6. Sensors, Placement & Deployment

Strain gauges & rosettes, FBG arrays & OFDR distributed fiber, sensor count & conditioning, placement optimization, weighted sparse iFEM, SEA pre-extrapolation, noise & temperature practice

Contents
1. From Strain to Shape: What the Hardware Must Deliver 2. Resistive Strain Gauges and Rosettes 3. Fiber Optics: FBG Arrays and OFDR Distributed Sensing 4. Sparse Coverage: Weighting for Missing Measurements 5. Sensor Count and Placement 6. SEA Pre-Extrapolation: Manufacturing Virtual Sensors 7. Noise, Temperature, and Mounting 8. A Deployment Recipe — and the Soft-Skin Outlook Interactive: FBG Interrogation & Temperature Cross-Sensitivity Interactive: Sensor-Layout Challenge Interactive: Walkthrough — Deploying an iFEM Network Flashcards

1. From Strain to Shape: What the Hardware Must Deliver

Modules 4 and 5 built the algorithm. This module builds the other half of the system: the physical network that feeds it. Restated from the hardware side, the iFEM contract is short. The algorithm needs, at discrete in-plane stations $\mathbf{x}_i$, the in-plane strain components $[\varepsilon_{xx}, \varepsilon_{yy}, \gamma_{xy}]$ measured at known through-thickness offsets $z$, plus the station coordinates, the shell thickness $2h$, the inverse mesh those stations sit in, and enough kinematic restraints to pin down rigid-body motion. Nothing else — no loads, no elastic moduli, no damping (Tessler & Spangler 2003, 2005). Do not read the restraints as optional bookkeeping: the functional scores strains only, and strains are blind to rigid-body modes, so the unrestrained global $\mathbf{K}$ is singular whatever the weights are — it is the restrained system that is positive-definite (Kefal et al. 2016, Eqs. 11d–12). That material- and load-blindness is precisely what makes iFEM deployable: the sensor system's only calibration burden is geometric (where is each gauge, at which $z$-plane) and metrological (how accurate is the strain reading).

Core Problem
Given a finite budget of strain channels on a real structure, deliver to iFEM the section-strain samples it needs — at the right places, in the right components, with surveyed geometry and calibrated accuracy — despite wiring limits, one-sided access, temperature drift, bonding physics, and channel failures.

In the observation-model language of Module 3, $\mathbf{y} = \mathcal{H}(\mathbf{u}, p, f, X_0) + b + \eta$, this module is entirely about the left half of the equation: what the operator $\mathcal{H}$ actually looks like for real sensors (a gauge-length average of one strain projection, not a point tensor), what sits in the bias term $b$ (temperature, drift, imperfect bonding), and how the placement of the rows of $\mathcal{H}$ decides whether the inverse solve is comfortable or hopeless.

The observation model, hardware edition
A single strain channel $i$ with axis direction $\theta_i$ and gauge length $L_g$ reports $$y_i = \frac{1}{L_g}\int_{\text{gauge}} \big[\varepsilon_{xx}\cos^2\theta_i + \varepsilon_{yy}\sin^2\theta_i + \gamma_{xy}\sin\theta_i\cos\theta_i\big]\, ds \;+\; b_i(\Delta T, \text{drift}, \text{bond}) \;+\; \eta_i$$ — the average over the gauge length of the axial projection of the surface strain field, plus bias, plus noise. No sensor in this module measures a strain tensor at a point; the projection formula $\varepsilon_\theta = \varepsilon_{xx}\cos^2\theta + \varepsilon_{yy}\sin^2\theta + \gamma_{xy}\sin\theta\cos\theta$ is why one sensing axis yields one scalar, and why three axes are needed for the three in-plane components.

Back-to-Back Pairs: the Canonical Instrumentation

For plate/shell inverse elements (iMIN3, iQS4 — Module 5), the canonical instrumentation is a pair of triaxial rosettes bonded back-to-back at the two outer surfaces $z = \pm h$. The linear through-thickness strain profile then splits cleanly — half-sum for membrane, half-difference over the full thickness for curvature (derived in Module 4; recapped here because everything in this module feeds it):

$$\mathbf{e}_i^{\varepsilon} = \frac{1}{2}\left(\boldsymbol{\varepsilon}_i^{+} + \boldsymbol{\varepsilon}_i^{-}\right), \qquad \boldsymbol{\kappa}_i^{\varepsilon} = \frac{1}{2h}\left(\boldsymbol{\varepsilon}_i^{+} - \boldsymbol{\varepsilon}_i^{-}\right), \qquad \boldsymbol{\varepsilon}_i^{\pm} = \begin{bmatrix} \varepsilon_{xx} & \varepsilon_{yy} & \gamma_{xy} \end{bmatrix}^{\mathsf{T}}\Big|_{z=\pm h}$$
Trap — the $2h$ convention
iFEM papers define the total thickness as $2h$, so the curvature is the surface-strain difference divided by $2h$. If your drawings use $t$ for total thickness, the formula is $\boldsymbol{\kappa} = (\boldsymbol{\varepsilon}^{+} - \boldsymbol{\varepsilon}^{-})/t$ — correct only if you keep $t = 2h$ consistently. Mixing the two conventions silently doubles or halves every curvature in the network, and nothing in the algebra will warn you: the reconstruction just comes out wrong by a factor of two in bending.
Why two planes are the minimum
A single surface reading is $\boldsymbol{\varepsilon}^{+} = \mathbf{e} + h\boldsymbol{\kappa}$: one equation, two unknowns per component. One accessible surface — a sealed wing box, a fuel tank — leaves the membrane/bending split ambiguous unless membrane action is provably negligible (then $\boldsymbol{\kappa} \approx \boldsymbol{\varepsilon}^{+}/h$) or a second sensing plane is embedded between plies. Wherever membrane strain and curvature coexist — any panel under combined in-plane and out-of-plane loading — single-surface data quietly aliases membrane strain into curvature. Unsymmetric laminates make this unavoidable even under pure in-plane load, because their laminate $\mathbf{B}$ matrix couples stretching and bending constitutively; a layup symmetric about the midplane — including a symmetric sandwich, identical facesheets about a symmetric core — has $\mathbf{B} = \mathbf{0}$ and is not intrinsically coupled. Module 5 §3 works the $z = \pm h$ case in full — all three membrane strains, all three curvatures, rosette transformations; Section 7 below generalizes it to two sensing planes at arbitrary $z_1 \neq z_2$.

