Perceptual Tolerances for 3D-Printed Resonators

Manufactured parts never match their digital blueprints exactly. For parts that make sound, such as bells, chimes, plates, and speaker housings, engineers usually judge whether a deviation matters with objective numbers: how far a resonant frequency moved, or how much the frequency response changed. Work on audio evaluation shows that such numbers often disagree with what listeners hear. They penalize changes nobody notices and miss artifacts that are obvious. This note proposes measuring manufacturing tolerance in perceptual units. The question becomes what fraction of real printed parts a listener cannot tell apart from the intended design, and which geometric deviations actually cause audible failure. We would answer this for 3D-printed stainless-steel plates and cylinders. The deviations would be generated by a statistical model fitted to laser scans of real printed parts, not by made-up noise.

We also make one specific, falsifiable bet. Metal printing leaves direction-dependent ripples. On symmetric resonators, these ripples should split pairs of vibration modes that ideally share one frequency. The result is an audible slow beating, the "warble" bell founders fight. Frequency-error metrics and spectral distances barely register it. If this is right, objective tolerance analysis will misjudge which parts pass and which design choices are robust, and the misjudgment will be systematic. A perceptual criterion calibrated with listeners would correct it.

Core Claim

For resonators whose modes are closely spaced or come in near-degenerate pairs, we claim three things. First, ranking as-built parts by objective acoustic error differs substantially from ranking them by human discriminability. Second, the disagreement is concentrated in symmetry-breaking deviations that isotropic noise models under-produce. Third, choosing a nominal design to maximize perceptual yield gives different, and audibly better, designs than maximizing objective yield. Perceptual yield is the probability that a printed part is indistinguishable from the nominal one.

Write $d$ for a nominal design, $g \sim p(g \mid d)$ for an as-built geometry drawn from the fitted manufacturing model, $s(\cdot)$ for the simulated strike sound of a geometry, and $\pi(a,b)$ for the probability that a listener hears sounds $a$ and $b$ as different. Perceptual yield is then

$$ Y_p(d) = \Pr_{g \sim p(g\mid d)}\big[\, \pi(s(g), s(d)) < \tfrac{1}{2} + \delta \,\big], $$

where $\delta$ is a small margin above chance that defines "not noticeably different." Objective yield replaces $\pi$ with a thresholded objective error, such as the maximum modal frequency shift in cents or the log-spectral distance.

Why This Intersection, and What Is New

Three lines of work already close the gap between design and manufacturing by optimizing under fabrication uncertainty:

These papers supply the method we reuse: sample as-built geometries and optimize expected performance. All three score outcomes with signal-fidelity or physical objectives (PSNR, SSIM, spot size, lift-to-drag). GAN-DUF's perturbations are also simulated jitter, not measured variation. None asks whether a human would perceive the degradation.

On the perception side, A Differentiable Perceptual Audio Metric Learned from Just Noticeable Differences measured listeners' "same or different" thresholds along perturbation directions in signal space, from about 55,000 judgments. ViSQOL v3 showed that full-reference metrics need explicit alignment fixes when sounds are perceptually fine but structurally different. These papers motivate the proposal: objective and perceptual scores diverge. They also constrain it. Their perturbations are digital (noise, reverb, codecs), so we cannot assume their learned metrics transfer to physically generated timbre changes. That is why we collect our own thresholds.

The new ingredient is to measure listener discrimination directly along manufacturing-generated deviation directions, and to use that measurement to define tolerances.

The deviation model comes from A Statistical Approach to Surface Metrology for 3D-Printed Stainless Steel. It fitted random-field models to laser scans of six printed 316L panels (about 300 mm square, 3.5 mm thick) and transferred the model to a held-out printed cylinder (about 170 mm diameter). Model selection by AIC chose an anisotropic model with smooth oscillatory noise, $M^*$, over isotropic alternatives. The same paper reports that its Gaussian assumption fails on the thin-wall tail of the thickness distribution. We treat that as a known weakness to probe, not hide. These exact geometries, thin plates and a thin-walled cylinder, are resonators, which is what makes the acoustic question tractable with existing data.

Research Program

Stage 1: acoustic simulation of printed parts. We draw thickness fields from $M^*$ on plate and cylinder geometries and compute vibration modes with shell finite elements that use a spatially varying thickness. We then render struck-object sounds by modal synthesis, which sums decaying sinusoids at each mode's frequency, damping, and strike-point amplitude. The nominal geometry gives the reference sound $s(d)$. Because every sample shares a strike position and onset, the reference is aligned by construction. This removes the alignment penalty ViSQOL had to engineer around and makes full-reference comparison fair.

Stage 2: listening thresholds in manufacturing space. Following the active-learning design of the JND paper, listeners hear pairs (nominal against as-built, and as-built against as-built) and answer "same or different." Queries are chosen to locate the threshold along directions in the space of deviation fields. We parameterize those directions by the low-order structure of $M^*$ samples, namely amplitude and orientation of the anisotropic component, and fit a model $\hat\pi$ that maps a pair of sounds to a discrimination probability. We budget on the order of several thousand judgments, far fewer than 55,000, because the perturbation space is low-dimensional and structured.

Stage 3: divergence analysis. For each design family we compare three yield estimates: from humans (on a held-out sample set), from $\hat\pi$, and from each objective metric. We report the disagreement and attribute it to physical causes. We compute mode-pair splitting explicitly so we can test whether the disagreement concentrates there.

