Measurement & characterization

How structure and properties are observed: spectroscopy, colorimetry, microscopy, particle sizing, rheometry, thermal analysis, EIS — and why the data is “dirty”.

You cannot read a material’s state the way you read a register. There is no port that returns “the microstructure” or “the degree of cure.” Every property is inferred: you excite a sample with a probe — light, electrons, heat, shear, an AC voltage — record the response, and invert it into a number. This module treats characterization as what it is: a bank of sensors observing one hidden plant, each returning a noisy, partial, indirect projection of the true state.

Measurement as a noisy, partial observation

Fix the frame once and every instrument becomes a special case. Let \(x\) be the true state — composition, morphology, defects, internal stress, cure. No instrument returns \(x\); each returns

\[ y = h(x) + \text{noise}, \]

an output map \(h\) exposing only some coordinates of \(x\), blurred and projected. This is the observability problem of control theory: the state hides behind the sensors. Recovering \(x\) from a set of readings \(\{y\}\) is an inverse problem, almost always under-determined — many distinct states give the same reading (P3, mediation through a hidden state).

Three facts recur. Scope: every technique resolves a finite window in length-scale, time-scale, chemistry, or depth and is blind outside it (P6). Degeneracy: even noise-free, a low-dimensional reading rarely pins a high-dimensional state (P3/P4). Dirt: real readings carry noise, drift, and operator variability. Indirect, partial, noisy — the origin of the “dirty data” this module builds toward.

Spectroscopy: probing chemistry by frequency

Most chemical information arrives as a spectrum: a response across a swept excitation frequency, each technique a different tap on the material’s optical transfer function.

UV-Vis sweeps a wavelength across the ultraviolet-to-visible band (~200–800 nm) and records absorption/reflectance. Absorption dips are notches in the transfer function — position tells you what color, depth how strongly. Transmission mode reads dyes, pigments, and band gaps; diffuse-reflectance handles opaque films and anchors opacity. It cannot say which molecule absorbs, and in a scattering film absorbance mixes concentration, path length, and scattering — separating them needs a model (Kubelka–Munk), so “how much pigment” is itself an inverse problem.

FTIR / IR probes molecular vibrations in the mid-IR (~4000–400 cm⁻¹). Picture each bond as a tiny mass-spring resonator at a characteristic frequency; the spectrum is the system’s resonance comb, acquired all at once by an interferometer and recovered by an inverse Fourier transform (the interferogram-to-spectrum step is the FFT you know). A disappearing epoxide or isocyanate peak quantifies cure; a growing carbonyl peak tracks degradation. The same band (~8–13 µm) is where surfaces radiate heat, so IR emissivity sets radiative-cooling function. ATR mode probes only the top ~µm, so a thin surface contaminant can misrepresent the bulk.

Raman illuminates with one laser tone and reads a faint frequency-shifted sideband — a weak modulation product around a strong carrier — whose shifts fingerprint vibrations. It sees bonds weak in IR (symmetric bonds, \(\ce{C=C}\), inorganic lattices) and works through water and glass. IR and Raman are partly orthogonal projections of the chemical state, so fusing them recovers more of the hidden composition (P3/P9); fluorescence can bury the signal.

NIR (~780–2500 nm) reads broad overtone and combination bands — the harmonics and intermodulation products of the fundamental vibrations. They are broad and low-amplitude: a feature vector you regress against, not a peak you assign. Broad bands penetrate deeper and tolerate turbid samples, making rugged in-line probes at the cost of specificity, so NIR is always paired with a chemometric (PLS/PCA) calibration. Here the measurement is itself a learned inverse model.

Hyperspectral imaging makes every pixel a spectrum — a data cube \(I(x,y,\lambda)\), the bridge between microscopy (where) and spectroscopy (what). It maps chemical/phase distribution across space, and a labeled cube is a dense supervised dataset (P12).

Color: a lossy three-channel projection of a spectrum

Color is the cleanest illustration of the framing, because the loss is exact. The eye is a three-channel sensor: three cone weighting functions over wavelength. Color is the inner product of the reflectance spectrum with three color-matching functions — the projection of an infinite-dimensional signal onto a fixed three-dimensional basis, CIE XYZ. Everything orthogonal to those three is discarded; what survives predicts appearance, not physics. \(L^*a^*b^*\) re-coordinatizes that 3-space to be perceptually uniform, so Euclidean distance approximates perceived difference.

