Open hardware research · Patent pending

Computing withthe shape of sound

Coherent Wave Memory stores information in the vibrational fingerprint of a piece of glass and reads it back through wave interference. Measured on a real bench, published with every failure, and built so you can replicate it for under $100.

See How It Works Enter the Lab

Scroll

Measured on the bench — not projected

100%

Fingerprint classification (80/80 trials)

4,096

Error-free states in one plate (12 bits)

56 dB

Peak mode signal-to-noise

16.5M

Cycles at 0.22% drift

All figures from the current validation paper (v19r, June 2026). Read it →

The physics, step by step

No equations required — every demo below runs the same math we use in the lab.

Everything rings

Tap a wine glass and it sings at one pitch. A flat glass plate is richer: one tap rings dozens of distinct patterns at once, each at its own exact frequency. These patterns — eigenmodes — are set by the plate's geometry and atomic structure. They are as repeatable as the laws of physics, because they are the laws of physics.

Watch the tap ring the plate, then pick any peak in the spectrum — the four labelled frequencies are real measurements from our bench.

mode pattern · 35,840 Hz — measured

auto-playing

pale lines = still · teal = vibrating

the same tap rings all of these at once — click a peak to see its pattern

The ring is a fingerprint

No two plates ring exactly alike. Microscopic differences from manufacturing shift each mode by a hair — and the full set of shifts forms a spectral fingerprint unique to that piece of glass.

We classified these fingerprints with 100% accuracy across 80/80 trials, with the closest wrong answer 193 standard deviations away. The fingerprint also barely moves: 0.22% drift over 16.5 million cycles. That combination — unique and stable — is the foundation for unclonable hardware keys.

Same glass, same drive signal — different fingerprint. Identity from physics, not serial numbers.

Mass writes, interference reads

Stick a tiny weight on the plate and every mode that moves at that spot slows down slightly — a rule worked out by Lord Rayleigh in 1877. Different positions slow different modes by different amounts, so the pattern of shifts encodes where the mass is and how heavy it is. Writing without electricity. Try it below.

Write a pattern with mass, read it in frequency

Δf/f = −Σ δmᵢ·φ²(xᵢ,yᵢ)/2M

click anywhere to place a putty dot

Mass per dot25 mg

35.8k

—

54.9k

—

57.0k

—

97.0k

—

Each dot slows every mode in proportion to the mode's local vibration amplitude at that spot — and the shifts add up. A multi-dot pattern produces a shift signature no single dot can mimic: that's a written word, not a written letter. Gray bars: bare plate. This is the WRITE mechanism we are validating on the bench (experiment E3).

Shrink it, and it computes

Here is the honest part: at desk scale, the plate stores and identifies — but it does not compute. Its vibrations fade in about a millisecond, and we can only listen at two points. Both limits are engineering, not physics.

Shrink the plate to one millimeter on a chip, seal it in vacuum, and the same equations flip in our favor: vibrations persist long enough to process sequences, and on-chip transducer arrays listen at sixteen points instead of two. That is the device on our roadmap — every prediction anchored to a number we measured at desk scale.

† projected from measured Q and validated scaling laws — not yet a device

We publish our failures

Falsification-first research: every hypothesis gets a kill criterion before it gets bench time. The kill list is why you can trust the rest.

✓ Survived validation

•Signal is acoustic — three independent null tests (0% electrical feedthrough)
•100% fingerprint classification, 193σ separation (80/80 trials)
•4,096 error-free states from 4 modes × 8 amplitude levels
•Classical non-separability: CHSH S = 2.73–2.83 across 5 mode pairs
•0.22% spectral drift over 16.5 million drive cycles

✗ Killed and documented

•Ferrofluid substrates — phase diffusion destroys coherence
•Reservoir computing at desk scale — vibrations fade 100× too fast
•"The plate computes" — at macro scale, decoders compute; the plate transforms
•Phase as a data channel at bench — accuracy collapsed to near-chance
•Audio-interface capture — loses the information a scope preserves

36 of 87 modeled extensions killed, mechanisms documented. Full ledger in the papers →

Where this is going

Four stops on one continuum — from the plate on our bench to the quantum end of the same physics. Each label tells you what is measured, what is projected, and what is the frontier.

Today · measured

A fingerprint written in glass

On our bench right now: a fused-silica plate whose vibrational spectrum identifies it perfectly, stores 12 bits with zero error, and exhibits the same non-separability mathematics that powers quantum information — at room temperature, for $38.

100%

classification, 80/80

4,096

error-free states

S = 2.83

CHSH non-separability

Next · projected from measured scaling laws

The millimeter machine

Shrink the plate 100×, seal it in vacuum, and the physics flips from limitation to leverage: vibrations persist long enough to process sequences, and a 16-element transducer array reads the full richness of the wave field. Memory, sensing, and signal processing in a chip with no transistors in the data path.

Q ≥ 10⁴

vacuum-packaged silica

channels vs 2 today

~µW

operating power

The frontier · gated on measured thresholds

Glass that computes

Pump each mode near its parametric threshold and it becomes a coin that physics flips: phase 0 or π — an Ising spin. Couple the modes and the network settles into the lowest-energy answer to optimization problems, the way optical Ising machines already do with kilometers of fiber. Ours would be a grain of glass: key, memory, and processor in one unclonable die.