Stage 4: design for perceptual yield. We optimize a small set of nominal parameters (wall thickness, taper, plate aspect ratio, and the print orientation relative to the geometry's symmetry axes) for $Y_p$ versus objective yield, with the same sampling budget as GAN-DUF-style robust design. The key comparison is whether the two optima differ, and whether listeners prefer the batches produced by the perceptual optimum. If budget permits, a small physical batch of printed cylinders would be recorded and compared against simulation. Without it, the claims stay explicitly simulation-grounded.

Baselines

Nominal yield (do nothing). Assume every part matches CAD, so yield is 100%. This is the reference point. Every other estimate is judged by how far it moves from this toward the human-measured yield.

Objective worst-case and statistical tolerance analysis. This is what an acoustics engineer would actually do. A part passes if every modal frequency stays within a threshold, set in cents, and the log-spectral distance stays below a threshold. We tune both thresholds to maximize agreement with human labels on a calibration split and then freeze them for the test split. This gives the objective metric its best possible chance. Beating it means lower yield-estimation error against human labels on held-out designs, and correct ordering of designs by perceptual yield where the objective rule orders them wrongly.

ViSQOL v3 used off the shelf. We score the as-built sound against the nominal reference with ViSQOL v3, in audio mode, with a pass threshold calibrated the same way. This is the strongest existing learned or engineered perceptual metric in the file that applies without new data. No reported result found on Lacuna for ViSQOL, or any metric, on manufactured-resonator discrimination. Its published numbers are codec and network MOS values (for example, 4.21–4.28 under good network conditions), which are not comparable to our task.

The JND-trained metric of Manocha et al. We apply it zero-shot. Its reported evidence is correlation gains over PESQ and ViSQOL, plus a human preference rate above 0.50 in denoising. It tests whether signal-space thresholds transfer to physical timbre changes.

Ablations.

  • Isotropic noise. Replace $M^*$ samples with isotropic Gaussian thickness noise of matched variance. This isolates whether directional printing variation drives the divergence.
  • Analytic threshold. Replace $\hat\pi$ with an analytic modal-splitting threshold, a beat-rate and amplitude rule derived from psychoacoustic beating. This tests whether a hand-built rule captures what the learned model captures.

All methods see identical sample sets, simulations, and calibration labels.

Why not just…?

Why not just use modal frequency error with a pitch threshold, as instrument makers do? This is the most serious threat. For a bar or plate dominated by one isolated mode, it may be enough, and we expect it to be. The proposal therefore includes such geometries as a control, where we predict no advantage. The claimed divergence is specific to near-degenerate mode pairs. There a split of a few cents can be invisible to pitch-error rules yet produce clearly audible beating. If the tuned frequency-threshold baseline matches human yield on cylinders as well, the core claim is falsified, and the note reduces to "objective tolerance analysis is adequate for printed resonators." That is itself a useful negative result.

Why not just use ViSQOL, since the references are aligned? Alignment was the main failure mode ViSQOL v3 had to patch, and our setup removes it, so ViSQOL may do well. Its mapping to quality was calibrated on speech and music degradations such as codecs and packet loss, not slow beating envelopes in decaying tones, and nothing in the file tests it there. The held-out comparison against human labels settles this directly. If ViSQOL matches $\hat\pi$, we would recommend it and drop the learned model.

Why not just model deviations as isotropic Gaussian noise on thickness? Oates et al. rejected isotropic models by AIC on real scans. Isotropic fields also break rotational symmetry less coherently than directional layer ripples, so they should under-produce mode splitting. The isotropic ablation tests whether the choice changes estimated yield and the optimal print orientation. If it does not, the simpler noise model suffices for this purpose.

Why not just ask an audio-language model to judge? CMI-RewardBench reports that Gemini 3 Pro reaches 65.80% agreement on fine-grained music judgments, against 78.20% for a model trained on expert preferences. The JND paper found classification-pretrained features worse than perceptually trained ones for discrimination. Near-threshold same-or-different judgment is the hardest version of this task, so we do not treat a zero-shot judge as a main baseline.

Proposed Experiments and Decisive Checks

The benchmark consists of simulated plates and cylinders at the dimensions of the scanned specimens in Oates et al., with thickness fields from $M^*$ and from the isotropic ablation. Human labels are split by design family into calibration and test sets, with listener identity held out.

The first decisive check is yield-estimation error against human labels on cylinders, compared with plates dominated by an isolated mode. The prediction is that $\hat\pi$ beats the tuned frequency-threshold baseline and ViSQOL on cylinders by a clear margin, with bootstrap intervals that exclude zero, and ties on plates.

The second check is attribution. Among parts that the objective rule passes but listeners reject, most should show mode-pair splits that produce beat rates in the audible range. The same pattern should appear far less often under isotropic noise.

The third check tests design. The perceptual-yield optimum and the objective-yield optimum should differ in print orientation or wall thickness. In a blind listening test, parts sampled from the perceptual-optimum batch should be judged indistinguishable from nominal more often than parts from the objective-optimum batch.

Any of the following would count against the proposal:

  • no gap on cylinders;
  • no concentration of the gap in mode splitting;
  • identical optima under the two yield definitions.