5. iFEM Elements & Formulations

Inverse Timoshenko beam, iMIN3, iQS4 — section strains and B-matrices, strain rosettes, the weighted least-squares element functional, sparse assembly, and the RZT extension for composites and sandwich

Contents
1. From Functional to Elements 2. Mindlin Kinematics & the Eight Section Strains 3. Strain Rosettes: Membrane + Bending from Surface Data 4. Interpolation & the B-Matrices 5. The Element Functional, kᵉ and fᵉ 6. The Inverse Timoshenko Beam 7. iMIN3: The Three-Node Inverse Mindlin Plate 8. iQS4: The Inverse Shell with Drilling DOF 9. Assembly, Boundary Conditions & the Real-Time Solve 10. RZT-iFEM: Zigzag Kinematics for Composites 11. The Element Zoo Today Interactive: Rosette → Section Strains Interactive: Inverse-Element Test Bench Interactive: Walkthrough — Anatomy of an Inverse Element Flashcards

1. From Functional to Elements: What an Inverse Element Is

Module 4 ended with a variational statement: the iFEM principle (Tessler & Spangler, NASA/TM-2003-212445; CMAME 2005) asks for the displacement field that minimizes a weighted least-squares mismatch between analytic section strains — computed from assumed first-order shear-deformation (FSDT/Mindlin) kinematics — and their experimentally derived counterparts. A statement about fields is not yet an algorithm. An inverse element is what makes it computable: it reuses the entire machinery of a standard $C^0$ displacement finite element — nodes, DOFs, shape functions, strain-displacement $\mathbf{B}$ matrices — but replaces the potential-energy objective with the element least-squares functional.

Core Problem
Given section-strain data at discrete stations, which discrete displacement space — which element, which interpolation, which constraint scheme — turns the least-squares principle into a well-posed linear solve whose reconstruction you can trust between the sensors? This module is about the elements: what they assume, how they are built, and how they fail.

Each element contributes a quadratic $\Phi_e$, and the gradient of that contribution defines the element arrays $\mathbf{k}^e, \mathbf{f}^e$ exactly as in direct FEM. Write $\mathbf{u}^e = \mathbf{S}_e\mathbf{U}$ for the gather of the element's DOFs from the global vector:

$$\Phi(\mathbf{U}) = \sum_{e} \Phi_e(\mathbf{S}_e\mathbf{U}), \qquad \frac{\partial \Phi_e}{\partial \mathbf{u}^e} = \mathbf{k}^e \mathbf{u}^e - \mathbf{f}^e, \qquad \frac{\partial \Phi}{\partial \mathbf{U}} = \sum_e \mathbf{S}_e^{\mathsf T}\bigl(\mathbf{k}^e \mathbf{S}_e \mathbf{U} - \mathbf{f}^e\bigr) = \mathbf{0}$$

— which is $\mathbf{K}\mathbf{U} = \mathbf{F}$ (§9). Read the middle expression as a definition of $\mathbf{k}^e$ and $\mathbf{f}^e$, not as an equation that holds element by element. Shared nodal DOFs are not independent across elements, so $\mathbf{k}^e\mathbf{u}^e - \mathbf{f}^e$ is the element's gradient contribution and does not vanish on its own at the solution; what vanishes is the assembled sum. The sources write the element form (Kefal et al. 2016, eq. 10a) in the same compact way, and it is worth knowing what it does and does not assert.

Three structural consequences follow from what $\Phi_e$ contains — and they define iFEM's character:

  1. No elastic moduli in $\mathbf{k}^e$ — in plain FSDT iFEM. The element matrix is built from sums of $w\,\mathbf{B}^{\mathsf T}\mathbf{B}$ terms only. Displacement reconstruction needs geometry, sensor positions and orientations, and displacement boundary conditions — $E$ and $\nu$ return only if you post-process reconstructed strains into stresses. Read the scope qualifier as load-bearing: the RZT family (§10) builds its zigzag functions $\phi^{(k)}(z)$ from the layerwise shear moduli, and those sit inside $\mathbf{B}$; the i3-RZT element also adds one non-$\mathbf{B}^{\mathsf T}\mathbf{B}$ diagonal block to keep its artificial DOFs invertible.
  2. The right-hand side $\mathbf{f}^e$ is assembled from measured strains, not loads. No knowledge of the loading is required, at any point. This is the one consequence that holds for every member of the family, RZT included.
  3. The global matrix $\mathbf{K}$ depends only on mesh + sensor layout + weights (plus, for RZT, the fixed layer moduli). It is factorized once, offline; each new measurement frame rebuilds the right-hand side and runs two triangular solves against the stored factors. This is what makes iFEM real-time capable (§9).

The Element Zoo, Previewed

Every member of the inverse-element family is a direct-FEM element with its objective swapped. All of them appear in detail below:

ElementNodes × DOFKinematicsNicheSource
Inverse Timoshenko beam2 × 6 = 12Timoshenko3D frames, slender membersGherlone et al. 2012
iMIN33 × 5 = 15FSDT (Mindlin)Flat platesTessler & Spangler 2005
iQS44 × 6 = 24FSDT + drilling DOFGeneral built-up shells — the workhorseKefal et al. 2016
iCS88-node, curvedSolid-shell / FSDT, curvilinearCylindrical marine geometryKefal 2019
i3-RZT3 × 9 = 27RZT: 7 physical fields + drilling $\theta_z$ + artificial zigzag drilling $\psi_z$Laminates, sandwichKefal, Tessler & Oterkus 2017

Two Independent Accuracy Axes

Reconstruction accuracy has two separate knobs, and conflating them causes endless confusion. Mesh resolution controls how well the $C^0$ interpolation can represent the deformation at all. Sensor density controls how much of that mesh is fed data. You can have a fine mesh starved of sensors (interpolation bridges the gaps — smoothly, but blind), or a coarse mesh drowning in sensors (data everywhere, but the displacement space cannot bend where the structure does).

Key insight
Between sensors, the element's assumed displacement field is the only physics filling the gap. That is why interpolation quality — the subject of this module — matters just as much as sensor count. A bad element turns good data into a bad shape.

2. Mindlin Kinematics and the Eight Section Strains

Module 4 used FSDT informally — "sections stay straight but may shear". Here is the formal version, because the element formulations live or die on its details. Take a plate of thickness $2h$: midplane at $z = 0$, surfaces at $z = +h$ and $z = -h$. FSDT (Mindlin) kinematics: in-plane displacements vary linearly through the thickness via the rotations of the normal; the transverse deflection is constant through the thickness:

$$u_x(x,y,z) = u + z\,\theta_x, \qquad u_y(x,y,z) = v + z\,\theta_y, \qquad u_z(x,y,z) = w$$
Convention — fix it once, before any algebra

Two conventions are load-bearing and silently corrupt results when mixed:

  • Thickness. Here surfaces sit at $z = \pm h$, so the plate thickness is $t = 2h$. Substituting the full thickness where $h$ (the half-thickness) belongs silently halves every curvature and corrupts the $(2h)^2$ factor of §5. Always check which convention a paper uses before copying a formula.
  • Rotations. We use the plane-labelled, displacement-generating convention declared in Module 2: $u_x = u + z\theta_x$, so $\theta_x$ is the rotation that moves material at $z>0$ in $+x$. This is not the lettering of the iFEM sources, which name rotations after the axis they turn about: Tessler & Spangler (NASA/TM-2003-212445, §2.1) write $u_x = u + z\theta_y$, $u_y = v + z\theta_x$, and Kefal et al. 2016 (eqs. 4a–4b) write $u_x = u + z\theta_y$, $u_y = v - z\theta_x$. Their $\theta_y$ is our $\theta_x$ in both cases; the other rotation matches only up to a sign, and the sign differs between the two sources — Tessler & Spangler's $\theta_x$ is our $+\theta_y$, Kefal et al.'s is our $-\theta_y$. Mixing the two swaps subscripts and flips curvature signs — and a sign-flipped $\boldsymbol{\kappa}^{\varepsilon}$ feeds the solver data from a mirror-image structure. Copying $\mathbf{B}$-matrix rows between papers without reconciling conventions is a classic silent failure. §6 deliberately switches to the axis-rotation convention for the 3D beam, and says so on the spot.