Transverse shear strains $\mathbf{g}$ have no direct surface counterpart: no axial projection of a surface strain field contains them, so no rosette reads $\mathbf{g}$ — not a better one, not more of them. They can be obtained indirectly, and the iFEM literature says so: Kefal et al. (2016) note that surface strains “cannot be directly used” for $\mathbf{g}$, but that smoothing the measured curvature field (SEA, Section 6) yields accurate first derivatives of $\boldsymbol{\kappa}$, which can then be used to obtain in-situ transverse shear strains — and add that for thin shells the $\mathbf{g}$ contribution is small enough to omit safely. Absent such an estimate, iFEM carries the analytic shear term with a zero datum and a small weight, purely as regularization; Section 4 makes that precise. Do not read $w_s$ as weighting a measurement: in that default treatment there is no measurement.

The Numbers You Design Against

Scale matters more than folklore. A 3 mm aluminum panel loaded to typical service levels sees surface strains of order $100$–$1000\ \mu\varepsilon$. Foil-gauge and FBG noise floors are of order $1$–$5\ \mu\varepsilon$. Raw signal-to-noise is therefore comfortable — roughly $20$ at the pessimistic end of both ranges ($100\ \mu\varepsilon$ against a $5\ \mu\varepsilon$ floor) to $1000$ at the optimistic end, i.e. one to three orders of magnitude depending on where you sit. The deployment battles are elsewhere: channel count (Section 2 does the arithmetic), coverage (Sections 4–6), temperature drift (a few kelvin can fake tens of $\mu\varepsilon$; Section 7), and strain transfer through the adhesive bond (Section 7). Every one of these is a systematic effect, not noise — averaging does not remove it.

Rosette Algebra

A rectangular 0°/45°/90° rosette reads three axial strains; inverting the projection formula gives the Cartesian components:

$$\varepsilon_{xx} = \varepsilon_{0^\circ}, \qquad \varepsilon_{yy} = \varepsilon_{90^\circ}, \qquad \gamma_{xy} = 2\varepsilon_{45^\circ} - \varepsilon_{0^\circ} - \varepsilon_{90^\circ}$$
Trap — engineering shear, not tensor shear
The rosette result $\gamma_{xy} = 2\varepsilon_{45^\circ} - \varepsilon_{0^\circ} - \varepsilon_{90^\circ}$ is the engineering shear strain, $\gamma = 2\varepsilon_{xy}$ — and engineering shear is what iFEM strain vectors expect, matching the FE strain-vector convention from Module 1. Feeding the tensor component $\varepsilon_{xy} = \gamma_{xy}/2$ into a formulation expecting $\gamma_{xy}$ halves the shear contribution of every station in the network. The bug is invisible in uniaxial bench tests (where $\gamma \approx 0$) and surfaces only under torsion-dominated loads.
midplane z = 0 2h ε⁺ at z = +h ε⁻ at z = −h back-to-back 0/45/90 rosettes membrane eᵋ = ½(ε⁺ + ε⁻) curvature κᵋ = (ε⁺ − ε⁻) / 2h transverse shear g — not directly measurable from surfaces → datum 0, small weight wₛ iFEM K U = F embedded fiber, z₁ embedded fiber, z₂ any two planes z₁ ≠ z₂ work: κ = (ε(z₂) − ε(z₁)) / (z₂ − z₁), e = ε(z₁) − z₁κ (Section 7) material- and load-blind: geometry, strains, restraints
Section-strain extraction: back-to-back rosettes at $z = \pm h$ deliver membrane strain by half-sum and curvature by half-difference over $2h$; transverse shear has no direct surface counterpart and enters iFEM as a zero-datum, small-weight regularization term. Inset: embedded fibers at any two distinct $z$-planes recover the same split.

2. Resistive Strain Gauges and Rosettes

The workhorse of experimental mechanics is the metal-foil resistive gauge: a zigzag conductor on a polymer backing, bonded to the surface, changing resistance as it stretches. The physics fits in one definition:

Gauge Factor and Bridge Output
$$S_g = \frac{\Delta R / R}{\varepsilon} \approx 2.0\text{–}2.2 \quad (\text{constantan}), \qquad \frac{V_o}{V_E} \approx \frac{S_g\, \varepsilon}{4} \;\; (\text{quarter bridge, linearized})$$ The gauge factor of constantan foil is dominated by geometry (the $1 + 2\nu$ term of a stretching conductor) plus a small piezoresistive contribution. The linearized quarter-bridge output is about $0.5\ \mu\text{V}$ per $\mu\varepsilon$ per volt of excitation — which is why 24-bit bridge DAQs resolve $\sim 1\ \mu\varepsilon$.

Standard gauge resistances are 120 Ω and 350 Ω, with 1–5 V excitation. Prefer 350 Ω and low excitation on poorly conducting or polymer substrates: excitation current heats the grid, and a substrate that cannot carry the heat away turns self-heating into a standing strain error. Two wiring disciplines matter in practice: three-wire quarter-bridge connections cancel lead-wire resistance to first order, and a half-bridge with a dummy gauge on an unloaded coupon of the same material gives passive temperature compensation for free — free of wiring cost, at least; Section 7 shows what it silently subtracts along with the drift.

Rosettes come in two habits: rectangular 0°/45°/90° (inversion formulas in Section 1) and delta 0°/60°/120°. Beyond the Cartesian components, the same three readings give the principal directions:

$$\tan 2\theta_p = \frac{2\varepsilon_{45^\circ} - \varepsilon_{0^\circ} - \varepsilon_{90^\circ}}{\varepsilon_{0^\circ} - \varepsilon_{90^\circ}}$$

The Error Budget of an Installed Gauge

Error sourceTypical magnitudeMitigation
Transverse sensitivityorder 1% of transverse strainmanufacturer correction factors
Alignment errorgrows with strain anisotropylayout templates; rosettes (self-checking)
Adhesive creep / hysteresisload-history dependentcyanoacrylate for lab campaigns; cured epoxy for long-term
Zero driftslow, thermal + agingscheduled zero-load re-referencing
Fatiguelimit ∼±1500 $\mu\varepsilon$ for >106 cyclesderate for vibration monitoring
Installed-system accuracytypically 1–3% of readingcalibration against known loads

The Wiring Wall

Now the deployment arithmetic that motivates the rest of this module. Each rosette is three bridge channels. Instrumenting both surfaces of a modest 30-element inverse mesh at one station per element costs $30 \times 2 \times 3 = 180$ channels — each one a run of shielded copper, a bridge completion, a connector, a failure point. Sensor cost is irrelevant; foil gauges are cheap. The wiring is what kills dense resistive networks on real structures.

Why it matters
The wiring wall forces every practical iFEM deployment into one of two exits: go sparse — few sensors, with the mathematics of Sections 4–6 making up the coverage — or go fiber-optic — Section 3, where one fiber replaces dozens of copper triples. The experimental comparison of shape-sensing methods on a wing-shaped plate by Gherlone, Cerracchio & Mattone (2018) is the anchor reference for instrumentation baselines in iFEM shape sensing: what was actually bonded, where, and what accuracy survived installation.

