's Voigt machinery. The iFEM core (Module 4) needs §5's compatibility-as-regularization insight plus the small-strain validity envelope. Material-identification content (Module 7) needs §10–11 and the fitting caveats: multi-mode data and stress-measure hygiene.

The one-sentence takeaway: kinematics is exact, balance is exact — so every material modeling sin lives in the constitutive law, and every kinematic-regime sin in pretending small-strain formulas hold at large deformation. That split is about the material and the regime only; geometry, boundary conditions, contact, dimensional reduction (§9), discretization, and the observation model $\mathcal{H}$ of §1 are the other places error enters a forward or inverse model.

Flashcards

References

Gerhard A. Holzapfel (2000). Nonlinear Solid Mechanics: A Continuum Approach for Engineering. John Wiley & Sons, Chichester.
Javier Bonet, Richard D. Wood (2008). Nonlinear Continuum Mechanics for Finite Element Analysis, 2nd Edition. Cambridge University Press.
R. W. Ogden (1997). Non-Linear Elastic Deformations. Dover Publications, Mineola, New York.
S. P. Timoshenko, J. N. Goodier (1970). Theory of Elasticity, 3rd Edition. McGraw-Hill, New York.
M. Mooney (1940). A Theory of Large Elastic Deformation. Journal of Applied Physics 11(9), 582–592. doi:10.1063/1.1712836
R. S. Rivlin (1948). Large elastic deformations of isotropic materials. I. Fundamental concepts. Philosophical Transactions of the Royal Society of London A 240(822), 459–490. doi:10.1098/rsta.1948.0002
R. S. Rivlin (1948). Large elastic deformations of isotropic materials. IV. Further developments of the general theory. Philosophical Transactions of the Royal Society of London A 241(835), 379–397. doi:10.1098/rsta.1948.0024
D. V. Georgiyevskii, B. Ye. Pobedrya (2004). The number of independent compatibility equations in the mechanics of deformable solids. Journal of Applied Mathematics and Mechanics 68(6), 941–946. doi:10.1016/j.jappmathmech.2004.11.015
Tessler, Alexander; Spangler, Jan L. (2005). A least-squares variational method for full-field reconstruction of elastic deformations in shear-deformable plates and shells. Computer Methods in Applied Mechanics and Engineering 194, 327–339. doi:10.1016/j.cma.2004.03.015
Hüsnü Dal, Kemal Açıkgöz, Yashar Badienia (2021). On the Performance of Isotropic Hyperelastic Constitutive Models for Rubber-Like Materials: A State of the Art Review. Applied Mechanics Reviews 73(2), 020802. doi:10.1115/1.4050978
Alexander Ricker, Peter Wriggers (2023). Systematic Fitting and Comparison of Hyperelastic Continuum Models for Elastomers. Archives of Computational Methods in Engineering 30, 2257–2288. doi:10.1007/s11831-022-09865-x
Kevin Linka, Ellen Kuhl (2023). A new family of Constitutive Artificial Neural Networks towards automated model discovery. Computer Methods in Applied Mechanics and Engineering 403, 115731. doi:10.1016/j.cma.2022.115731
Vahidullah Taç, Kevin Linka, Francisco Sahli-Costabal, Ellen Kuhl, Adrian Buganza Tepole (2024). Benchmarking physics-informed frameworks for data-driven hyperelasticity. Computational Mechanics 73, 49–65. doi:10.1007/s00466-023-02355-2
Zisheng Liao, Mokarram Hossain, Xiaohu Yao (2020). Ecoflex polymer of different Shore hardnesses: Experimental investigations and constitutive modelling. Mechanics of Materials 144, 103366. doi:10.1016/j.mechmat.2020.103366
Siddhesh S. Kulkarni, Nahom M. Bayre, Kamran Khan (2025). Modelling visco-hyperelastic response of Silicone based elastomers for soft robotics and foldable structure applications. International Journal of Engineering Science 211, 104253. doi:10.1016/j.ijengsci.2025.104253

Where do the recent references fit? The 2020–2025 activity relevant to this primer is not new strain or stress measures — that layer is settled — but a professionalization of the constitutive layer on top of it: rigorous multi-mode fitting protocols (Dal et al. 2021; Ricker & Wriggers 2023), machine-learned constitutive models rebuilt on the classical invariant machinery, so that the invariants of this module — $I_1$, $I_2$, and their isochoric counterparts — are what those networks take as inputs (Linka & Kuhl 2023; Taç et al. 2024), and soft-silicone characterization moving beyond pure hyperelasticity into rate-dependent regimes (Liao et al. 2020; Kulkarni et al. 2025).

2. The Forward Finite Element Method →