Differentiate the kinematics and every strain in the plate is generated by eight section strains (strain measures), functions of $(x, y)$ only — three membrane strains, three bending curvatures, two transverse shears:

$$\mathbf{e}=\begin{bmatrix}u_{,x}\\ v_{,y}\\ u_{,y}+v_{,x}\end{bmatrix},\qquad \boldsymbol{\kappa}=\begin{bmatrix}\theta_{x,x}\\ \theta_{y,y}\\ \theta_{x,y}+\theta_{y,x}\end{bmatrix},\qquad \mathbf{g}=\begin{bmatrix}w_{,x}+\theta_x\\ w_{,y}+\theta_y\end{bmatrix}$$

The in-plane strain at any height is the affine profile

$$\boldsymbol{\varepsilon}_{\parallel}(x,y,z) = \mathbf{e}(x,y) + z\,\boldsymbol{\kappa}(x,y), \qquad z \in [-h, +h]$$

— a constant membrane part plus a linear bending part. The transverse shear strain is constant through the thickness in FSDT; the real distribution is roughly parabolic, and FSDT carries the average. In direct FSDT FEM this is patched with a shear correction factor ($5/6$) inside the constitutive matrix. iFEM has no constitutive matrix in the reconstruction — so no shear correction factor appears in it. (The claim is about the reconstruction, not about everything on this page: the direct ancestor element iMIN3 borrows its interpolation from does use one, §7.)

Key insight — the meeting point
The section strains are simultaneously (a) computable from any candidate displacement field, by differentiation, and (b) — for six of the eight — derivable from strain measurements on the surfaces, by the rosette algebra of §3. The whole method lives in that overlap. Be exact about the arithmetic, though: back-to-back rosettes yield the three membrane strains and the three curvatures, six numbers. The two transverse shears $\gamma_{xz}, \gamma_{yz}$ are not measured by surface rosettes at all; they stay analytic quantities carried by the model, pinned in the functional by a small-weight zero target — a prior, not a measurement. So the honest statement is six measured, eight modeled: the model speaks eight numbers and the data answers on six of them.

Count check. Plate/shell element: $3 + 3 + 2 = 8$ measures. The beam analogue (§6) has 6. RZT (§10) enlarges the set with zigzag measures. The pattern — a small set of section strains generating the full 3D strain field — is what every member of the family shares.

3. Strain Rosettes: Membrane + Bending from Surface Data

Now the data side. Module 4 pulled the half-sum/half-difference trick for a beam; this section is the canonical plate machinery. At station $i$, instrument the top surface ($z = +h$) and bottom surface ($z = -h$) with three-gauge strain rosettes, giving the full in-plane strain vector on each surface: $\boldsymbol{\varepsilon}^{+}_i = \{\varepsilon^{+}_{xx}, \varepsilon^{+}_{yy}, \gamma^{+}_{xy}\}$ and $\boldsymbol{\varepsilon}^{-}_i$. For a rectangular 0/45/90 rosette the conversion from gauge readings to strain components is:

$$\varepsilon_{xx}=\varepsilon_{0},\qquad \varepsilon_{yy}=\varepsilon_{90},\qquad \gamma_{xy}=2\varepsilon_{45}-\varepsilon_{0}-\varepsilon_{90}$$
Trap — the factor 2
Forgetting the factor 2 in $\gamma_{xy}=2\varepsilon_{45}-\varepsilon_{0}-\varepsilon_{90}$ is a classic error. Writing $\varepsilon_{45}-\tfrac{1}{2}(\varepsilon_{0}+\varepsilon_{90})$ gives the tensor shear $\varepsilon_{xy}$ — exactly half the engineering shear $\gamma_{xy}$ that the FE strain vector uses. The mistake halves one component of your data and is invisible until validation.

Because the through-thickness profile is affine (§2), the half-sum and half-difference of the two surface readings separate it exactly:

$$\mathbf{e}^{\varepsilon}_i = \tfrac{1}{2}\left(\boldsymbol{\varepsilon}^{+}_i + \boldsymbol{\varepsilon}^{-}_i\right), \qquad \boldsymbol{\kappa}^{\varepsilon}_i = \tfrac{1}{2h}\left(\boldsymbol{\varepsilon}^{+}_i - \boldsymbol{\varepsilon}^{-}_i\right)$$

This is the experimental membrane+bending decomposition — the data-side twin of the kinematic decomposition $\boldsymbol{\varepsilon}_{\parallel} = \mathbf{e} + z\boldsymbol{\kappa}$. The superscript $\varepsilon$ marks quantities derived from measurement, distinguishing them from their analytic counterparts $\mathbf{e}(\mathbf{u})$, $\boldsymbol{\kappa}(\mathbf{u})$.

z = 0 z = +h z = −h z rosette ε⁺ rosette ε⁻ ε⁺ ε⁻ ε(z), affine = membrane e (ε⁺+ε⁻)/2 + bending z·κ κ = (ε⁺−ε⁻)/(2h) transverse shear γxz: constant in z — produces no in-plane surface strain, invisible to surface rosettes
The affine through-thickness profile $\varepsilon(z) = e + z\kappa$ and its exact separation from two surface rosettes: half-sum isolates the membrane block, difference over $2h$ isolates the bending wedge. The transverse shear leaves no trace on either surface.

Noise Propagation: Why Thin Structures Are Hard

The half-sum averages noise down; the difference, divided by $2h$, amplifies it. Be exact about what carries the noise: the formulas below are for one already-resolved surface-strain component, with independent noise $\sigma_{\varepsilon}$ on that component at each of the two surfaces — not for the raw gauge readings a rosette delivers.

$$\sigma_{e} = \frac{\sigma_{\varepsilon}}{\sqrt{2}}, \qquad \sigma_{\kappa} = \frac{\sqrt{2}\,\sigma_{\varepsilon}}{2h}$$

Worked numbers. With $\sigma_{\varepsilon} = 5\,\mu\varepsilon$ on a resolved component: at $2h = 2$ mm, $\sigma_{\kappa} \approx 3.5\times10^{-3}\ \text{m}^{-1}$; at $2h = 0.5$ mm it is $14\times10^{-3}\ \text{m}^{-1}$. Curvature noise scales as $1/\text{thickness}$ while membrane noise is thickness-independent. Thin structures amplify bending noise — and this, not sensor cost alone, drives sensor-budget decisions on thin skins.

Trap — the shear component is noisier than the normal ones
Propagate raw gauge noise instead, $\sigma_g$ independent per gauge on a 0/45/90 rosette, and the components stop being equal. $\varepsilon_{xx}=\varepsilon_0$ and $\varepsilon_{yy}=\varepsilon_{90}$ inherit $\sigma_g$ unchanged, but $\gamma_{xy}=2\varepsilon_{45}-\varepsilon_{0}-\varepsilon_{90}$ has $\operatorname{Var}=(2^2+1+1)\sigma_g^2 = 6\sigma_g^2$. Carrying that through the half-sum/half-difference gives $\sigma_{e_\gamma}=\sqrt{3}\,\sigma_g$ and $\sigma_{\kappa_\gamma}=2\sqrt{3}\,\sigma_g/(2h)$ — a factor $\sqrt{6}$ worse than the normal components. At $\sigma_g = 5\,\mu\varepsilon$ and $2h = 2$ mm that is $8.7\times10^{-3}\ \text{m}^{-1}$ on the twist curvature, against $3.5\times10^{-3}$ on $\kappa_{xx}$. The rosette components are also correlated ($\operatorname{Cov}(\gamma_{xy},\varepsilon_{xx})=-\sigma_g^2$), which is why a scalar $\sigma_{\varepsilon}$ is a simplification, not the full story. The widgets on this page use the resolved-component convention; a real error budget uses the full covariance.

What Surface Rosettes Cannot Give

The transverse shear section strains $\mathbf{g}$ produce no in-plane surface strain signature: in FSDT they are constant through the thickness and built from $w_{,x}+\theta_x$, not from surface stretching. No arrangement of surface rosettes measures them. iFEM's standard treatment (§5): drop the data term for $\mathbf{g}$ and give the analytic shear a small regularization weight instead.