3. Fiber Optics: FBG Arrays and OFDR Distributed Sensing

A fiber Bragg grating (FBG) is a periodic modulation of the refractive index, pitch $\Lambda$, written into the core of an optical fiber. It reflects a narrow spectral peak at the Bragg wavelength:

$$\lambda_B = 2\, n_{\text{eff}}\, \Lambda$$

With $n_{\text{eff}} \approx 1.45$, a grating pitch of $\Lambda \approx 535$ nm reflects at $\lambda_B = 1550$ nm — the telecom C-band, where interrogator hardware is a commodity. Strain stretches the pitch and changes the index; temperature does both too:

FBG Strain–Temperature Response (KEY EQUATION)
$$\frac{\Delta\lambda_B}{\lambda_B} = (1 - p_e)\,\varepsilon + (\alpha_f + \xi)\,\Delta T$$ with effective photoelastic coefficient $p_e \approx 0.22$, giving $\approx 1.2\ \text{pm}/\mu\varepsilon$ at 1550 nm. The thermo-optic coefficient $\xi \approx 6.7\times10^{-6}/\text{K}$ dominates silica's own CTE $\alpha_f \approx 0.55\times10^{-6}/\text{K}$, giving $\approx 10$–$13\ \text{pm/K}$. Keep the ratio in your head: one kelvin looks like roughly nine microstrain. Section 7 and the interactive below deal with the consequences.

Wavelength-division multiplexing (WDM) puts tens of gratings on one fiber, each parked at its own nominal wavelength inside the interrogator's C-band window; the multiplexing architecture goes back to the foundational fiber-grating-sensor work of Kersey et al. (1997). Each grating needs a guard band covering its expected strain range — $\pm1000\ \mu\varepsilon$ maps to $\pm1.2$ nm — and interrogation rates reach kHz, which suits dynamic shape sensing.

The embedded-sensing advantage is what made FBGs the smart-structure default (the multiplexed fiber-grating sensing case is made in Kersey et al. 1997): a 125 μm glass fiber — roughly 150–250 μm outside diameter once coated, depending on the coating chemistry — lays between composite plies with minimal intrusion, is immune to EMI, needs no bridge completion, and replaces dozens of copper triples with a single lead. For what a shape-sensing installation delivers as a system — accuracy once it is on the structure, rather than fiber hardware specifications — the anchor is the experimental comparison of Gherlone, Cerracchio & Mattone (2018) on a wing-shaped plate; that benchmark is not itself an embedded-FBG demonstration, so read it for shape-sensing performance and not as evidence for what embedding specifically buys.

Trap — a fiber is not a rosette
An optical fiber reads only the axial strain along its own path — the projection formula of Section 1 with $\theta$ fixed to the local fiber direction, averaged over the grating length. It is not a point measurement of a tensor component: the grating returns one wavelength for the gauge-length-averaged axial strain over its few-millimetre extent. Feeding iFEM's three in-plane components from fibers therefore requires rosette-like routing — legs at 0°/45°/90° — or an explicitly justified uniaxial-dominance assumption (a slender spar in bending, for instance). One straight fiber down a plate does not deliver $[\varepsilon_{xx}, \varepsilon_{yy}, \gamma_{xy}]$, no matter how many gratings it carries.

OFDR: From Sparse Points to a Quasi-Continuous Profile

Optical frequency-domain reflectometry (OFDR) drops the gratings entirely. A swept laser interrogates the fiber's intrinsic Rayleigh backscatter — the frozen-in random fingerprint of the glass — and cross-correlates it against an unstrained reference trace. The local spectral shift obeys the same strain/temperature form as the FBG equation, but is available quasi-continuously: sub-millimetre to millimetre gauge pitch over tens of metres of fiber, at full-profile rates of order 10–250 Hz.

That turns one fiber into thousands of virtual strain gauges and converts iFEM's sparse-data problem into a dense-data one. Zhao, You & Ren (2024) drive iFEM crack detection from OFDR strain grids; Wu et al. (2025) pair distributed fiber with solid-element iFEM to move beyond plate kinematics entirely. Brillouin-based systems reach kilometres of range but with metre-scale spatial resolution — generally too coarse to serve as inverse-element strain stations.

Foil rosette networkFBG array (WDM)OFDR distributed
Points per line1 rosette = 3 channelstens of gratings / fiberthousands (mm pitch)
RatekHzkHz∼10–250 Hz full profile
Strain components3 in-plane per stationaxial only, per gratingaxial only, continuous
Wiring per line3 shielded copper runs / rosetteone fiber leadone fiber lead
EMI / embeddingvulnerable / surface onlyimmune / embeds between pliesimmune / embeds between plies
Best ataccessible metal surfaces, short campaignsdynamics, harsh environmentsspatial density, damage-scale gradients
Foil rosette 0° 45° 90° Wheatstone bridge Sₖ ≈ 2 · 3 channels/rosette · ±1–3% · kHz FBG array (WDM) 5 gratings, one fiber, one lead λ → peaks + guard bands λB=2nΛ · ~1.2 pm/με · tens/fiber · kHz ~10–13 pm/K ⇒ compensate! OFDR distributed one serpentine fiber = a strain field scan Rayleigh backscatter → continuous ε(s) mm pitch · thousands of points · 10–250 Hz fibers sense axial strain only → rosette-style routing, or an explicitly justified uniaxial assumption
The three sensing technologies. Foil rosettes deliver the full in-plane strain state but cost three shielded channels per station; FBG-WDM arrays multiplex tens of axial stations on one lead at kHz rates; OFDR reads the fiber's Rayleigh fingerprint quasi-continuously at millimetre pitch. Both fiber technologies share the strain–temperature cross-sensitivity and the axial-only limitation.
Interactive Tool — FBG Interrogation & Temperature Cross-Sensitivity

One sensing grating at $\lambda_B = 1550$ nm, $K_\varepsilon = 1.2$ pm/$\mu\varepsilon$, $K_T = 11$ pm/K. Apply strain and temperature, watch the reflected peak move, and see what the interrogator thinks the strain is — with and without a strain-free reference grating. The reference is parked at its own nominal 1552 nm: WDM requires every grating to own a guard band wide enough for its expected shift, so a reference sharing 1550 nm with the sensing grating would be optically indistinguishable from it. Mind the consequence of the key equation above: the thermal response is fractional, so a reference at $\lambda_r$ shifts by $K_T\,\lambda_r/\lambda_B$ per kelvin — 11.014 pm/K at 1552 nm against 11 pm/K at 1550 nm. Subtracting the two raw picometre shifts would therefore leave $-0.71\ \mu\varepsilon$ at $\Delta T = 60$ K, so the widget subtracts fractional shifts $\Delta\lambda/\lambda$ instead. Read the strain slider literally: it is the fiber's own mechanical strain, which is what a free reference grating cancels down to exactly, so compensation here is clean by construction. On a grating bonded to a host that quantity is $\varepsilon_{\text{host}} - \alpha_f\Delta T$, and the leftover $\alpha_f\Delta T \approx 0.55\ \mu\varepsilon$/K against host strain is the residual Section 7 discusses.

