Calibratable Versus Irreducible Variance in Cleanroom-Free Sensors

Cheap ways of making sensors without a cleanroom now work well enough to use. Examples include drawing graphite with a pencil on paper, cutting foil with a craft plotter, extruding conductive grease from a desktop printer, casting composites in 3D-printed molds, and patterning metallized plastic with a UV laser. Their repeatability claims, though, rest on eyeballing curves from about three devices. This proposal asks a narrower question with more practical weight than "are these sensors repeatable?" We ask how much of the device-to-device spread is a stable per-device offset or gain, which a one- or two-point calibration removes for nearly nothing, and how much is drift, hysteresis, or curve-shape difference that no cheap calibration can fix. We answer it with a nested variance-components analysis: an analysis of variance (ANOVA) that splits measurement scatter into contributions from batches, devices within batches, days, and loading cycles. The same crossed protocol is applied to five fabrication routes that span a range of process complexity.

The claim we will test is this. Removing fabrication steps mostly inflates the calibratable part of variance, meaning offset and gain. The non-calibratable part is set mainly by the transducing material and its packaging, not by how carefully it was deposited. If that holds, "low fabrication complexity plus per-device calibration" is a defensible design rule. The decomposition would also tell a lab which single controlling step, such as degassing, precise filler weighing, or controlled curing, is worth adding back, because it shrinks the component that calibration cannot touch. If it fails, the field's "high consistency" claims are masking variance that simplicity itself creates.

Why this question, and what exists

The contradiction motivating the work is already in the literature. The fabrication of electrochemical geophone based on FPCB process technology laser-cuts commercial flexible circuit board layers and reports "nearly identical" responses across three geophones. Low-Cost and Paper-Based Micro-Electromechanical Systems Sensor for the Vibration Monitoring of Shield Cutters draws graphite gauges into laser-engraved grooves and likewise asserts consistency across three prototypes. On the other side, Low-cost fabrication of flexible tactile sensor arrays builds plotter-cut Velostat taxel arrays and reports sensitivity to foil batch, encapsulation thickness, and room temperature. Stretchable Capacitive and Resistive Strain Sensors: Accessible Manufacturing Using Direct Ink Writing attributes poor baseline-resistance repeatability and 21.88% hysteresis in printed resistive gauges to extrusion-rate fluctuation. High-Performance Paper-Based Capacitive Flexible Pressure Sensor and Its Application in Human-Related Measurement sees high-pressure degradation after a month of storage. None of these papers separates batch, device, time, and cycle contributions, so the claims are not comparable. None distinguishes an offset that a user can null from a hysteresis loop that a user cannot.

The obvious idea at this intersection is to run an orthogonal-array experiment over process parameters and rank their main effects by ANOVA. Design and Experiment of Horizontal Transplanter for Sweet Potato Seedlings does exactly this for a seedling transplanter, using an $L_9(3^4)$ array and SPSS ANOVA to rank ribbon spacing above forward speed and screw speed. That design optimizes the mean response and estimates only main effects. It says nothing about whether the remaining scatter is correctable. We deliberately do not propose this Taguchi-style mean optimization, because it is decades-old practice in manufacturing and would add little.

Our conceptual template comes instead from What causes the test error? Going beyond bias-variance via ANOVA. Lin and Dobriban split the variance of a network's test error into main and interaction effects of sample, label noise, and initialization, and found that the damaging variance lives in interactions rather than main effects. The analogue here is that a device's main effect, a constant offset, is harmless once calibrated. A device-by-load interaction is only harmless if it is linear. Device-by-time and cycle-level terms are the ones that hurt. The functional-ANOVA identifiability constraints of Neural Decomposition: Functional ANOVA with Variational Autoencoders motivate how we keep interaction terms from absorbing main effects. We impose zero-mean constraints on device-specific curves, but we do so with a linear mixed model rather than a neural decoder, because our data are tens of devices, not images.