Be precise about the word "directly", though, because the primary sources are. Kefal et al. 2016 states that the surface strains "cannot be directly used to calculate the in-situ transverse shear strains", and then describes the indirect route: Smoothing Element Analysis (SEA) smooths the measured curvatures $\boldsymbol{\kappa}^{\varepsilon}$ so that their first derivatives can be computed accurately, and those derivatives are then used to obtain the transverse-shear measures. The same paper adds that for thin shells the $\mathbf{g}$ contributions are small enough to omit safely — which is why the small-weight prior is the default in practice, not the only option. So: $\mathbf{g}$ is never a measurement; it can be an inference. A source presenting shear section strains as read straight off the rosettes has it wrong; a source recovering them from smoothed curvature fields is doing something else, and saying so.

Practical variants. Distributed fiber-optic sensing (FBG arrays) measures axial strain along the fiber path only — rosette-equivalent data needs three fiber directions per station. Single-surface access (sealed tanks, bonded skins) leaves $\mathbf{e}$ and $\boldsymbol{\kappa}$ entangled: the standard escapes are assuming a membrane- or bending-dominated response, or the strain pre-extrapolation strategies of the recent literature. Instrumentation-side detail — gauge-length averaging, adhesive strain transfer, FBG line integration — lives in Module 6.

Interactive Tool — Rosette → Section Strains

Set the two surface strains and the plate thickness; the profile splits into membrane + bending live. Then add noise and thin the plate: watch $\sigma_{\kappa}$ blow up as $1/(2h)$ while $\sigma_{e}$ stays put. Everything here is one resolved surface-strain component — $\sigma_{\varepsilon}$ is that component's noise, not a raw gauge's (the twist component is $\sqrt{6}$ noisier; see the trap box above).

4. Interpolation and the Strain-Displacement Matrices B

The discrete link between nodal DOFs and section strains. Within element $e$, the kinematic variables are interpolated from the nodal DOF vector $\mathbf{u}^e$ with $C^0$ shape functions. Because the section strains are first derivatives (or values) of those variables, each is a linear map of $\mathbf{u}^e$:

$$\mathbf{e}(\mathbf{u}^{e})=\mathbf{B}^{m}\mathbf{u}^{e},\qquad \boldsymbol{\kappa}(\mathbf{u}^{e})=\mathbf{B}^{b}\mathbf{u}^{e},\qquad \mathbf{g}(\mathbf{u}^{e})=\mathbf{B}^{s}\mathbf{u}^{e}$$

The $\mathbf{B}$ matrices are exactly the same objects as in direct FEM — rows of shape-function derivatives — and for the classic inverse elements they are inherited unchanged from the direct ancestor element. An inverse element is "$\mathbf{B}$-matrix reuse with a different objective". Everything you know about shape-function quality from Module 2 transfers wholesale.

Anatomy, with iMIN3 as the Concrete Case

iMIN3 carries 15 DOF (three nodes × $(u, v, w, \theta_x, \theta_y)$, §7). Its three matrices:

For iQS4 the same three matrices are $3\times24$, $3\times24$, $2\times24$, with additional drilling-rotation columns entering $\mathbf{B}^m$ through the Allman-type membrane interpolation (§8):

$$\mathbf{q}_{i}=\left[\,u_{i}\;\;v_{i}\;\;w_{i}\;\;\theta_{xi}\;\;\theta_{yi}\;\;\theta_{zi}\,\right]^{\mathsf T},\qquad \mathbf{u}^{e}_{\mathrm{iQS4}}=\left[\,\mathbf{q}_{1}^{\mathsf T}\;\;\mathbf{q}_{2}^{\mathsf T}\;\;\mathbf{q}_{3}^{\mathsf T}\;\;\mathbf{q}_{4}^{\mathsf T}\,\right]^{\mathsf T}\quad (4\times 6 = 24\ \text{DOF})$$

Where the B-Matrices Are Evaluated

The element functional samples the $\mathbf{B}$ matrices at the measurement stations: $\mathbf{B}_i = \mathbf{B}(x_i, y_i)$. Nothing in the formulation requires stations to sit at Gauss points or centroids — the centroid is merely the common convention when one rosette pair per element is used ($n=1$). Sensors go where the structure allows; the discrete functional (§5) accepts arbitrary in-element stations. Claiming centroids are required confuses a convention with a constraint.

Dimensional bookkeeping
Entries of $\mathbf{e}$ and $\mathbf{g}$ are dimensionless strains; entries of $\boldsymbol{\kappa}$ have units of 1/length. Three residuals with different units cannot be summed as they stand — this is exactly what forces the $(2h)^2$ factor in the next section. Keep the units in view and the factor stops looking like a fudge.

5. The Element Functional, kᵉ and fᵉ

The weighted least-squares element functional (Tessler & Spangler 2003/2005; Kefal et al. 2016 eq. 8) compares the three analytic section-strain vectors against their measured counterparts — where "measured" is exact for $\mathbf{e}$ and $\boldsymbol{\kappa}$ and a stand-in zero for $\boldsymbol{\gamma}$, whose target no surface rosette supplies:

$$\Phi_{e}(\mathbf{u}^{e})=w_{m}\bigl\lVert \mathbf{e}(\mathbf{u}^{e})-\mathbf{e}^{\varepsilon}\bigr\rVert^{2}+w_{b}\bigl\lVert \boldsymbol{\kappa}(\mathbf{u}^{e})-\boldsymbol{\kappa}^{\varepsilon}\bigr\rVert^{2}+w_{s}\bigl\lVert \mathbf{g}(\mathbf{u}^{e})-\mathbf{g}^{\varepsilon}\bigr\rVert^{2}$$

The norms are area-integrated squared mismatches. With $n$ discrete measurement stations per element, the practical discrete form multiplies the station-summed squared residual by $A_e/n$ — for the curvature term:

$$\bigl\lVert \boldsymbol{\kappa}(\mathbf{u}^{e})-\boldsymbol{\kappa}^{\varepsilon}\bigr\rVert^{2}=\frac{(2h)^{2}A_{e}}{n}\sum_{i=1}^{n}\bigl[\mathbf{B}^{b}_{i}\mathbf{u}^{e}-\boldsymbol{\kappa}^{\varepsilon}_{i}\bigr]^{\mathsf T}\bigl[\mathbf{B}^{b}_{i}\mathbf{u}^{e}-\boldsymbol{\kappa}^{\varepsilon}_{i}\bigr]$$
Trap — the $(2h)^2$ factor is part of the definition
The curvature norm carries $(2h)^{2}$ so that all three terms have identical units of $(\text{strain})^2 \times (\text{area})$. It is not an optional scaling. Omit it and the weights $w_m, w_b, w_s$ stop being comparable: bending is mis-weighted relative to membrane by a factor of order $1/(2h)^2$ — for a 2 mm plate, $1/(0.002\,\text{m})^2 = 2.5\times10^{5}$, better than five decades. Implementations that "simplify" the functional by dropping it produce reconstructions dominated by whichever term happens to have the biggest numbers.

The Weight Logic

Where sensor data exists, the corresponding weight is $1$ (or a component-wise diagonal weight matrix in refined implementations). For strain measures with no experimental counterpart, the data term is dropped and a small positive weight — $\lambda$ typically in the range $10^{-3}$ to $10^{-5}$, with $10^{-4}$ the common default — multiplies the squared analytic strain alone. Two such cases are standard:

Mind $n$ in the second case. $n$ counts measurement stations, so a sensorless element has $n = 0$ and the $A_e/n$ station average of the displays above is undefined, not merely small. The source resolves this by dropping the data term and keeping the plain area integral (Kefal et al. 2016, eq. 9): each squared norm becomes $\int_{A_e}\lVert\boldsymbol{\alpha}(\mathbf{u}^e)\rVert^{2}dA$ at weight $\lambda$, so the prior contributes $\lambda\int_{A_e}\mathbf{B}^{\mathsf T}\mathbf{B}\,dA$ to $\mathbf{k}^e$ and nothing at all to $\mathbf{f}^e$. An implementation that would rather collocate needs a quadrature count $n_q \ge 1$ of its own choosing; $n_q$ is not $n$.