As the figure shows, an entire reflectance curve collapses onto three numbers — and that collapse has consequences.

A spectrophotometer measures the reflectance spectrum feeding colorimetry. Two taps complete the appearance vector: a gloss meter reads the specular spike (the mirror-direction return at a standard angle — how much energy stays coherent rather than scattering), and haze measures the diffuse leakage around it. Partial projections of one surface state (P6/P9): gloss at 60° can look perfect while grazing 85° reveals orange-peel, so the scope here is literally an angle.

Microscopy and imaging: resolution versus representativeness

To see structure you sample space, and every imaging tool trades how finely it resolves against how much it represents — the dominant axis here (P6).

Optical and confocal microscopy spatially sample in visible light, diffraction-limited to ~λ/2 (a few hundred nm) — a hard aperture/Nyquist ceiling no pixel count beats. Confocal adds a pinhole rejecting out-of-focus light, building depth-resolved z-stacks. What looks uniform optically can be wildly heterogeneous at the controlling scale.

SEM swaps photons for electrons (shorter wavelength, finer limit) and raster-scans the beam like a CRT, mapping secondary-electron intensity to pixel. It excels at surface morphology (particle shape, texture, fracture, pores) and with EDS adds elemental maps — but it is mostly a 2-D surface projection needing vacuum and a conductive coating.

TEM sends electrons through a <~100 nm slice — an electron shadowgraph resolving internal structure to near-atomic scale (lattice fringes, nanofiller dispersion, phase boundaries). It is the deepest look over the least representative volume: generalizing from a few nm-scale slices to a whole batch is a severe sampling leap.

AFM drags or taps a sharp tip across a surface; cantilever deflection (sensed by a reflected laser) feeds a control loop tracing the height field. Tapping/force modes also probe local stiffness and adhesion — a mechanical impedance per point — so AFM is a rare technique reporting both structure and a property at the same scale (P9 at the nanoscale).

Particles, rheology, and impedance: inverting the response

Disperse-phase and mechanical characterization are nearly pure inverse problems: you never see the particles or the network, you see a signal they produce and invert it.

Particle sizing infers a size distribution from scattering or settling. DLS watches scattered-intensity fluctuations — small particles jiggle faster, so the autocorrelation decay time encodes size (sub-µm, as hydrodynamic diameter). Laser diffraction inverts a static angular scattering pattern over a wide µm range. Sedimentation times settling via Stokes’ law. All assume a shape/optical model (usually spheres) and return a distribution through an ill-conditioned inversion (worst for broad/multimodal samples), so priors do real work — DLS is the exemplar under-determined inverse problem (P3). Because DLS is intensity-weighted, a few large particles dominate and hide the population you care about: never read one mean diameter as “the size.”

Zeta-potential applies a field, watches charged particles drift (electrophoresis by laser Doppler), and backs out the slip-plane potential as a proxy for colloidal stability — high magnitude resists clumping, near zero aggregates. An indirect, model-laden (Smoluchowski) readout, a measured proxy for a hidden state (P3/P9).

Rheometry is system identification of a mechanical network. Rotational mode is a step/ramp — steady shear in, stress out — giving viscosity versus shear rate, a nonlinear DC gain that reveals shear-thinning. Oscillatory mode is a frequency sweep: a small sinusoidal strain yields a complex response split into storage (elastic) \(G'\) and loss (viscous) \(G''\), with

\[ \tan\delta = \frac{G''}{G'}, \]

literally a Bode plot of a mechanical impedance. It characterizes sag, leveling, sprayability, and gelation; at the gel point \(G'\) crosses \(G''\) — a canonical marker of a state transition (P13). It is protocol-dependent: a viscosity without its shear rate and history is nearly meaningless.