Core model

For log output $y$ of device $d$ in batch $b$, on day $t$, loading cycle $c$, at applied stimulus level $x_k$ (pressure, strain, or acceleration), we fit

$$y_{bdtck} = \mu(x_k) + (\alpha_b + \alpha_{d(b)}) + (\beta_b + \beta_{d(b)})\,x_k + \delta_{d(b)}(x_k) + \gamma_{d(b),t} + \varepsilon_{bdtck}.$$

Each term has a physical reading. $\mu(x)$ is the route's mean response curve. $\alpha$ terms are batch and device offsets, and $\beta$ terms are batch and device gain deviations; together these form the calibratable component that a two-point affine calibration removes. $\delta_{d(b)}(x)$ is a device-specific curve-shape deviation, constrained to be orthogonal to constant and linear functions of $x$. $\gamma_{d(b),t}$ is day-to-day drift of each device after its day-0 calibration. $\varepsilon$ is cycle-to-cycle scatter, which includes hysteresis; hysteresis is also modeled separately via a loading/unloading direction indicator. Working on the log scale addresses the heteroscedasticity flagged for sensor data, where spread grows with signal. Variance components are estimated by restricted maximum likelihood, with cluster bootstrap intervals over devices. The summary quantity for each route is the calibratable fraction, $(\sigma^2_\alpha+\sigma^2_\beta)$ divided by total between-device-plus-temporal variance. Its complement, the irreducible residual in physical units, is the other headline number.

Measurement-rig noise is separated out by running a commercial reference load cell through the identical protocol. Any variance the rig contributes to the reference is subtracted from, and reported alongside, each route's $\varepsilon$.

Research Program

Five routes are fabricated in one lab by two operators. Pencil-graphite gauges on photo paper follow the shield-cutter paper sensor. Plotter-cut aluminum foil around Velostat follows Low-cost fabrication of flexible tactile sensor arrays. Direct-ink-written carbon grease in silicone, in both resistive and capacitive variants, follows Cha et al. Graphene/PDMS cast in FFF-printed molds follows Wireless Readout System for Pressure Monitoring Using FFF-Printed Mold Fabricated Flexible Piezoresistive Sensors. UV-laser-patterned Al/PET follows UV-laser-machined stretchable multi-modal sensor network for soft robot interaction. These span the complexity range in the file, from hand drawing to a single-tool digital process.

For each route we make four batches of five devices, with batches separated by at least a week and each using fresh material lots. Operator is crossed with batch so its effect is estimable. Each device receives 50 load cycles at six stimulus levels on a motorized stage on days 0, 7, and 30, with temperature and humidity logged and entered as covariates. That gives 20 devices and about 18,000 readings per route. The sample size is chosen so the device-level variance component is estimated with roughly ±30% relative precision. This count follows from the chi-square distribution of a variance estimate with about 15 degrees of freedom. It is modest, but it is five to seven times the $N=3$ norm.

The causal arm is a complexity ablation within two routes, mold casting and direct ink writing. In each we add back exactly one controlling step and rerun a reduced two-batch design. For mold casting the candidate steps are vacuum degassing, a 0.1 mg balance for filler, or oven curing. For ink writing they are a positive-displacement extruder instead of a syringe pump, or a humidity-controlled enclosure. The hypothesis predicts in advance which component each step should shrink. Process controls should shrink $\alpha$ and $\beta$, while humidity and encapsulation controls should shrink $\gamma$. These predictions are registered before data collection.

The downstream arm tests whether the decomposition matters for use. For each route we train a simple force regressor on three devices and test it on held-out devices from a different batch. We do this with and without a two-point calibration on the target device, and on day 0 versus day 30. The decomposition should predict the cross-device transfer error before we measure it.

Baselines

The naive reference point is what the cited papers already do: overlay response curves from three devices and report mean ± standard deviation of sensitivity, as in the FPCB geophone and paper vibration studies. We reproduce it on a random three-device subset of our data for each route. This shows how often that practice would have declared "high consistency" for a route whose irreducible variance is actually large. Beating it means finding at least one route where the three-device summary and the full decomposition disagree on the repeatability ranking.

