Three ways to build it

The dose definition we are building to

Version 0, derived from the evidence. Every value has a tolerance and a measurement rule, because a number without a measurement rule is not reproducible.

The three architectures compared

A · Adapted commercial bath

Buy a laboratory ultrasonic bath whose nominal frequency is near 33 kHz, keep its generator, and add three things: a gating controller on the enable line, thermal control, and acoustic verification.

What you can actually buy

No EU or US vendor sells 32.2–33 kHz off the shelf. 33 kHz is a mass-produced Asian frequency — it is the resonance of the standard 60 W bolted Langevin with a ~48 mm face. Western labs cluster at 35 kHz (Bandelin), 37 kHz (Elma) and 40 kHz (Branson); Crest has 30 kHz push-pull; Guyson offers custom frequencies on request.

Hardware to add

An opto-isolated relay or solid-state switch on the generator's enable/remote input (never switch the mains — the microcontroller reboots every 3 s); aquarium heater plus PID thermostat; two independent thermometers; hydrophone; oscilloscope; acoustic absorber lining.

Software

A small ESP32/RP2040 firmware: pulse timing, session timer, temperature interlock, CSV session log over USB or Wi-Fi. No power control — that stays on the bath's own panel, and its setting becomes part of the calibration record.

Verdict: good first bench, not a good reference instrument. Use it to learn the calibration workflow while B is being built.

B · Open driver + Langevin transducer recommended reference build

A bolt-clamped (Langevin) piezo transducer of the ultrasonic-cleaning family, driven by our own low-power stage under microcontroller control, in a tank we define. The drive electronics are trivial in power terms — reaching 4–8 kPa in a few litres of water needs watts, not hundreds of watts — which is exactly why building our own is easier here than in most ultrasonic projects.

What to buy — verified parts and prices

Rest of the build

RP2040/STM32/ESP32 controller (integer PWM division lands within ±3 Hz of 32.248 kHz, which is well inside the transducer bandwidth); gate driver UCC27714 or IR2184, whose shutdown pin doubles as the hardware 1.5 s gate; low-voltage MOSFETs on a 24–48 V bus — there is no reason to put 300–400 V on a bench for a 42 mW experiment; tank from a 4 L beaker to a foot bath with absorber lining; heater and PID in an outer bath, so no moving water in the acoustic path.

Software

Firmware: DDS or PWM frequency synthesis, optional sweep, precise gating, resonance tracking, amplitude regulation against the calibration table, thermal and cavitation interlocks, CSV logging. Host tools in Python: field-mapping over a 3-axis gantry, spectrum analysis for cavitation, calibration-file generation, and export of a session log in the schema used by biohack.it.

Verdict: this is the build that makes the project reproducible, and it is what the builder kit will document.

C · Multi-channel array tub

The human-scale configuration, mirroring the photographs of the UT device: many individually driven transducers around a tub, sequenced and swept, so the field is built from many weak overlapping sources rather than one strong one.

Hardware

8–16 potted transducers on arms or walls; one B-stage driver per channel, galvanically isolated, with per-channel current monitoring; a SELV supply where possible, otherwise a medical-grade isolation transformer; RCD at 10 mA plus an insulation-monitoring device; a neutral-generator array topology; an all-plastic tub with no earthed metal in the water; redundant temperature sensors; emergency stop; a single-board computer coordinating the channel MCUs.

Software

Channel scheduler (group sequencing, phase and sweep), interlocks and watchdogs, leakage and temperature monitoring with automatic shutdown, per-position calibration tables, full session logging, and an operator checklist enforced by the UI before a session can start. A separate interlock microcontroller that can only cut power, so that a bug in the drive host cannot defeat the safety layer.

The acoustics of a tub are different from a beaker

• A 250 L tub has ~11,000 modes below 33 kHz — about one per hertz. It is a diffuse, statistical field, not a modal one: you cannot tune the standing waves out, you can only average over them. In such a field ~2% of the volume exceeds twice the nominal pressure.
• Decorrelation length is about λ/2 ≈ 2.3 cm, so asking the subject to shift position every couple of minutes is a free and effective averaging mechanism.
• De-synchronise deliberately: give each channel a small frequency offset (7–40 Hz), re-draw phases every 10–20 ms, place transducers aperiodically. Twelve incoherent sources plus sweeping can cut the spatial spread of time-averaged intensity from ~100% to a few per cent — but this must be verified by mapping, not assumed. Water circulation does not homogenise an acoustic field.
• The body is nearly transparent at 33 kHz (about 0.3 dB through a torso). This is whole-body exposure, not surface exposure, and the only strong internal reflectors are gas bodies — lung and bowel. Note that the mouse protocol immersed only the lower half of the animal.
• An acrylic tub is preferable to steel or glass: the plastic absorbs, which damps the field — and you have drive power to spare.
• You cannot properly degas 250 litres. Vacuum needs a sealed vessel, boiling is absurd, and ultrasonic degassing works by cavitating. The honest options are a membrane contactor on a recirculation loop (€600–2,000, ~8 h) or filling hot at 50–55 °C and letting it cool to 36 °C, which leaves the water undersaturated. Since the entire cavitation safety case rests on the absence of nuclei, this is a first-order unresolved problem, not a detail.
• Drain and refill every session: for a couple of euros of water it solves sanitation, dissolved gas, conductivity and corrosion at once. Do not use ozone (it aerates) and prefer no halogen sanitiser (it raises conductivity and pits stainless).
• Best thermal answer: no in-tub heater at all — fill blended from the domestic hot water. A covered 250 L tub drifts only 1–2 °C in 30 minutes, and this deletes the largest electrical hazard in the system.