The small weights act as Tikhonov-style regularization (Module 3): they restore rank that the missing data terms would otherwise cost, and let the displacement interpolation bridge uninstrumented regions smoothly, while barely biasing instrumented ones. Setting them to zero is a bug wherever those rows are load-bearing — the assembled $\mathbf{K}$ can lose rank, and does exactly when dropping them destroys the full column rank of the remaining weighted-$\mathbf{B}$ stack on the free DOFs (§9 states the condition properly). The test bench below is a case where it does: zeroing $w_s$ there costs six directions. Note the limit of what a positive weight can do, though: it only reaches modes that appear in some weighted strain row. A displacement mode that every $\mathbf{B}$ block annihilates — iQS4's constant drilling mode, §8 — stays in the null space for any $\lambda$ whatsoever, and has to be removed by the assembly or the boundary conditions instead. Setting them large is the opposite bug — and it is worth being exact about what the prior actually does: $\lambda\lVert\mathbf{g}(\mathbf{u}^e)\rVert^2$ pushes the analytic transverse shear toward zero, which is the Kirchhoff constraint imposed as if it were data — not a push toward zero displacement. Pure bending of a thin plate has $\mathbf{g} = 0$ identically while deflecting freely. In a locking-free space a large $\lambda$ therefore bites only in the thick regime, where it erases the genuine transverse-shear deflection — and since that deflection vanishes as the structure thins, the thin limit is precisely where the prior is harmless. In such a space the damage barely depends on how large $\lambda$ is, because the deflection DOFs enter no data term for the prior to compete against (the bench's tip-force case at low slenderness, unchanged across the whole $w_s$ slider). Where a large $\lambda$ does destroy the thin limit is a locking-prone equal-order space: there it collapses the reconstruction outright (§7), because that space cannot bend without parasitic shear. Where a whole element is sensorless, all three measures carry $\lambda$ — and there the bias really is toward no deformation. Deployment-side strategies for missing measurements — pre-extrapolation, virtual strains, weight tuning — are Module 6's territory.

Stationarity: The Normal Equations

$\Phi_e$ is quadratic in $\mathbf{u}^e$, so its gradient is affine: $\partial\Phi_e/\partial\mathbf{u}^e = \mathbf{k}^{e}\mathbf{u}^{e}-\mathbf{f}^{e}$, which is what defines the element arrays (§1 — the element residual is a contribution to the assembled normal equations, not an equation that holds on its own):

$$\mathbf{k}^{e}=\frac{A_{e}}{n}\sum_{i=1}^{n}\Bigl(w_{m}\mathbf{B}^{m\mathsf T}_{i}\mathbf{B}^{m}_{i}+w_{b}(2h)^{2}\mathbf{B}^{b\mathsf T}_{i}\mathbf{B}^{b}_{i}+w_{s}\mathbf{B}^{s\mathsf T}_{i}\mathbf{B}^{s}_{i}\Bigr)$$
$$\mathbf{f}^{e}=\frac{A_{e}}{n}\sum_{i=1}^{n}\Bigl(w_{m}\mathbf{B}^{m\mathsf T}_{i}\mathbf{e}^{\varepsilon}_{i}+w_{b}(2h)^{2}\mathbf{B}^{b\mathsf T}_{i}\boldsymbol{\kappa}^{\varepsilon}_{i}+w_{s}\mathbf{B}^{s\mathsf T}_{i}\mathbf{g}^{\varepsilon}_{i}\Bigr)$$
Small print — this is the collocation form, not the source's integral form
Both displays above sample $\mathbf{B}$ at the $n$ stations and multiply by $A_e/n$. Kefal et al. 2016 eq. (10c) instead keeps the area integral, $\mathbf{k}^{e}=\int_{A_e}\bigl(w_m\mathbf{B}^{m\mathsf T}\mathbf{B}^{m}+w_b(2h)^2\mathbf{B}^{b\mathsf T}\mathbf{B}^{b}+w_s\mathbf{B}^{s\mathsf T}\mathbf{B}^{s}\bigr)\,dA$, with the station average appearing only in $\mathbf{f}^e$ (eq. 10d). Constant $\mathbf{B}$ guarantees the two agree — true of iMIN3's CST-like $\mathbf{B}^m$ and $\mathbf{B}^b$, false of iQS4's Allman-enriched $\mathbf{B}^m$ and of any anisoparametric $\mathbf{B}^s$. But constancy is sufficient, not necessary: a varying $\mathbf{B}$ still agrees exactly whenever the station rule happens to integrate $\mathbf{B}^{\mathsf T}\mathbf{B}$ exactly. The bench below is its own counterexample — it samples at the two Gauss points of the element, which integrates the product of its $\xi$-linear $\mathbf{B}^s$ to machine precision. So exactness depends on station locations, quadrature weights, and the data treatment together, not on $\mathbf{B}$ alone. Station collocation is an approximation you are choosing, so choose it knowingly.

$\mathbf{k}^e$ is symmetric positive semidefinite; uniqueness comes only after global assembly and boundary conditions (§9). Note the asymmetry of roles: $\mathbf{k}^e$ is data-independent — geometry, stations, weights — while $\mathbf{f}^e$ is where the measurements live. The entire time-dependence of a monitoring problem flows through $\mathbf{f}^e$.

Why it matters
This split is the real-time argument in embryo: everything expensive ($\mathbf{k}^e$, assembly, factorization) is data-free and happens once; everything data-dependent ($\mathbf{f}^e$) is a cheap weighted sum of the incoming strains. §9 finishes the argument.

6. The Inverse Timoshenko Beam Element

The 1D member of the family, developed by Gherlone, Cerracchio, Mattone, Di Sciuva & Tessler for 3D frame structures (IJSS 2012) and validated experimentally (Smart Materials and Structures 2014). This is the complete beam derivation, run in the same four moves as the plate: kinematics → section strains → interpolation → functional.

Kinematics. For a beam along $x$ with cross-section coordinates $(y, z)$, Timoshenko kinematics read:

$$u_x = u + z\,\theta_y - y\,\theta_z, \qquad u_y = v - z\,\theta_x, \qquad u_z = w + y\,\theta_x$$

— six kinematic variables per section (three translations $u, v, w$ and three rotations $\theta_x, \theta_y, \theta_z$), all functions of $x$ only. Cross-sections translate, rotate, and may shear relative to the axis, but do not deform in their own plane.

Convention switch — the §2 trap, live on this page
Note which rotation multiplies $z$: here it is $\theta_y$, not the $\theta_x$ of §2. The 3D beam is written in the axis-rotation convention (rotations named after the axis they turn about), because that is what the frame literature and the beam's three-rotation set make natural; §2–§5 use the plane-labelled convention. Both are self-consistent, and the two are related by $\theta_x^{\S 2} = \theta_y^{\S 6}$ together with $\theta_y^{\S 2} = -\theta_x^{\S 6}$ — a swap and a sign flip on the second rotation (match the two displayed kinematics at $y = 0$ and read it off). Carry a $\mathbf{B}$-matrix row across the boundary without both halves of that rule and the curvature lands in the wrong slot with the wrong sign — exactly the failure §2 warns about, happening inside one document.

Section strains. Differentiating gives six section strains: axial stretch, two bending curvatures, torsion rate, two transverse shears:

$$\mathbf{E}(x) = \begin{bmatrix} u' & \theta_y' & \theta_z' & \theta_x' & v'-\theta_z & w'+\theta_y \end{bmatrix}^{\mathsf T}$$

Sign conventions for the shear entries vary between texts; the physical content — 1 axial + 2 bending + 1 torsion + 2 shear = 6 — does not. (The test bench below uses the planar convention $\gamma = w' - \theta$, with $\theta$ measured in the same sense as $w'$. Note what does not distinguish the two: pure bending gives $\gamma = 0$ in either, because zero transverse shear in the Euler–Bernoulli limit is a property of Timoshenko kinematics, not of a sign choice. Only the sign of $\theta$ differs.)

Interpolation. The inverse element is two-node with 6 DOF per node (12 DOF), and uses locking-free interdependent (anisoparametric) interpolation. The orders are not uniform across the six variables, and the sloppy summary "deflections cubic, rotations quadratic" gets two of them wrong: the axial displacement $u$ and the torsional rotation $\theta_x$ are linear, the bending rotations $\theta_y, \theta_z$ are quadratic, and the transverse deflections $v, w$ are cubic — tied together so that each element represents a state of constant transverse shear exactly, the beam analogue of the constraint schemes of iMIN3/iQS4. (The pattern follows the shear definitions $e_4 = w_{,x}+\theta_y$, $e_5 = v_{,x}-\theta_z$: only the variables that meet inside a shear measure need matching orders, which is why $u$ and $\theta_x$ stay linear. Gherlone et al., NASA/TP-2011-217315, §3.1.) Equal-order interpolation would lock, least squares or not (§7).

Functional. The element functional is the weighted least-squares sum over the six measures sampled at stations along the axis. Here the beam parts company with the plate, and in the opposite direction from the obvious guess. The published formulation treats all six measures as data-carrying: its weight vector is $\{w_k^{0}, \ldots\}$ with uniform $w_k^{0} = 1$ $(k = 1,\ldots,6)$ in the numerical examples, transverse shears included. Eight axial gauge readings per element determine the six measures of the constant-shear element directly; six suffice if $e_4, e_5$ are instead solved from the Timoshenko equilibrium and constitutive relations rather than measured — which the authors describe as a way of reducing the gauge count, not as a zero prior (Gherlone et al., NASA/TP-2011-217315, §4.2, eqs. 30). Dropping beam shear to a small weight is a later deployment choice: Esposito et al. 2023, for instance, declare the beam's shear and torsion measures unknown and set those weights to $10^{-4}$. Behind all of it is a kinematic difference worth keeping: plate transverse shear leaves no surface signature at all (§3), whereas the beam shears do appear in the ±45° wall-shear readings — see the algebra below. On a beam they go undata'd by gauge budget and convention, not by kinematics.

The Data Side: Gauges Around the Cross-Section

Section strains come from axial strain gauges arranged around the cross-section at each station. For a thin-walled circular tube of radius $R$:

Gauge arrangementExtracted measureHow
Axial gauges at 0°/90°/180°/270°axial stretch $u'$mean of the four readings
Same four gaugesboth bending curvatures $\theta_y', \theta_z'$half-differences of opposite pairs, divided by $R$ — equivalently, differences divided by $2R$: $\theta_y' = (\varepsilon_{90}-\varepsilon_{270})/2R$, $\theta_z' = -(\varepsilon_{0}-\varepsilon_{180})/2R$. Signs set by gauge ordering
Gauge pairs at ±45°, at diametrically opposite stationstorsion rate $\theta_x'$wall shear strain from each pair, $\gamma = \varepsilon_{+45}-\varepsilon_{-45}$ (the difference cancels both normal strains), then average the two opposite stations: $\theta_x' = \bigl[\gamma(\phi)+\gamma(\phi+\pi)\bigr]/2R$. A single pair is not enough — see below. Signs set by gauge ordering and circumferential sense

Why one ±45° pair does not isolate torsion. Differentiate the §6 kinematics on the tube wall at circumferential angle $\phi$ — measured from $+y$ toward $+z$, the same labelling the row above uses, so $y = R\cos\phi$, $z = R\sin\phi$ — and the in-plane wall shear comes out as

$$\gamma(\phi) \;=\; (w' + \theta_y)\cos\phi \;-\; (v' - \theta_z)\sin\phi \;+\; R\,\theta_x'$$

The difference $\varepsilon_{+45}-\varepsilon_{-45}$ does kill both normal strains exactly — that part of the trick is sound — but it does not kill the two transverse shears, which ride on $\cos\phi$ and $\sin\phi$. So $\theta_x' = \gamma/R$ from a single station holds only under pure torsion, or wherever the transverse shears are assumed zero. Averaging a diametrically opposite pair cancels the $\cos\phi,\sin\phi$ terms identically and leaves $R\theta_x'$; differencing the same two stations throws the torsion away and returns the shear combination instead. That is the concrete sense in which the beam shears are inferable rather than invisible: a distributed gauge set around the section determines all six measures jointly, and the published formulation solves that system rather than reading each measure off one station at a time.

Frames. Assembly of members at arbitrary orientation uses standard rotation transformation matrices per member, identical to direct frame FEM — and just as in direct FEM, forgetting the transformation (or applying a wrong local axis convention) silently corrupts the assembled system. The test bench has a toggle for exactly this failure.

Experimental credibility. The 2014 paper established the approach on a thin-walled cantilevered beam subjected to different static and dynamic loads, comparing reconstructed deflections and twist rotations against independently measured values (Gherlone et al. 2014).

7. iMIN3: The Three-Node Inverse Mindlin Plate Element

Ancestry first, because the inverse element inherits everything that matters. MIN3 (Tessler & Hughes, CMAME 1985) is a three-node Mindlin plate bending element whose defining trick is anisoparametric interpolation — deliberately different interpolation orders for different variables: rotations $\theta_x, \theta_y$ linear, deflection $w$ quadratic.

The quadratic deflection terms are not independent: continuous shear edge constraints force the edge-tangential transverse shear strain to be constant along each element edge. This eliminates the parasitic linear shear terms that cause shear locking in equal-order $C^0$ elements, and condenses the midside deflection parameters out, leaving a 3-corner-node element. The result: the thin-plate (Kirchhoff) limit is captured without locking and without reduced integration or its hourglass pathologies.

Credit where it is due, though: Tessler & Hughes attribute much of MIN3's transverse-shear performance to a second ingredient alongside the edge constraints — an element-level shear correction factor concept, in which the shear rigidity carries an element-appropriate factor rather than the constitutive $5/6$. That is a direct-FEM device: it lives in a stiffness matrix, which the inverse functional does not have. So §2's "no shear correction factor appears in the reconstruction" holds for iFEM, and the ancestor element it borrows its interpolation from does use one. iMIN3 inherits the anisoparametric interpolation and the edge constraints; it does not inherit the rigidity correction, because there is no rigidity to correct.

$$\gamma_{sz}\big|_{\text{edge}} = \text{const} \quad \text{(edge-tangential shear; continuous shear edge constraints, MIN3)}$$

iMIN3 (Tessler & Spangler, NASA/TM-2003-212445; CMAME 2005) is the inverse counterpart, extended with membrane behavior: linear $u, v$ interpolation (CST-like), so the nodal DOF set is $(u, v, w, \theta_x, \theta_y)$ — 5 DOF/node, 15 DOF total. The deflection interpolation has the characteristic anisoparametric form

$$w = \sum_{i=1}^{3} N_i w_i + \sum_{i=1}^{3} \left( L_i\,\theta_{xi} + M_i\,\theta_{yi} \right), \qquad N_i \;\text{linear},\; L_i, M_i \;\text{quadratic}$$

with $L_i, M_i$ quadratic functions built by the edge constraints — nodal rotations contribute directly to the deflection field. That coupling is the whole anti-locking mechanism in one line.

node 1 node 2 node 3 each node: (u, v, w, θx, θy) — 5 DOF, no θz γsz = const along edge (continuous shear edge constraint) rotations θ: linear  ·  deflection w: quadratic, tied to nodal rotations via Lᵢ, Mᵢ thickness → (thin limit at left) normalized deflection MIN3-type anisoparametric: correct Kirchhoff limit equal-order element: locks, deflection collapses as t → 0 1.0 0
iMIN3's anisoparametric interpolation: linear rotations, quadratic deflection tied to nodal rotations by continuous-shear edge constraints (left). Inset: what the constraints buy — a locked equal-order element loses its deflection as the plate thins; the MIN3 scheme keeps the correct thin-plate limit.
Key insight — locking survives the change of objective
Shear locking is a property of the discrete displacement space, not of the objective functional. Least squares does not exorcise it: if the interpolation cannot represent pure bending without parasitic shear, the least-squares fit is distorted in exactly the same way as the direct stiffness solution. That is why every serious inverse element inherits an anti-locking scheme from its direct ancestor — MIN3 heritage in iMIN3/iQS4, interdependent interpolation in the beam. Verify it yourself in the test bench: switch the formulation to "equal-order" and thin the structure.

The limitation that motivates §8: iMIN3 has no drilling rotation $\theta_z$. Assembling non-coplanar shell models — folded plates, stiffened panels, curved surfaces approximated by facets — leaves a singular sixth equation at nodes where all adjacent elements are coplanar, requiring ad hoc fixes. A flat-plate element is not yet a shell element.

8. iQS4: The Four-Node Inverse Shell with Drilling DOF

The workhorse of applied iFEM: iQS4 (Kefal, Oterkus, Tessler & Spangler, Engineering Science and Technology 2016). Four-node quadrilateral inverse shell, six DOF per node — $u, v, w, \theta_x, \theta_y$ plus the hierarchical drilling rotation $\theta_z$ — 24 DOF total. Two interpolation ingredients:

(1) Membrane. Bilinear in-plane displacement enriched with Allman-type quadratic edge modes driven by differences of nodal drilling rotations ($\theta_{zj} - \theta_{zi}$ along each edge):

$$\mathbf{u}_{\text{mem}} = \sum_{i=1}^{4} N_i \begin{bmatrix} u_i \\ v_i \end{bmatrix} + \sum_{i=1}^{4} \mathbf{L}_i\, \theta_{zi} \qquad \text{(Allman-type edge modes)}$$

This both improves the in-plane bending response over the plain bilinear quad and — decisively — equips every node with the full 6-DOF set, so that shell assemblies with folds, stiffener junctions, and arbitrary orientation changes produce a nonsingular global system in suitably constrained, non-degenerate assemblies, without artificial drilling springs. The qualifier is not decoration; see the drilling null-space box below.

(2) Bending. An anisoparametric constrained scheme in the MIN-element tradition: the deflection interpolation is coupled to the bending rotations such that the transverse shear along each edge is constant, keeping the thin limit locking-free — the same medicine as §7, applied to a quadrilateral.

The element functional is exactly the standard three-term least squares of §5 — membrane, bending, transverse shear, and nothing else. It is worth being explicit about this, because the obvious guess is wrong: iQS4 does not add a drilling penalty or an artificial drilling spring. The drilling rotations are not stabilised by a fourth term in the functional at all; they enter through the membrane interpolation itself. Allman-type quadratic edge modes are driven by differences of nodal $\theta_z$, so the drilling DOFs appear directly in $\mathbf{B}^m$ and pick up stiffness from the non-uniform part of the drilling field.

Trap — difference-driven enrichment does not kill the constant drilling mode; it creates it

It is tempting to write "the Allman edge modes remove the zero-energy drilling mode". That is backwards, and the paper's own appendix shows why. Write the two scalar components of $\mathbf{L}_i$ above as $L_i$ (acting on $u$) and $M_i$ (acting on $v$). With $L_1 = y_{14}N_8 - y_{21}N_5$, $L_2 = y_{21}N_5 - y_{32}N_6$, $L_3 = y_{32}N_6 - y_{43}N_7$, $L_4 = y_{43}N_7 - y_{14}N_8$ and the matching $M_i$ in the $x_{ij}$ (Kefal et al. 2016, eqs. A.9–A.16), each sum telescopes:

$$\sum_{i=1}^{4}L_i = 0, \qquad \sum_{i=1}^{4}M_i = 0$$

So set every $\theta_{zi} = c$ with all translations zero: the enrichment contributes nothing, $\mathbf{u}_{\text{mem}} \equiv 0$, and $\mathbf{B}^m\mathbf{u}^e = \mathbf{0}$. Since the $\theta_z$ column of both $\mathbf{B}^b$ and $\mathbf{B}^s$ is identically zero (Kefal et al. 2016, eq. B.4), that vector is in the null space of $\mathbf{k}^e = \sum w\,\mathbf{B}^{\mathsf T}\mathbf{B}$ — for any weights, so the small $\lambda$ of §5 does not touch it either. This is the classical Allman constant-drilling mode.

What iQS4 actually buys is the full 6-DOF node, so folded and stiffened assemblies are kinematically compatible without ad hoc fixes. What removes the constant-drilling mode is the assembly: at a fold, one element's $\theta_z$ is a neighbour's bending rotation and inherits its stiffness — plus the displacement boundary conditions of §9. On a flat, weakly constrained patch the mode survives. Kefal et al. claim exactly this and no more: drilling DOFs mean "singular solutions can be simply avoided when modelling complex shell structures". Complex, i.e. non-coplanar.

1 2 3 4 u v w θx, θy θz (drilling, about shell normal) x y z measurement station (ε⁺, ε⁻) Allman membrane quadratic edge modes from Δθz Anisoparametric bending constant transverse shear along each edge 4 nodes × 6 DOF = 24 DOF full 6-DOF nodes → folds assemble without ad hoc fixes
iQS4 anatomy: four corner nodes each carrying $(u, v, w, \theta_x, \theta_y, \theta_z)$, a centroidal measurement station with top/bottom rosettes (the $n = 1$ convention), Allman-type membrane enrichment driven by drilling-rotation differences, and constrained anisoparametric bending.

Assumptions and Failure Modes — Not a Catalogue of Successes

AssumptionHow it failsRemedy
FSDT kinematics, monolithic sectionsandwich cores and thick laminates warp in zigzag; reconstruction degrades through the thicknessRZT family (§10)
Flat element geometrycurved surfaces need fine faceting → many more elements, each of which must then be instrumented or regularized awaycurved/isogeometric elements (§11)
Drilling DOFs carry no measurement of their own$\theta_z$ enters only through the differences driving the Allman edge modes, so a spatially uniform drilling rotation is strain-free and a flat, weakly constrained patch keeps it in the null spacerely on non-coplanar assembly and drilling boundary conditions; check the drilling columns of the assembled system rather than reaching for an artificial spring
Per-element local axeswrong or inconsistent local frames scramble the assembled system — silently on flat regions, catastrophically at foldstransformation discipline (§9); see the bench's "skip transformation" toggle
Small strains, linear kinematicslarge deflections / elastomeric substrates violate linearityco-rotational and large-deformation variants (§11 pointer)

Validation and role. The 2016 paper reports two validation problems — a cantilevered plate under static transverse force near the free tip, and a short cantilever beam under shear loading. Robustness to noise is demonstrated separately, on a third problem: a thin-walled cylinder, re-analysed with 5% and 10% random noise added to the simulated strain data, for which the authors report only insignificant loss of accuracy (Kefal et al. 2016). The same group applied the element to full-scale marine monitoring problems (containership and tanker studies), establishing iQS4 as the default choice for general built-up thin-walled structures — the role the 4-node isoparametric shell plays in direct FEM. Baseline instrumentation is one top/bottom rosette pair at the element centroid ($n = 1$), though the discrete functional accepts arbitrary in-element stations (§4).

Interactive Tool — Inverse-Element Test Bench

The same test cases — membrane, pure bending, bending + shear, a graded mesh, a folded frame — pushed through a planar 3-DOF-per-node inverse beam, with the plate elements' two rival design choices selectable: equal-order interpolation (what iMIN3/iQS4 refuse to use) vs an interdependent anisoparametric scheme (constant shear per element). Thin the structure, raise the shear weight, zero it, skip the local-axis transformation — and watch the panel, not just the picture. Failures are the curriculum here.

What this bench is and is not. It is a deliberately reduced surrogate, not the published inverse Timoshenko beam. Here each element carries $(u, w, \theta)$ per node with element-constant curvature — so rotations are linear and the interdependent deflection is quadratic, one order below the published element's quadratic bending rotations and cubic deflections (§6), which reach that order via extra interior DOFs that are then condensed out. The bench also drops torsion and the out-of-plane bending pair, and it weights the curvature term by $(2h)^2$ — the plate convention of §5 — where the 3D beam functional uses section ratios $I/A$ and $J_T/A$ (for a rectangle of depth $2h$, $I/A = (2h)^2/12$). The mechanisms on display — locking, the $w_s = 0$ rank collapse, the missing local-axis transformation — are properties of the discrete space and the assembly, and survive the reduction; the absolute error percentages are the surrogate's, not the published element's.

Test case (truth)
Element formulation
Mesh (6 elements)

9. Assembly, Boundary Conditions, and the Real-Time Solve

Pure standard FE machinery — which is precisely the point. Element contributions are rotated from local to global frames and scatter-added into the global system:

$$\mathbf{K}\,\mathbf{U} = \mathbf{F}, \qquad \mathbf{K} = \sum_e \mathbf{T}^{e\mathsf T} \mathbf{k}^e \mathbf{T}^e, \qquad \mathbf{F}(t) = \sum_e \mathbf{T}^{e\mathsf T} \mathbf{f}^e(t)$$

$\mathbf{K}$ is sparse and symmetric positive semidefinite. Its null space always contains the rigid-body modes, and zeroing a weight can add more (it need not: a weight whose strain rows are already covered by the remaining terms can go to zero without costing rank). Say the condition precisely, because the loose version is wrong on this very page: after boundary conditions, $\mathbf{K}$ is positive definite iff the complete stack of weighted strain operators $\{\sqrt{w_\alpha}\,\mathbf{B}_\alpha\}$ has full column rank on the free DOFs. Suppressing rigid-body motion is necessary, not sufficient. Any spurious mode — iQS4's constant drilling mode on a flat weakly constrained patch (§8), an inextensional mechanism, a deflection field left dangling when $w_s = 0$ — is non-rigid and survives every boundary condition aimed only at rigid-body motion, and survives every positive weight too, because no strain row sees it. Those must be removed by the assembly (fold geometry, neighbouring elements) or constrained explicitly. A free-floating structure yields only deformation relative to the constrained reference — absolute displacement requires known supports.

With $\mathbf{K}$ positive definite, a single sparse Cholesky factorization is computed offline. The monitoring loop then runs on back-substitution alone:

$$\mathbf{K} = \mathbf{L}\mathbf{L}^{\mathsf T} \;\text{(factor once)}, \qquad \mathbf{U}(t) = \mathbf{L}^{-\mathsf T}\mathbf{L}^{-1}\mathbf{F}(t) \;\text{(per frame)}$$

Geometry, mesh, sensor map, and weights are fixed at design time, so $\mathbf{K}$ and its factorization never change. At each measurement frame, only $\mathbf{F}$ is rebuilt — linear in the incoming strain data, trivially cheap — and one forward/back substitution yields the full-field nodal displacements: comfortably real-time even for large meshes. Squeezing this loop harder — element economy, reduced models — is an active 2024–2026 thread (Dong, Yuan & Huang 2024; Xiao et al. 2026; §11).

Trap — $\mathbf{F}$ is not a load vector
The right-hand side is a weighted projection of measured strains, $\mathbf{f}^e = \tfrac{A_e}{n}\sum_i w\,\mathbf{B}_i^{\mathsf T}(\text{data})_i$ — it has no relation to applied forces, and no load knowledge is required anywhere in the pipeline. Likewise $\mathbf{K}$ is not a stiffness matrix: in plain FSDT iFEM no elastic moduli appear in it at all, and even in RZT-iFEM the moduli that do enter arrive as fixed kinematic coefficients, not as a constitutive law relating stress to strain. Treating $\mathbf{K}\mathbf{U}=\mathbf{F}$ as "stiffness times displacement equals force" imports intuitions (about load paths, about stiffness scaling) that are simply false here.
strain sensors rosettes + FBG fiber station data ε⁺, ε⁻ at (xᵢ, yᵢ) known positions, known orientations section strains eε = ½(ε⁺+ε⁻) κε = (ε⁺−ε⁻)/2h g: no data → weight λ element arrays kᵉ = (Aᵉ/n) Σ w BᵗB fᵉ = (Aᵉ/n) Σ w Bᵗ(data) B blocks sparse K U = F K factorized once (offline) F updated every frame (online) → full-field displacement U reconstructed shape heatmap optional post-processing: stress recovery σ = C(E, ν) · ε(U) material properties enter HERE only — no loads anywhere (plain FSDT iFEM; RZT also needs the layer shear moduli, §10)
The iFEM pipeline: sensors → station data → section strains → element arrays → one offline factorization plus a per-frame right-hand side. Loads never enter anywhere. Stress recovery is the only place material properties enter for plain FSDT iFEM; the RZT family also needs the layer shear moduli up in the kinematics box (§10).

Post-Processing and Practical Craft

Post-processing. With $\mathbf{U}$ in hand: strains anywhere in the structure from $\mathbf{B}\mathbf{U}$; stresses through the constitutive law — in plain FSDT iFEM the first and only place material properties appear, though RZT-iFEM has already spent the layer shear moduli up in its kinematics (§10); damage metrics from discrepancies between reconstructed and measured strain fields.

Craft, three items. (a) Elements without sensors get the small $\lambda$ weights — but very large uninstrumented patches with tiny $\lambda$ degrade conditioning and accuracy together, so the sensor layout should track expected strain gradients. (b) Deformation patterns that generate no signal in the instrumented set — torsion of a plate instrumented only with axial gauges on the neutral axis — are reconstructed arbitrarily badly. Observability is a placement property; systematic placement optimization is treated in Module 6. (c) Absolute displacement requires known supports; anchor the pose or report relative deformation only.

10. RZT-iFEM: Zigzag Kinematics for Composites and Sandwich

Why FSDT breaks. In sandwich and thick laminated structures, the transverse shear moduli of face sheets and core differ by orders of magnitude. The cross-section then does not remain straight — it warps in a piecewise, layer-by-layer "zigzag" pattern. FSDT's straight-normal assumption reconstructs both displacements and through-thickness strain distributions poorly. Feeding perfect strain data into the wrong kinematic theory returns the wrong shape: the kinematic assumption is part of the model, and here it fails first.

The fix, at minimal cost. The Refined Zigzag Theory (RZT), developed by Tessler, Di Sciuva & Gherlone — first for beams (2009), then plates (J. Mechanics of Materials and Structures, 2010) — adds per-direction zigzag amplitudes $\psi_x(x,y), \psi_y(x,y)$, multiplied by piecewise-linear through-thickness zigzag functions $\phi^{(k)}(z)$ computed from the layerwise transverse shear moduli:

$$u_x^{(k)} = u + z\,\theta_x + \phi_x^{(k)}(z)\,\psi_x, \qquad u_y^{(k)} = v + z\,\theta_y + \phi_y^{(k)}(z)\,\psi_y$$

Seven kinematic variables $(u, v, w, \theta_x, \theta_y, \psi_x, \psi_y)$, $C^0$ continuity, no shear correction factors, and layerwise-quality response at fixed DOF count. The in-plane strain gains a third profile term and the transverse shear gains piecewise-constant $\beta^{(k)}$ terms:

$$\varepsilon_{xx}^{(k)} = u_{,x} + z\,\theta_{x,x} + \phi_x^{(k)}\,\psi_{x,x}, \qquad \phi^{(k)}(\pm h) = 0$$
$$\gamma_{xz}^{(k)} = w_{,x} + \theta_x + \beta_x^{(k)}\,\psi_x, \qquad \beta_x^{(k)} = \frac{d\phi_x^{(k)}}{dz} = \frac{G_x}{G_x^{(k)}} - 1$$

with $G_x$ the thickness-weighted harmonic mean $\bigl[\tfrac{1}{2h}\sum_k h^{(k)}/G_x^{(k)}\bigr]^{-1}$ of the layer transverse-shear stiffnesses (Kefal, Tessler & Oterkus 2017, eqs. 5b–5c). Read that formula before drawing the zigzag. For the thick-soft-core sandwich drawn below — core at about two thirds of the thickness, $G_{\text{face}}/G_{\text{core}} = 100$ — the harmonic mean sits close to the softest layer, and $|\beta^{(k)}|$ comes out largest in the stiff face sheets ($\beta_{\text{face}} \approx -0.98$) and smaller in the compliant core ($\beta_{\text{core}} \approx +0.51$). The steep-looking part of the physical displacement profile is still the core — that is where the shear strain is — but the steep part of $\phi(z)$ itself is the faces. Identical layers necessarily get identical $\beta$.

Do not promote that ordering to a rule, though: it depends on the thickness fractions as much as on the moduli. Since $\bar{G}$ is a harmonic mean it lies between the smallest and largest layer stiffness, so the stiffest layer always has $-1 < \beta \le 0$, hence $|\beta| < 1$; the softest layer always has $\beta \ge 0$, with no upper bound. Shrink the soft layer to a tenth of the section at the same modulus ratio and $\bar{G} = 9.17\,G_{\text{soft}}$, giving $\beta_{\text{soft}} = 8.17$ against $\beta_{\text{stiff}} = -0.91$ — the ordering flips. Thick compliant core: steepest $\phi$ in the faces. Thin compliant interlayer: steepest $\phi$ in that interlayer.

stiff face sheet compliant core stiff face sheet Gface >> Gcore FSDT: straight normal RZT: u + zθ + φ(z)ψ in-plane displacement uₓ(z) through a sandwich section φ z z=+h z=−h φ(z): piecewise linear, slope β⁽ᵏ⁾ = G/G⁽ᵏ⁾ − 1, steepest in the faces here; pinned φ(±h) = 0 ⇒ surface strains carry no direct zigzag signature
FSDT vs RZT through a sandwich section: the straight normal cannot follow the layerwise warping; RZT's zigzag term $\phi^{(k)}(z)\psi$ can. The displacement profile bends most in the compliant core, because that is where the transverse shear strain is. Inset: $\phi(z)$ itself, whose piecewise slope is $\beta^{(k)} = G/G^{(k)} - 1$ — equal in the two identical faces, and largest there rather than in the core for this thick-soft-core geometry (with a thin compliant layer instead, the soft layer carries the larger $|\beta|$ — see the text). $\phi$ is pinned to zero at both outer surfaces, which is exactly why surface rosettes cannot see the zigzag amplitudes.

The Inverse Extension

Cerracchio, Gherlone, Di Sciuva & Tessler (Composite Structures 2015) built the first RZT-based inverse formulation for displacement and stress monitoring of sandwich structures. Kefal, Tessler & Oterkus (Composite Structures 2017) developed the enhanced formulation around a novel three-node $C^0$-continuous inverse-shell element named i3-RZT, targeting multilayered composite and sandwich structures including through-thickness stress reconstruction. Formally the change is contained but not trivial: the §5 functional grows from 8 weighted measures to 13, $\Phi_e = \sum_{\alpha=1}^{13} w_\alpha \varphi_\alpha$ (Kefal, Tessler & Oterkus 2017, eq. 29) — 3 membrane, 3 bending, 3 zigzag, plus two transverse-shear groups of two: the FSDT-like pair $\boldsymbol{\gamma}$ and an RZT-specific second pair $\boldsymbol{\eta}$. Each carries its own weight, and neither the zigzag group nor either shear group has a surface-rosette counterpart. The element itself is correspondingly heavier: 9 DOF per node — the seven RZT fields plus a drilling $\theta_z$ and an artificial zigzag drilling $\psi_z$ — so 27 DOF for a three-node element, against iMIN3's 15. (Applications and results live in Module 12; this section owns the formulation.)

Small print — the one term that is not a $\mathbf{B}^{\mathsf T}\mathbf{B}$ block

Count the DOFs against the measures and something does not add up: none of the 13 physical strain measures involves $\psi_z$, so the three artificial $\psi_{zi}$ DOFs have identically zero columns in every $\mathbf{B}_\alpha$. The raw element matrix is therefore singular by construction — and, exactly as in §5, no positive weight $w_\alpha$ can fix it, because a weight only scales rows that already exist. Kefal, Tessler & Oterkus handle this explicitly (2017, Appendix C): a diagonal block is added to the element matrix,

$$\boldsymbol{\Gamma}_{\psi_z}=\operatorname{diag}\bigl(k^{1}_{\psi_z},\,k^{2}_{\psi_z},\,k^{3}_{\psi_z}\bigr),\qquad k^{i}_{\psi_z}=\lambda\,\min\bigl(k^{i}_{\psi_x},\,k^{i}_{\psi_y}\bigr),\quad \lambda \approx 10^{-5}$$

where $k^i_{\psi_x}, k^i_{\psi_y}$ are the existing diagonal entries for the real zigzag-amplitude DOFs, and the corresponding right-hand-side entries are zero. This is not a fourteenth physical strain mismatch: it is an artificial stabilization, and it is the one place on this page where the element matrix is not purely a sum of $w\,\mathbf{B}^{\mathsf T}\mathbf{B}$ terms (§1). Its cousin in iQS4 is the constant drilling mode of §8, with the opposite resolution: iQS4 leans on the assembly, i3-RZT adds the diagonal. Be precise about why, because the tempting explanation is wrong. It is not that assembly can never reach $\psi_z$ — Appendix A of the same paper builds $\mathbf{T}^e$ as nine copies of one $3\times3$ direction-cosine block (eq. A.9), so each nodal $(\psi_x, \psi_y, \psi_z)$ triplet rotates exactly like the translation and rotation triplets do. At a fold, a direction that is pure local $\psi_z$ in one element generally has visible local $\psi_x, \psi_y$ components in its neighbour, and does pick up stiffness. The diagonal is there because that rescue is unreliable: on an isolated or coplanar patch there is no fold to perform it, and the element matrix is singular by construction before assembly is even reached. The added diagonal regularizes every locally invisible $\psi_z$ direction unconditionally — which is what a general-purpose element has to do.

Trap — RZT observability from the surface
In the standard RZT normalization $\phi(\pm h) = 0$, so surface in-plane strains reduce to the FSDT form $\mathbf{e} \pm h\boldsymbol{\kappa}$ and carry no direct zigzag-amplitude information. Presenting RZT-iFEM as fully observable from surface rosettes alone is wrong: measures lacking experimental counterparts enter with small weights (the same device used for transverse shear), and richer observability requires strain data inside the stack — e.g. embedded FBGs at the face–core interfaces.
Trap — the philosophical price
Plain iFEM is material-independent. RZT-iFEM is not: constructing $\phi^{(k)}(z)$ requires the layerwise transverse shear moduli — material information enters through the kinematics, before any stress recovery. The composite extension trades away a piece of iFEM's cleanest property to fix its kinematic blind spot. State both halves: "$\mathbf{K}$ contains no $E, \nu$" is true for FSDT-iFEM, and only half-true for the RZT family.

11. The Element Zoo Today: Curved, Higher-Order, Hybrid

Where the element technology has gone since the classics — all of it composing with, rather than replacing, the iQS4/beam/RZT core.

Curved geometry. iCS8 (Kefal, Ocean Engineering 2019) is an eight-node curved inverse-shell element combining solid-shell kinematics with FSDT assumptions in a curvilinear isoparametric coordinate system. Its stated motivation, in the paper's own words: the curvilinear coordinate system lets iCS8 model cylindrical and curved geometry "with a coarse discretization", and this "can allow a relatively sparse placement of sensors" (Kefal 2019). Read the modal verb. Can allow, not implies — the two densities remain the independent axes of §1, and the same paper demonstrates coarse iCS8 discretizations with both dense and sparse sensor deployments. Coarse geometry is what makes a sparse layout practical; it does not make one automatic.

Higher-order flat elements. An eight-node quadrilateral inverse element for 2D problems (Li, Oterkus & Oterkus, Sensors 2023) continues the pattern of porting the direct-FEM element catalogue, one entry at a time.

Isogeometric. Del Priore & Lampani (Thin-Walled Structures 2024) developed a methodology for applying isogeometric iFEM — NURBS shape functions, exact CAD geometry — to stiffened thin-shell structures: the coarse-mesh/sparse-sensor logic taken to its geometric conclusion.

Hybrid modeling of built-up structures. Esposito, Roy, Surace & Gherlone (Sensors 2023) formulated a hybrid shell–beam iFEM — inverse shells for skins, inverse beams for stiffeners — with experimental validation on a composite wing-shaped panel: the practical answer for realistic stiffened airframes.

Scale and speed. Applications now reach full aerostructure scale: morphing wing shape monitoring (Biscotti, Roy & Gherlone, Computers & Structures 2025); real-time shape sensing of large-scale honeycomb antennas with a displacement-gradient-based variable-size iFEM (Dong, Yuan & Huang, Composite Structures 2024); and efficiency-focused formulations for real-time shape sensing of aerospace plate and shell structures (Xiao et al., Thin-Walled Structures 2026).

Choosing an Element in Practice

SituationElement choiceWhy
General thin-walled assembliesiQS4 (default)6-DOF nodes, folds and junctions assemble cleanly; the field's workhorse
Slender frame members, stiffenersinverse Timoshenko beam — or the hybrid shell–beam strategy1D economy; stiffeners modeled at their own scale
Strong geometric curvaturecurved (iCS8) or isogeometric elementscoarse meshes stay accurate, which makes a sparse sensor layout practical
Core shear compliance, thick laminatesRZT family (i3-RZT)zigzag kinematics; accepts the material-input price
Why it matters — the bridge for this site
Classical iFEM assumes small strains and structural materials. Transferring shape sensing to elastomeric tactile skins pushes against both assumptions at once — and large-deformation (e.g. co-rotational) iFEM variants appearing in the very recent literature are the relevant frontier to watch. The element formulations of this module are the baseline any such extension must reduce to in the small-strain limit.

Flashcards

References

← 4. The iFEM Principle 6. Sensors, Placement & Deployment →