The second baseline is the across-device coefficient of variation of sensitivity and baseline value, computed on all 20 devices. This is the strongest simple summary an engineer would reach for. It uses the same data at zero extra cost. It fails our question only if it ranks routes differently from the irreducible residual, and we will report whether it does.

The strongest competitor is per-device calibration followed by reporting post-calibration residual error on held-out cycles and days. It directly measures what a user experiences, and for the narrow question "is this sensor good enough after calibration?" it may be sufficient. It gets identical data and the identical two-point calibration. The decomposition beats it only if it does something the residual cannot. That something is attribution: correctly predicting, before the ablation runs, which added process step reduces post-calibration error and which only reduces pre-calibration spread. It must also predict which routes degrade by day 30 from day-7 components alone.

The literature-standard parametric baseline is a Taguchi-style orthogonal-array main-effects ANOVA in the manner of Design and Experiment of Horizontal Transplanter for Sweet Potato Seedlings, run on the mold-casting route over mold depth, filler fraction, and cure temperature with a matched run budget. It ranks factors by their effect on mean sensitivity. We count our approach as better if the factor it ranks first is not the one that most reduces irreducible variance.

The key ablation is the same nested ANOVA without the offset/gain/shape split, a plain batch–device–day–residual decomposition. This isolates whether separating calibratable structure adds predictive value for cross-device transfer beyond ordinary variance components.

As reference numbers for route performance, the file gives the following. Cha et al. report gauge factor 16.83 and hysteresis 21.88% for printed resistive gauges, and gauge factor 0.95 with 1.36% hysteresis for capacitive ones. Laser Fabrication and Performance of Flexible Pressure Sensors with Ridge-Mimicking Spatially Ordered Microstructures reports 1.65 kPa⁻¹ sensitivity and under 5% sensitivity error over 20–80 °C. Ham et al. report under 1% trace resistance change under extreme stretch. For between-device or between-batch variance components on any of these routes, no reported result was found on Lacuna. Establishing those numbers is part of the contribution.

Why not just…?

Why not just calibrate every device and report the residual error? This is the serious threat, and we treat it as the main baseline rather than a straw man. If post-calibration residual error alone ranks routes identically and predicts the ablation outcomes as well as the decomposition, then the decomposition is descriptive overhead, and we will say so. Our expectation that it falls short rests on the residual mixing rig noise, hysteresis, and aging into one number. Cha et al. attribute their resistive-gauge problems to extrusion jitter, while Li et al.'s paper capacitor degrades with storage. These are different fixes that a single residual cannot distinguish. The pre-registered ablation predictions settle it.

Why not just run a factorial or orthogonal-array design over process parameters? That approach optimizes the mean, and Design and Experiment of Horizontal Transplanter for Sweet Potato Seedlings shows it estimates main effects from nine runs without testing interactions. It treats each run as one number and so cannot separate device-to-device variance within a recipe from variance between recipes, which is exactly the repeatability question. We include it on one route with a matched budget so the comparison is direct.

Why not just learn the process with a flexible model, such as a Bayesian functional-ANOVA network or a learned controller? Bayesian Neural Networks for Functional ANOVA Model finds fourth-order interactions, but its demonstrations use Madelon-scale data. The file notes it depends on informed MCMC proposals and has long runtimes. Twenty devices per route cannot support that. Closed-loop control of direct ink writing via reinforcement learning cuts deposition-error spread by about 16 μm, but it needs a purpose-built simulator and evaluates bead geometry, not sensor output. We use it as motivation for the extrusion-control ablation rather than as a competitor. As a robustness check, we fit a Gaussian-process version of the device-shape term $\delta$ to confirm the linear mixed model is not missing curvature.

Decisive checks

The central hypothesis fails if, in the complexity ablation, process-control steps shrink the drift or cycle components as much as or more than the offset and gain components. That would mean simplicity creates irreducible variance. It also fails if the calibratable fraction does not rise as routes get simpler. The practical payoff fails if the decomposition's day-7 estimates do not predict day-30 cross-device transfer error better than post-calibration residuals do. If the hypothesis holds, the output is a comparable repeatability number for each route, with a confidence interval, and a tested rule for which control step to buy back first.