4. Sparse Coverage: Weighting for Missing Measurements

Now the formalism that lets a network survive being sparse. This module is the canonical home of the weighted-iFEM machinery; Module 4 gave the motivating paragraph. Recall the element-level weighted least-squares functional (Tessler & Spangler 2005; Kefal et al. 2016 for iQS4): it compares the analytic section strains — membrane $\mathbf{e}(\mathbf{u})$, bending $\boldsymbol{\kappa}(\mathbf{u})$, transverse shear $\mathbf{g}(\mathbf{u})$ — against their measured counterparts,

$$\Phi_e(\mathbf{u}^e) = w_m \left\| \mathbf{e}(\mathbf{u}^e) - \mathbf{e}^{\varepsilon} \right\|^2 + w_b \left\| \boldsymbol{\kappa}(\mathbf{u}^e) - \boldsymbol{\kappa}^{\varepsilon} \right\|^2 + w_s \left\| \mathbf{g}(\mathbf{u}^e) - \mathbf{g}^{\varepsilon} \right\|^2$$

with the squared norms defined, following Kefal et al. (2016, Eqs. 8b–d), as normalized Euclidean norms: the sum over the element's $n$ strain stations divided by $n$, integrated over the element mid-plane area $A_e$, and applied component by component. Note what the normalization is — station count, not area. Read the three scalar weights the same way: Kefal et al. implement the reduced (small-weight) form “on the component-by-component basis”, so a lone $w_m$ means one common value when all three membrane components are measured; an installation that reads $\varepsilon_{xx}$ but cannot see $\gamma_{xy}$ demotes only the component it is missing, which the single-symbol notation hides. The curvature term carries the $(2h)^2$ factor to homogenize units — curvature is strain per length:

$$\left\| \boldsymbol{\kappa}(\mathbf{u}^e) - \boldsymbol{\kappa}^{\varepsilon} \right\|^2 = \frac{(2h)^2}{n} \int_{A_e} \sum_{i=1}^{n} \left[ \boldsymbol{\kappa}(\mathbf{u}^e)\big|_i - \boldsymbol{\kappa}_i^{\varepsilon} \right]^2 dA$$

Minimizing $\sum_e \Phi_e$ produces element matrices built from B-matrix products weighted by $(w_m, w_b, w_s)$ and element vectors linear in the measured strains; assembly plus the problem's boundary restraints yields a symmetric positive-definite system:

$$\mathbf{k}^e = \int_{A_e} \left( w_m\, \mathbf{B}_m^{\mathsf{T}} \mathbf{B}_m + (2h)^2 w_b\, \mathbf{B}_b^{\mathsf{T}} \mathbf{B}_b + w_s\, \mathbf{B}_s^{\mathsf{T}} \mathbf{B}_s \right) dA, \qquad \mathbf{K}\,\mathbf{U} = \mathbf{F}$$

$\mathbf{K}$ depends only on geometry, mesh, and weights — factorize once per weight pattern; each new strain frame then costs one cheap back-substitution. That is the real-time property, and it survives sparsity untouched. Read the qualifier literally: the weights are part of $\mathbf{K}$, so demoting a dead channel's element to the small-weight regime (the dropout discipline of Section 7) changes $\mathbf{K}$ and invalidates the stored factorization. Refactorize on a weight change, or precompute one factorization per anticipated failure pattern. Absorb the same dropout by pre-extrapolation instead (Section 6) and the weights can be held still — the element still gets a datum — so $\mathbf{K}$ survives and only the smoother is refitted. Two conditions on that, both load-bearing: enough live stations must remain to fit the smoother at all (lose the last one and there is nothing to extrapolate from, so every element falls back to the small weight and $\mathbf{K}$ moves after all), and the extrapolated datum must be carried at the same weight as the reading it replaced — implementations that deliberately discount extrapolated values (Section 6) change $\mathbf{K}$ the moment a physical station becomes a virtual one.

Where Sparse Sensing Enters: the Weights

If an element has no strain station, simply dropping its contribution can leave $\mathbf{K}$ singular — whole patches of the mesh lose their connection to any data. The Tessler–Spangler remedy, spelled out for “strain-less” inverse elements by Kefal et al. (2016), keeps the analytic term with a small weight: drop the measured value from the squared norm — equivalently, set the datum to zero — and set the weight to $10^{-4}$ (the literature uses $10^{-3}$–$10^{-5}$) instead of $1$ for instrumented components:

$$w = \begin{cases} 1 & \text{station present (measured component)} \\ 10^{-4} \text{ (typ.)} & \text{no measurement — analytic term retained, datum set to } 0 \end{cases}$$
Trap — small weight, never zero weight
Setting the weight of uninstrumented elements — or of the transverse-shear term — to exactly zero deletes equations from the least-squares system, and can render $\mathbf{K}$ singular. The scheme requires a small nonzero weight on the retained analytic term. The small-weight term is a soft zero-strain prior, not a measurement and not a smoothness penalty: its target was set to zero in the equation above, so it pulls uninstrumented regions toward zero section strain. (The spatial coupling that lets instrumented neighbours inform them comes from FE continuity and assembly, not from this term; a genuine smoothness regularizer would penalize a derivative of the field.) It buys rank and pays bias, and the bias is largest exactly where you have no data to notice it — so the mesh “bridges” the gaps by interpolation, while this term only supplies the rank that makes the bridged DOFs solvable at all. In the default treatment the transverse-shear term lives in this regime, since $\mathbf{g}$ has no direct surface measurement — there, $w_s$ never weighted data to begin with. Supply an indirect $\mathbf{g}^{\varepsilon}$ estimate — SEA-smoothed curvature derivatives, Section 1 — and $w_s$ becomes an ordinary weight again, expressing how much you trust that estimate.

What Sparse Coverage Costs

Four consequences to design against:

The evidence base, with the two strands kept apart. For small-weight fill-in itself, Kefal et al. (2016) strip the back-to-back rosettes from 2160 of their iQS4 elements — leaving 240 instrumented stations — carry every stripped element at $w = 10^{-4}$, and still land within 3% of the high-fidelity FEM maximum displacement. For placement, Kefal & Yildiz's wing-shaped sandwich panel study (2017) sweeps sensor density and alignment under bending, torsion, and membrane loading; note what it actually runs, because it is easy to miscite: SEA pre-extrapolation (Section 6) is applied for every sensor configuration, and its small weights ($10^{-3}$) mark strain components the uniaxial sensors cannot see, not sensor-free elements filled with zero. Both strands agree on the caveat: accuracy survives sparsity only provided the sensors sit where the strain gradients are informative. Which is the cue for the next section.

5. Sensor Count and Placement

Two regimes: engineering heuristics you can apply on day one, and formal optimization when the layout is worth automating. Both answer the same underlying question from Module 3: each sensor contributes rows to the observation operator, and a layout is good when those rows keep the least-squares problem well-conditioned for the deformations the load envelope can actually produce. Placement is observability engineering by other means.

Heuristics

Engineering heuristics, consistent with what Kefal & Yildiz (2017) and Roy et al. (2020) report but stated here as rules of thumb rather than as findings of either paper:

Roy et al. (2020) made the pattern question systematic: varying sensor pattern and density on a rectangular plate reconstructing bending and torsional deformation modes, they showed that well-chosen sparse patterns lose little accuracy versus dense coverage — boundary-distributed sensors suffice for relatively simple deflection shapes, cross-diagonal patterns improve the reconstruction of more complex deformation patterns, and accuracy converges toward the reference FE solution as density grows.

Trap — “more sensors is always better”
Under measurement noise, returns diminish once the informative regions are covered. Placement quality dominates raw count: an optimized sparse layout can approach dense-coverage accuracy at a fraction of the budget, while a channel spent in a quiet region returns almost nothing for its cost — not because the extra station poisons the rest of the network, but because the same channel on a load path would have bought real accuracy. (Locally it does trade the fill-in's bias for measurement variance; Section 4 has that version of the argument.) Buy placement first, count second.

Formal Optimization: the GA Template

Ghasemzadeh & Kefal (2022) formalized placement as a discrete optimization over the inverse mesh. Encode a candidate layout as a binary vector over the $N_e$ inverse elements, with a hard sensor budget:

$$\mathbf{s} \in \{0,1\}^{N_e}, \qquad \sum_{e=1}^{N_e} s_e = N_s \ (\text{sensor budget})$$

The fitness of a layout is the weighted sum of squared differences between the candidate's iFEM solution — weights set by $\mathbf{s}$: $w = 1$ where $s_e = 1$, $10^{-4}$ elsewhere — and the full-coverage reference solution, summed over the $N$ nodes for the load case being optimized:

$$f(\mathbf{s}) = \sum_{i=1}^{N} \left[ \phi_1\left(u_i - u_i^{\text{ref}}\right)^2 + \phi_2\left(v_i - v_i^{\text{ref}}\right)^2 + \phi_3\left(w_i - w_i^{\text{ref}}\right)^2 + \phi_4\left(\theta_{x,i} - \theta_{x,i}^{\text{ref}}\right)^2 + \phi_5\left(\theta_{y,i} - \theta_{y,i}^{\text{ref}}\right)^2 \right] \;\rightarrow\; \min$$

and the population evolves with standard selection, crossover, and mutation. The trick that makes this affordable: each fitness evaluation is one small SPD solve — the cheapness of the iFEM forward pass, exploited as an inner loop. Note the index set: the sum runs over nodes only, and Ghasemzadeh & Kefal run one optimization per load case. Covering an envelope of $L$ load cases means an outer sum over them, and then each evaluation costs one factorization of the candidate's $\mathbf{K}$ plus one back-substitution per load case — still cheap, but $L$ times the arithmetic, and it is the version the caveat box below demands. Their plate, stiffened-plate, and curved-shell results show GA-optimized sparse layouts approaching fully instrumented accuracy at a fraction of the sensor count. Multi-objective and swarm-based variants (trading error against sensor count) appear in the surrounding literature, and Li et al. (2025) fold placement optimization and strain pre-extrapolation (Section 6) into a single framework.

Two caveats that bind every “optimal” layout
(1) Optimality is conditional on the assumed load envelope. Optimize across load cases and inject noise into the fitness, or the layout overfits one load and fails on the next. (2) The fitness is measured against a model, so placement quality inherits the model's form error — a layout tuned on a pristine FE reference does not know about the stiffener the model simplified away. A layout presented as universally optimal is a layout whose training envelope has not been stated. Practical recipe: seed the optimizer with the heuristic layout, not with random chromosomes.
(a) dense — baseline error: low (b) uniform sparse error: moderate (c) tip-clustered error: worst — sensors in ε ≈ 0 (d) root-biased / GA error: close to (a) same budget in (b), (c), (d) — only the placement differs; shading in (d): strain-magnitude ghost for the load envelope GA fitness f(s) = weighted squared displacement/rotation error vs full-coverage reference, budget Σ sₖ = Nₛ (after Roy et al. 2020; Ghasemzadeh & Kefal 2022 — qualitative)
Four layouts, one budget (panels b–d). Tip-clustered sensors sit in near-zero strain and waste the budget; root-biased or GA-optimized layouts track the informative strain regions and approach dense-coverage accuracy. Try exactly this experiment live in the Sensor-Layout Challenge below.

6. SEA Pre-Extrapolation: Manufacturing Virtual Sensors

The second sparse strategy repairs the measurement set before the inverse solve, instead of patching the functional with small weights. Its engine predates iFEM by a decade: Smoothing Element Analysis (SEA) was built by Tessler, Riggs & Macy (1994) to recover $C^1$-continuous stress fields from the discrete, inter-element-discontinuous stress points of a forward FEM, with a posteriori error estimation; Tessler, Riggs, Freese & Cook (1998) refined the variational formulation. Pre-extrapolating measured strains is a later repurposing of the same machinery, not what the original papers set out to do. The mechanism:

The SEA Functional (KEY EQUATION)
Over a dedicated smoothing mesh (typically triangles), interpolate the smoothed field $\hat{\varepsilon}$ and an independent gradient field $\boldsymbol{\theta}$ with $C^0$ shape functions, and minimize $$\Phi_{\text{SEA}}(\hat{\varepsilon}, \boldsymbol{\theta}) = \sum_e \left\{ \frac{1}{n_e} \sum_{i=1}^{n_e} \left[ \hat{\varepsilon}(\mathbf{x}_i) - \varepsilon_i \right]^2 + \lambda \int_{A_e} \left( \nabla\hat{\varepsilon} - \boldsymbol{\theta} \right)^{\mathsf{T}} \left( \nabla\hat{\varepsilon} - \boldsymbol{\theta} \right) dA + \beta\, A_e \int_{A_e} \left[ \theta_{x,x}^2 + \theta_{y,y}^2 + \tfrac{1}{2}\left( \theta_{x,y} + \theta_{y,x} \right)^2 \right] dA \right\}$$ — assembled element by element over the smoothing mesh: a normalized least-squares misfit at the $n_e$ data points falling inside each element (read that term piecewise — a smoothing element containing no data point drops it entirely rather than evaluating $1/n_e$ at $n_e = 0$, and contributes only the two penalties, which is exactly how a smoothing mesh finer than the sensor network stays well posed; Del Priore & Lampani 2024); a penalty forcing $\boldsymbol{\theta} \to \nabla\hat{\varepsilon}$; and a second penalty on the gradients of $\boldsymbol{\theta}$, which is what actually controls how hard the field is smoothed. Two hyperparameters, not one. Do not drop the element area $A_e$ in front of the $\beta$ integral: $\theta$ has units of 1/length, so that integral scales as $1/\text{length}^2$, and only the $A_e$ factor leaves $\beta$ (like $\lambda$) dimensionless and the smoothing strength consistent across element sizes on a nonuniform mesh (the elementwise form is spelled out in Del Priore & Lampani 2024, Eq. 12). The $\lambda$ penalty structure mirrors Mindlin-plate shear penalties; in the large-$\lambda$ limit (typically $10^2$–$10^6$) the fit is effectively $C^1$, and the solve is a small sparse symmetric system — positive semidefinite in general, since any affine $\hat{\varepsilon} = a + bx + cy$ with $\boldsymbol{\theta} = (b, c)$ costs both penalties nothing, and positive-definite once the misfit term pins that affine mode down (three non-collinear stations suffice). Crucially, the fitted field can then be evaluated anywhere.

“Evaluated anywhere” is the whole idea. The pre-extrapolation pipeline for sparse iFEM (Roy et al. 2022; Oboe et al. 2021):

  1. Measure strains at the sparse physical stations.
  2. Run SEA separately on each of the six section-strain components — three membrane from half-sums, three curvature from half-differences.
  3. Sample the smoothed fields at every inverse element's strain stations — these are the virtual sensors.
  4. Run iFEM with full coverage: every element now carries a datum from the fit, so the zero-filled gaps are gone. The weights are a separate, implementation-dependent choice — unity at the physical stations, and commonly a reduced weight on elements whose datum was extrapolated rather than measured (Poloni et al. 2023 use $10^{-1}$–$10^{-3}$ there), because an extrapolated strain is not as trustworthy as a reading. Replacing the zero target is the part SEA guarantees; raising the weight to unity is not.

The evidence. Oboe et al. (2021) compared pre-extrapolation techniques — SEA against polynomial fitting — on a composite plate in compression buckling, a deliberately nonuniform strain field, and found that pre-extrapolation markedly improves both shape and strain reconstruction over the plain small-weight approach at equal sensor count. Roy et al. (2022) showed SEA-aided iFEM on plates reaching accuracy comparable to much denser physical networks: sensor count bought back in software. The family keeps growing: Oboe, Sbarufatti & Giglio (2022) introduced a physics-based pre-extrapolation technique that informs the fill-in with structural knowledge rather than generic smoothness, and Poloni et al. (2023) extended pre-extrapolation to variable-thickness structures — thickness steps induce strain-field discontinuities that violate naive smoothing, and extrapolating in a thickness-normalized space, where the thickness-induced trends are removed, restores accuracy.

Trap — virtual sensors are not new information
Pre-extrapolation manufactures pseudo-data, and pseudo-data is correlated: every virtual strain is a deterministic function of the same few physical readings, so it shares their noise and adds the smoother's bias on top. A thousand virtual stations do not out-vote ten physical ones — the information budget is fixed by the physical network. Two corollaries. First, pre-extrapolation cannot restore unobserved information: if a deformation pattern is invisible to the physical layout, SEA fills the gap with a plausible invention, and iFEM will confidently reconstruct the invention. Second, never report the reconstruction residual at virtual stations as validation — the fit agreeing with itself proves nothing. Validate at withheld physical sensors (Section 8).
Trap — what SEA is not
Two common mislabelings. SEA pre-extrapolation is not post-smoothing of the iFEM output — it repairs the measurement set before the inverse solve, which is why it can change what the solve converges to. And SEA is not generic polynomial regression — it is a penalty-constrained $C^1$ finite-element least-squares fit with its own mesh and its own two penalty parameters ($\lambda$ and $\beta$ above); Oboe et al. (2021) tested it head-to-head against polynomial fitting precisely because the two behave differently on nonuniform fields.

Limits to state plainly: SEA interpolates well inside the sensor hull but extrapolates poorly outside it — uninstrumented edges remain unreliable regardless of smoothing; the smoothing-mesh density and both penalty parameters are hyperparameters someone must choose; and dynamic data needs a per-frame refit — cheap, since the SEA system matrix factorizes once, like iFEM's own.

$$\text{virtual sensor:}\quad \mathbf{e}^{\varepsilon}(\mathbf{x}_v) = \hat{\mathbf{e}}(\mathbf{x}_v), \quad \boldsymbol{\kappa}^{\varepsilon}(\mathbf{x}_v) = \hat{\boldsymbol{\kappa}}(\mathbf{x}_v) \quad \forall\, \text{iFEM stations } \mathbf{x}_v$$
sparse measured strains (few physical stations) Path A — small-weight fill-in w = 1 (colored), w = 10⁻⁴, datum 0 (grey) iFEM sag toward 0-strain fill in dark regions Path B — SEA pre-extrapolation smooth fit ε̂, misfit + λ∫(∇ε̂−θ)² dA virtual sensors everywhere, datum ≠ 0 iFEM close to truth at equal sensor count dashed = true deflection. Same physical sensors in both paths — only the treatment of the gaps differs (Oboe et al. 2021; Roy et al. 2022). Caution: Path B's virtual stations are correlated pseudo-data — the information still comes from the three orange dots.
Two treatments of the same sparse measurement set. Path A retains uninstrumented elements at $w = 10^{-4}$ with zero datum — robust, but the reconstruction sags toward zero strain in dark regions. Path B fits a $C^1$ SEA field through the physical readings and samples it at every element (“virtual sensors”), so no element is left with a zero datum — markedly more accurate at equal sensor count on smooth-but-nonuniform fields. What weight an extrapolated datum then carries is an implementation choice: unity at the physical stations, often deliberately less where the value was invented.
Interactive Tool — Sensor-Layout Challenge

A cantilever strip (L = 500 mm, thickness 2h = 3 mm, 10 inverse elements) with a 12-channel reconstruction budget — the six withheld validation gauges described below are test-only instrumentation and are deliberately not counted against it, so the physical channel count on the specimen is however many reconstruction channels the layout actually installs, plus six — 18 only when the 12-channel budget is exactly filled, and the layout you land on may well sit under or over it (the counter above reads the installed total). Click an element chip to cycle its sensor: none → back-to-back axial pair (2 ch, gives $\mathbf{e}$ and $\boldsymbol{\kappa}$) → top-only axial gauge (1 ch, no membrane reading of its own) → FBG grating (1 ch, 5 mm gauge-length average, $\Delta T$-sensitive). Only the FBG averages over a gauge length here; the foil pairs, the top-only gauges, and the withheld validators are idealized as point samples, so the gauge-length physics of Section 1 shows up in exactly one place rather than everywhere. The strip is a 1-D bending problem, so one axial component per station is all the model can use — that is why a pair costs 2 channels here. Instrument a plate with triaxial rosettes instead and the accounting triples: 3 channels per rosette, 6 for a back-to-back pair. Then inject faults — temperature, debonding, dropout — and judge the layout by the withheld-gauge error: six held-out top-surface gauges (◇) that the solve never sees, each carrying its own independent noise, exactly as a real validation channel would. Even a flawless, fully instrumented network leaves this metric at $\approx 20$–$25\ \mu\varepsilon$ here: each inverse element consumes one constant curvature sampled at its midpoint, so the reconstructed surface strain is a staircase across a field that genuinely varies, and a withheld gauge sitting away from its element's midpoint reads that discretization bias directly. Compare it against the training residual, which is scored per live structural-strain channel against the datum that channel handed the solver — already reference-subtracted where FBG compensation is switched on, and with the temperature-reference grating left out, since it carries no structural datum of its own (a dark channel drops out of the count too). Under the small-weight strategy that residual comes out identically zero for every instrumented layout here — the fit passes exactly through the readings it consumed, however wrong the reconstruction is between them, which is the sharpest possible statement of “small residual $\neq$ valid reconstruction”; under pre-extrapolation the solve consumes the smoothed field instead of those readings, so the residual reports the smoother's disagreement with them rather than the solve's. The panel tracks measurements, unknowns, assumptions, conditioning, and validation throughout.

Stations (click to cycle type)
Presets:
Load case (truth)
Gap strategy

7. Noise, Temperature, and Mounting

Noise

Foil bridges and FBG interrogators both deliver noise floors in the 1–5 $\mu\varepsilon$ class. iFEM's least-squares structure averages independent station noise, and the standard assessment is Monte Carlo: superimpose zero-mean Gaussian noise — a fixed $\mu\varepsilon$ level, or a percentage of peak strain — on synthetic measurements and report reconstruction-error statistics across many trials. This is the widely used verification practice in numerical iFEM studies. Each reroll of the interactive above draws one realization from that distribution — press it repeatedly to see the spread by eye; the statistics over trials are the part you would have to add for a publishable noise study.

Synthesis note — not yet standard iFEM practice
If per-channel noise variances genuinely differ, setting $w_e \propto 1/\sigma_e^2$ is the maximum-likelihood weighting under Gaussian noise — the whitening argument of Module 3. The iFEM literature, however, usually keeps binary weights ($1$ or $10^{-4}$); treat variance-proportional weighting as a defensible extension, not established practice.

Temperature: Two Distinct Issues

(a) Sensor-level contamination. Foil gauges exhibit a thermal output

$$\varepsilon_{\text{app}} = \left( \frac{\beta_g}{S_g} + \alpha_s - \alpha_g \right) \Delta T$$

from the gauge alloy's temperature coefficient of resistance $\beta_g$ and the CTE mismatch between substrate ($\alpha_s$) and gauge ($\alpha_g$). Countermeasures: self-temperature-compensated alloys matched to the substrate CTE, half-bridge dummy gauges, or correction from logged temperature — the first two remove more than the drift, as the box below spells out. FBGs read $\approx 10$–$13$ pm/K, so an uncompensated 1 K masquerades as $\approx 9\ \mu\varepsilon$; the standard fix is a co-located strain-free reference grating — a grating in a loose capillary, feeling the temperature but not the strain — subtracted in software:

$$\varepsilon = \frac{1}{1 - p_e} \left( \frac{\Delta\lambda_B}{\lambda_B} - (\alpha_f + \xi)\,\Delta T \right)$$

(b) Structure-level thermal deformation is not an error. iFEM is kinematic: genuinely thermally induced strains reconstruct into genuinely thermally induced displacements, with no thermal constitutive input required. Cerracchio, Gherlone & Tessler (2015) demonstrated iFEM displacement monitoring of a composite stiffened panel under mechanical and thermal loads. If the sun bends your panel, iFEM correctly reports a bent panel.

Key insight — which temperature effect are you fighting?
Separate the two ruthlessly. Sensor-level thermal output is a bias in $b$: the structure did not move, the reading did. Structure-level thermal deformation is signal in $\mathbf{u}$: the structure really moved, and iFEM should say so. Compensation schemes must remove the first without touching the second — which is exactly what a strain-free reference grating does: it feels $\Delta T$ but not the structure's strain, so subtracting it removes the sensor-level term — down to the fiber's own expansion. The bonded grating follows the host, so its mechanical strain is $\varepsilon_{\text{host}} - \alpha_f \Delta T$; differencing against the free reference recovers that, not $\varepsilon_{\text{host}}$. A residual $\alpha_f \Delta T \approx 0.55\ \mu\varepsilon$/K survives — about 6% of the $\approx 9\ \mu\varepsilon$/K it removed, and itself correctable once $\alpha_f$ is calibrated. The foil analogue is the same trap an order of magnitude larger. A self-temperature-compensated alloy, or a half-bridge whose dummy sits on an unloaded coupon of the same material, is arranged to output zero when that material sits unloaded and freely expands; what survives compensation is therefore the strain relative to free thermal expansion, $\varepsilon_{\text{tot}} - \alpha_s \Delta T$ — the stress-producing part, not the kinematic total iFEM reconstructs from. Add the free thermal strain back from logged temperature and a known $\alpha_s$ (order $20\ \mu\varepsilon$/K for aluminium alloys, against the fiber's $0.55$), and treat anisotropy and through-thickness gradients as their own calibration problem. Compensate the sensor, then restore the structure's thermal strain: skip the second half and you have deleted precisely the deformation (b) says iFEM should be reporting.

Mounting: Strain Must Survive the Bond

Strain reaches the sensing element through an adhesive layer working in shear. Shear lag means the sensor under-reads near its ends: the strain carried by the gauge or fiber rises from zero at each end over a transfer length before plateauing — below the structure's strain if the bondline is thick or compliant. Practical rules: keep bondlines thin and stiff; prefer polyimide-coated fiber for bonding and embedding (soft acrylate coatings and loose buffers degrade transfer); align embedded fibers with the ply direction to avoid resin pockets.

strain position along sensor εₛ (structure) sensor strain — plateau below εₛ transfer length substrate (structure), strain εₛ →→→ adhesive — works in shear (thickness exaggerated) sensor (foil backing / coated fiber) thick / soft bondline → longer transfer length, systematic under-reading soft substrate + stiff sensor → local strain shadowing: calibrate! mitigations: • thin, stiff bondlines • polyimide-coated fiber • fibers along ply direction • calibrate installed system
Shear lag in the bond: the structure's strain enters the sensor through adhesive shear, so the sensed strain rises from zero at each end over a transfer length and plateaus below $\varepsilon_s$ when the bondline is thick or compliant. On very soft substrates the stiff sensor additionally shadows the local strain field.

Sensing Planes Away from the Surfaces

Embedded fibers rarely sit exactly at $z = \pm h$. No matter: with the linear through-thickness kinematics $\boldsymbol{\varepsilon}(z) = \mathbf{e} + z\boldsymbol{\kappa}$, any two distinct planes $z_1 \neq z_2$ recover both section strains:

$$\boldsymbol{\varepsilon}(z) = \mathbf{e} + z\,\boldsymbol{\kappa} \;\Rightarrow\; \boldsymbol{\kappa} = \frac{\boldsymbol{\varepsilon}(z_2) - \boldsymbol{\varepsilon}(z_1)}{z_2 - z_1}, \qquad \mathbf{e} = \boldsymbol{\varepsilon}(z_1) - z_1 \boldsymbol{\kappa}$$

This is how embedded-fiber layouts replace surface rosette pairs — and note where the sensitivity actually sits. On exact data the quotient stays equal to $\boldsymbol{\kappa}$ however small the separation, because the numerator vanishes at the same rate as the lever arm. What blows up is the noise: with independent errors of standard deviation $\sigma_1, \sigma_2$ on the two readings, $\operatorname{Var}(\hat{\boldsymbol{\kappa}}) = (\sigma_1^2 + \sigma_2^2)/(z_2 - z_1)^2$, so the curvature error grows as $1/|z_2 - z_1|$ and plane separation is itself a design variable.

Registration, Synchronization, Drift, Dropout

Four operations-side effects complete the error budget. None is exotic; each has bitten a real deployment.

EffectWhat it does to iFEMDiscipline
Registrationa station whose surveyed coordinates are wrong is a correct reading attached to the wrong row of the observation operator — a systematic error no averaging removessurvey positions and orientations into the model frame; treat coordinates as calibration data
Synchronizationchannels sampled at different instants mix load states within one “frame” — harmless for statics, poison for dynamicscommon clock / triggered acquisition across interrogators and bridges
Driftslow zero-shift accumulates into a phantom quasi-static deformationscheduled zero-load re-referencing; temperature logging; drift audits
Dropouta dead channel that keeps reporting zero is a fake measurement of zero strain at full weight — far worse than an acknowledged gapdetect and demote: mark the element as a gap and let the active gap strategy handle it — small-weight fill-in at $w = 10^{-4}$, or a pre-extrapolated virtual sensor (Section 6) — but never ingest zeros as data. Which system needs rebuilding depends on which route you took: under small-weight fill-in the element's weight just went $1 \to 10^{-4}$, so $\mathbf{K}$ changed and must be refactorized; under pre-extrapolation the element still receives a datum, and provided that datum keeps the weight the reading had and at least one live station remains to fit the smoother, $\mathbf{K}$ stands and it is the SEA fit — whose live measurement set just shrank — that has to be rebuilt. Lose the last live station, or carry extrapolated values at a reduced weight, and you are refactorizing after all
Synthesis note — soft substrates
On very soft substrates — elastomeric skins, the tactile-sensing regime of this site's own research context — the sensor's own stiffness locally shadows the strain field it is trying to measure, and the shear-lag picture above becomes the dominant error source rather than a correction. Calibration against a reference measurement is mandatory there; treat this extrapolation beyond the plate-metal literature as synthesis, not established result.

8. A Deployment Recipe — and the Soft-Skin Outlook

The module, condensed into an executable checklist:

  1. Build and validate a forward FEM; define the load envelope. Everything downstream — placement, weighting, validation — is conditional on this envelope.
  2. Choose the inverse mesh — usually much coarser than the forward mesh. Every element must either host a station or sit within bridging distance of instrumented neighbours (Section 4).
  3. Select the sensing technology by channel count, bandwidth, embedding, environment: foil rosettes for accessible metal surfaces and short campaigns; FBG-WDM for dynamics, harsh EMI, embedded composite lines; OFDR for maximum spatial density at moderate rates (Sections 2–3).
  4. Place sensors: heuristic layout from the forward model's strain map, then GA refinement (Ghasemzadeh & Kefal 2022, whose reported runs each score a single reference load state and then re-test one optimized layout under several dynamic histories). Covering a load envelope means extending that fitness with an outer sum over the load cases; injecting measurement noise into it is prudent practice beyond the original noise-free study. Li et al. (2025) unify this step with the next.
  5. Choose the sparse strategy: small-weight fill-in ($w \approx 10^{-4}$) when coverage is moderate and fields are smooth; SEA pre-extrapolation when the network is sparse or the gradients are structured (Roy et al. 2022; Oboe et al. 2021; Poloni et al. 2023 for variable thickness).
  6. Calibrate: apply known static loads; compare iFEM output against dial gauges, DIC, or laser displacement sensors; tune weights and station coordinates.
  7. Operate: pre-factorize $\mathbf{K}$ for the current weight pattern; each frame costs assembling $\mathbf{F}$ from the strain vector plus one back-substitution — kHz-capable. Refactorize whenever a weight changes — a strategy switch, or a dropout handled by small-weight demotion; a dropout absorbed by pre-extrapolation changes the smoother's input rather than $\mathbf{K}$ — but only while a live station survives to seed the fit and extrapolated data keep the weight the readings had.
  8. Maintain: scheduled zero-load re-referencing, temperature logging, drift audits, dropout detection (Section 7).
Trap — validation that actually validates
Two rules for step 6 and for every accuracy claim you publish. Score on withheld sensors, not the training residual: the residual at the sensors the solve consumed measures self-consistency, and systematic errors — a debonded gauge, an uncompensated $\Delta T$ — leave it small while the reconstruction is badly wrong (demonstrate this in the Sensor-Layout Challenge above). And never split validation data by random frames: adjacent frames of the same loading history are nearly identical, so a random frame-level train/test split leaks the test set into training and certifies nothing. Split by specimen, loading path, or experiment.

Outlook: Strain-Based Skins

Synthesis / outlook — flagged as such
Strain-based iFEM with embedded distributed fiber is the tactile-skin analogue of vision-based deformation sensing: OFDR's millimetre gauge pitch matches contact-scale strain fields, OFDR-dense data already powers damage localization (Zhao et al. 2024), and solid-element iFEM driven by distributed fiber (Wu et al. 2025) targets the thick, geometrically complex three-dimensional regime where plate kinematics fail — a necessary step toward soft bodies, not a sufficient one. The open problems for soft skins are the ones this module's assumptions expose: large-strain kinematics beyond iFEM's linear theory, extreme sensor–substrate stiffness mismatch (the shadowing effect of Section 7), and elastomer bond hysteresis. None of these is solved in the cited literature; they are the research gap.

Flashcards

References

← 5. iFEM Elements & Formulations 7. Material Identification →