Rough bill of materials, 12 channels

The cost distribution is the real finding: transducers and drive electronics are about 15% of a defensible build. Metrology and electrical safety dominate, and they are exactly the parts nobody can skip — a build with €500 of transducers and no calibrated hydrophone cannot tell 4 kPa from 400 kPa.

Verdict: last, not first. It requires A or B to be verified, an independent electrical check, and a written safety case.

Power budget: this is a watts machine, not a kilowatt machine

The single most liberating number in this project. Working from plane-wave intensity and the vessel cross-sections:

Consequences that shape every architecture:

• A SELV design becomes feasible (≤30 V DC), which is the single biggest safety win available for a device that puts electricity near an immersed person.
• Acoustic heating is irrelevant: about 0.004 °C over a whole session in a full tub. The water heater, not the ultrasound, is the thermal system.
• The drive is mostly reactive: a piezo ring of a few nanofarads at 33 kHz and 200 V peak-to-peak draws tens of milliamps, a couple of volt-amperes. A small piezo driver is the right class of part, not an RF power amplifier.
• The published "135–275 V" is consistent with exactly this: a modest, high-voltage, low-current drive.

Acoustic safety: the real limits at 33 kHz

Most of the reassuring numbers people reach for are defined for megahertz diagnostic ultrasound and do not apply here. What the physics and the standards actually say:

Calibration: the shared bottleneck

A professional needle-hydrophone system of the type used in the original work costs more than everything else in this project put together. The way out is not to skip calibration, but to share it.

1. One reference measurement. Get access to a calibrated hydrophone at a university acoustics or biomedical-engineering lab (or a metrology institute) for a single session.
2. Transfer calibration. Build a simple PZT hydrophone, measure it side by side with the reference at 33 kHz, and derive its sensitivity in V/Pa with an uncertainty figure.
3. Circulate. Keep one transfer hydrophone as the project's travelling reference; each new builder either borrows it or cross-calibrates their own against it.
4. Map, don't spot-check. A converted 3-D printer makes a perfectly good 3-axis traverse for a field map on a 1 cm grid.
5. Publish the map with every build, alongside the drive settings that produced it. A build without a published map does not contribute biological data.

Which hydrophone, and the counter-intuitive rule

Where a 33 kHz calibration actually exists. Queried against the international database of recognised calibration capabilities, only two national labs in the world hold one at 32 kHz: NPL in the UK (2 kHz–500 kHz, 0.5 dB) and VNIIFTRI in Russia. PTB, NIST, INRiM and the rest all start at 0.5–1 MHz because they are medical-ultrasound oriented. Commercially, Onda calibrates from 30 Hz upward. The practical path is one calibrated passive reference, then comparison-calibrate DIY units in your own tank. An uncalibrated DIY hydrophone carries ±3–6 dB of uncertainty — and since our target window (4 → 8 kPa) is only 6 dB wide, that is not good enough on its own.

Reading it. A 12-bit bench scope (Siglent SDS804X HD, ≈€409) for field mapping, plus a scriptable USB instrument (Digilent Analog Discovery 3, $379, or PicoScope 2206B, €463) for automated cavitation snapshots. Avoid 8-bit scopes for cavitation work, and never use a 48 kS/s DAQ — 32.2 kHz would alias to 15.8 kHz.

Degassing, cheaply. Henry's law means you only need about 0.2 atm, and water vapour sets the floor, so a $136 vacuum chamber with a single-stage pump is entirely sufficient — a two-stage pump buys nothing. Boiling for 5 minutes and cooling sealed also reaches under 2 mg/L for free. Target dissolved oxygen ≤ 2 mg/L against about 7 mg/L at saturation, and measure it at the start and end of every run with a €170–250 meter, because it rises again during the session.

Cavitation detection costs almost nothing: a bare 27 mm piezo disc potted in epoxy (€1–5) as a passive detector, a high-pass filter or notch at 32 kHz before the amplifier (the fundamental is 40–70 dB above what you are looking for), then watch the subharmonic at 16.1 kHz, the ultraharmonics, and broadband noise. Cheap qualitative controls — aluminium-foil erosion, luminol sonochemiluminescence, iodide dosimetry — should all be negative at 4–8 kPa, and you can prove the test works by running them positive in a €40 cleaner.