EIS (electrochemical impedance spectroscopy) is the flagship of “measurement = impedance spectroscopy,” and you already know it. Apply a small AC excitation across a coated metal in electrolyte, sweep frequency, record the complex impedance \(Z(\omega)\); plot Bode (\(|Z|\), phase) or Nyquist (\(-\operatorname{Im} Z\) vs. \(\operatorname{Re} Z\)). The metal behaves as an equivalent circuit: an intact film is a high-resistance capacitive barrier (large \(|Z|\) at low frequency); as electrolyte penetrates, resistance drops and new RC elements (interfacial corrosion) appear. You fit a circuit exactly as in EE and watch it drift — coating capacitance reports water uptake, pore/charge-transfer resistance reports degradation and delamination (P3/P6).

Thermal, geometric, and durability probes

The rest ramp a control input or stress the sample and watch a response — the same observe-the-output idea in other domains.

DSC ramps temperature and records heat flow against a reference — a thermal transfer-function sweep. A baseline shift marks the glass transition \(T_g\), an endotherm marks melting, an exotherm marks cure (onset, peak, enthalpy → degree of cure). It measures energetics, not identity; \(T_g\) shifts with heating rate and history. TGA runs the same ramp with mass as output: each weight-loss step is a process (solvent loss, decomposition, oxidation), and the high-temperature residue estimates inorganic filler/ash content — a compositional projection (P3). Thermal/IR cameras image the ~8–14 µm band, converting per-pixel intensity to temperature via Planck’s law and emissivity, closing the loop from a spectral property (IR emissivity, from FTIR) to delivered function (a cooler surface under sun); wrong emissivity causes large errors.

Film geometry uses several taps: eddy-current/magnetic gauges sense coating-to-substrate spacing as an impedance readout (non-destructive); cross-sectioning images the film directly (destructive ground truth); ellipsometry inverts the polarization change of reflected light into nm-scale thickness and refractive index (a multi-solution inversion — wrong optical model, wrong thickness); profilometry traces height for roughness. The recurring trade is non-destructive (inferred) versus destructive (direct) (P6). Contact-angle imaging fits the droplet angle — low spreads (hydrophilic), high beads (hydrophobic) — a geometric proxy for surface energy (P9), badly misled by roughness and advancing/receding hysteresis.

Accelerated weathering and salt-spray are burn-in tests: QUV/xenon-arc cycle UV + heat + moisture, salt-spray (ASTM B117) holds samples in hot salt fog. They over-drive the stressor to extrapolate long-time behavior, yielding relative durability rankings (gloss loss, \(\Delta E\), chalking, blistering, rust creep) — long-horizon dirty proxies for a hidden degradation trajectory (P3) with an uncertain acceleration factor (multi-fidelity, P12). Chamber hours are not field years without correlation data.

Dirty data: fusing partial observations into a state

Pull back to the whole stack. Every readout above is a partial projection of one hidden state through a noisy channel; across many samples, instruments, and days the result is structurally messy, and modeling must treat the mess as part of the signal.

The deepest issue is informational, not statistical: each technique resolves only a scope (P6) and returns a projection of the true structure (P3), so even noise-free no single measurement determines the state. Two hard constraints sharpen it — destructive vs. non-destructive (trade ground-truth fidelity for keeping the specimen) and the resolution↔︎representativeness trade (TEM sees atoms over a vanishing volume; an in-line NIR probe sees a huge volume vaguely). This is why the field needs learned, probabilistic, multi-fidelity models, not lookup tables.

Recap

• Every characterization is an output map \(y = h(x) + \text{noise}\); reading the hidden state \(x\) back out is an under-determined inverse problem (P3).
• Spectroscopies are frequency sweeps — UV-Vis notches, FTIR’s resonance comb via inverse FFT, Raman sidebands, NIR as learned regression, hyperspectral as a data cube — different taps on one transfer function.
• Color is a lossy three-channel projection onto the CIE basis, and metamerism is aliasing: spectra collapse into the null space of the color-matching functions, breaking when the illuminant changes the basis (P9, P3/P4).
• Microscopy trades resolution for representativeness (optical → SEM → TEM); rheometry and EIS are system identification, Bode/Nyquist plots whose crossover or fitted circuit stands in for the hidden state (P9, P13).
• Dirty data is first-class: a sensor bank with noise, drift, domain shift, and missingness over a ragged matrix. Fusing these scope-limited projections is sheaf-theoretic data fusion (math:08) — P6/RD6 — never the reading of one trusted dial.