0 / π

phase-bistable spins

MaxCut

native problem class

1 die

PUF + memory + solver

The horizon · established physics, our bridge

The quantum continuum

This is not a metaphor: cooled to 10 millikelvin, high-Q acoustic resonators are real quantum objects — single phonons have been coupled to superconducting qubits since 2017. Our architecture lives on that continuum. Every degree of structure we validate at 300 K is a design rule for the hybrid acoustic processors at the cold end of the dial.

ħω ≈ kT

the crossover we ride

2017

qubit–phonon coupling (Science)

architecture, both regimes

scroll to travel the roadmap ↓

The full roadmap — every phase, kill criterion, and bill of materials — is public on GitHub →

Why it matters

Computing's bill is coming due.
Glass doesn't charge interest.

Every digital operation moves electrons through resistance and pays in heat — data centers already consume more power than most countries. A wave-based substrate spends almost nothing: the medium itself holds the state, transforms the signal, and — at the frontier — settles into answers. Here is what that buys, from nearest to farthest:

within reach · 2026

Hardware identity that cannot be counterfeited

Every plate's spectral fingerprint is born in manufacturing randomness and stable for millions of cycles. Chips, passports, medical devices, and supply chains get keys that physics itself refuses to clone — no stored secret to steal.

How the architecture gets there

Each plate rings differently

measured

100% classification across 80/80 trials, closest wrong answer 193σ away — measured on our bench.

And keeps ringing the same way

measured

0.22% drift over 16.5 million cycles. The fingerprint is a property of the glass, not the day.

Randomness no one can dial in

physics

The fingerprint comes from atomic-scale manufacturing variance — to clone it you would have to place defects atom by atom.

A key that is also the lock

projected

Challenge the device with a drive signal; only the authentic glass returns the right spectrum. No stored secret exists to be stolen.

Multi-device uniqueness study (5 plates) is Phase B of the public roadmap.

click to see how →

within reach · 2026

Sensors that hear where and how much at once

One plate, many modes: a single drop of mass shifts each mode differently, encoding position and weight in one measurement. Air-quality monitors, structural health patches, lab assays — sensing rich enough to replace arrays with a single piece of glass.

How the architecture gets there

Rayleigh, 1877

physics

Added mass slows each mode in proportion to how much that mode moves at that exact spot. 150-year-old, never-broken law.

Four modes, four different ears

measured

Our plate’s modes have distinct spatial patterns — the same dot of mass shifts each one by a different, predictable amount.

Invert the shift pattern

projected

Four shifts form a vector; solving backwards recovers both mass and position from one electrical measurement — experiment E3 on the bench now.

One plate replaces an array

projected

Where today you tile many single-point sensors, one resonator covers the surface — fewer parts, lower power, richer data.

Position-inference mapping (5×5 grid) is scheduled in Phase B.

click to see how →

the chip · 2027+

Edge intelligence on a micro-watt budget

A millimeter resonator that preprocesses signals as they arrive — hearing aids that run for months, implants that never need charging, satellites that think with the power of a wristwatch. The data path has no transistors to feed.

How the architecture gets there

The plate transforms signals for free

measured

Drive it and the response is a rich, fixed mixing of the input — a physical feature map. Energy cost: the drive tone itself.

Bench honestly fails at sequences

measured

Vibrations fade in ~1–4 ms but our desk electronics update 100× too slowly. We published this failure — it is a speed limit, not a physics limit.

Shrinking flips the inequality

physics

τ = Q/πf. At 1 mm and vacuum Q ≥ 10⁴, echoes persist ~300 µs while symbols arrive every ~100 µs — the memory the bench lacked.

Preprocessing before the processor

projected

Audio, vibration, and RF features computed in glass; the digital chip wakes only to read answers. Months of battery, not days.

MEMS design study with full error budgets is Phase D, gate G5 of the roadmap.

click to see how →

the frontier

Optimization without the megawatts

Logistics, drug discovery, and grid balancing all reduce to the same problem class our parametric spin network would solve natively — the way water finds the lowest valley. Quantum-inspired answers at room temperature, in a grain of glass.

How the architecture gets there

A pumped mode picks a phase

physics

Drive a mode at twice its frequency and it oscillates locked at phase 0 or π — a coin physics flips. Demonstrated in optics and MEMS for decades.

Coins that feel each other

physics

Coupled modes prefer agreeing or disagreeing — exactly an Ising spin system, the native language of optimization problems.

The network settles into the answer

projected

NTT’s optical Ising machine proves the principle with km of fiber. Ours would be a grain of glass at micro-watts — gated on crossing the parametric threshold at Q = 10⁴.

The structure is already there

measured

Our CHSH result (S = 2.83) shows the plate’s modes share the non-separable mathematical structure these machines exploit.

Parametric threshold search is gate G7 — the final gate of the public roadmap.

click to see how →

Partner with the research

The next experiment costs $300.
The MEMS chip needs a partner.

Everything on this page — the data, the failures, the roadmap, the patent filing — is open for inspection. We're looking for fabrication partners, research collaborators, and investors who want in before the millimeter machine exists.

Inspect the Evidence

U.S. Provisional Patent No. 64/023,264 · All data and code public on GitHub

Build one. Measure it. Make it science.

Every plate rings differently — which means every device you build adds a data point no lab can produce alone. Your measurements feed the public dataset behind the next paper.

🔬