Electrical safety for human immersion

We went looking for the standard that governs a bathtub with an immersed person and a dozen energised 33 kHz transducers in the water. There isn't one. The device falls between three standards, and each one excludes it for a different reason:

The architecture this forces

1. Floating supply, 2 × MOPP from mains. At 250 V working voltage that means 4 kV a.c. withstand, 8 mm creepage and 5 mm clearance, with the safety source outside the wet zones. Nothing in the water bonded to protective earth, so a first insulation fault creates no return path — the medical IT-system model.
2. SELV in the water, and step up inside the potted head. The installation rules cap in-water parts at 12 V a.c. / 30 V d.c., but a Langevin transducer wants 100–275 V. The resolution is to carry SELV along the cable and put the step-up transformer or resonant matching network inside the sealed, potted transducer assembly, so that the only high voltage in the tub is behind two layers of insulation and has no path to the water.
3. Neutral-generator topology, from the immersion patent: arrays whose feed voltages sum to zero, built with alternating coil direction or inverted piezo crystals, so even an insulation failure presents no net voltage to the water.
4. Insulation monitoring, not RCD tripping. Continuous measurement of insulation resistance to the water, alarming on degradation (the hospital Group 2 model), plus a Type B RCD upstream of the isolating transformer as a backstop for the mains-side wiring only.
5. IPX7/IPX8 potting with bend protection at the cable entry. With 8–16 immersed heads you have 8–16 independent barriers, glands and flexing cables in warm, soapy, chloride-laden water; the failure rate is multiplied, and the 33 kHz drive is itself a fatigue mechanism at the potting interface. This is a wear-out problem, not a one-shot one.
6. Supplementary equipotential bonding of every conductive part the bather can reach: 2.5 mm² sheathed or 4 mm² unsheathed copper.

Thermal limits, which are stricter than people expect

• Water ≤ 40 °C (the consumer hot-tub ceiling), with 37.8 °C considered comfortably safe. Above body temperature the failure mode is drowsiness, then unconsciousness, then drowning.
• Heater inlet ≤ 50 °C, and transducer radiating surface ≤ 41 °C after a 15-minute continuous immersed run at rated power.
• Two independent layers: a control thermostat plus a separate, non-self-resetting thermal cut-out that removes power from heater and transducers independently of the controller. Two sensors with disagreement detection; a failed sensor must read as over-temperature, not as cold.
• Hard session timer regardless of temperature, and pregnancy is an absolute exclusion (hyperthermia above ~38 °C in early pregnancy carries a documented miscarriage risk, before any ultrasound question).
• Bulk water temperature is not a proxy for tissue temperature near a transducer face.

Sources and clause-by-clause detail in research note 06. Several figures there are flagged as unverified; none of this is a substitute for a qualified electrical engineer signing off a real build.

The acceptance test

The same test for every architecture and every builder. It is the definition of "this build counts".

1. Frequency within ±0.3 kHz of the declared value, or a sweep within the declared range (FFT of the hydrophone signal).
2. Pressure at the target plane inside 4–8 kPa, reported as peak with RMS alongside, with the hydrophone's calibration reference and uncertainty stated.
3. Uniformity: over the target zone, maximum/minimum ≤ 2 on a ≤1 cm grid.
4. No inertial cavitation: subharmonic and broadband noise below the declared threshold at the working amplitude, including after 30 minutes of operation, measured by the same noise-onset method used in the ultrasonic-cleaning literature.
5. Dissolved gas: dissolved oxygen measured and recorded before every session — stable cavitation of resonance-sized bubbles has a threshold of only a few kPa in air-saturated water, so degassing is part of the dose.
6. Pulse timing 1.5 s ± 20 ms, verified on both the electrical gate and the acoustic envelope.
7. Thermal: water held at 32–35 °C ± 0.5 °C for the whole session; transducer face temperature logged.
8. For human immersion: field mapped with a body-equivalent phantom in place, not in an empty tank; head out of water; hearing protection; a written acknowledgement that the exposure exceeds the ACGIH waterborne limit.
9. Electrical: insulation resistance to the water measured and logged; leakage current measured; floating-supply integrity verified; for architecture C, SELV in the water, an insulation monitor, and an independent check by a qualified engineer.
10. Repeatability: the same measurement repeated on another day within ±10%.

What it actually costs

Prices captured 2026-09-16 from live vendor pages where possible; several marketplaces block automated access and those entries are flagged in the research note. Add Italian import VAT and shipping for the Chinese parts.

From the first build to many builds

The point of the project is the hundredth builder, not the first one. Design rules that follow from that: