Quiescent accelerator

Administrator

Administrator - 04/05/2026

I am rebuilding a deliberately difficult switching circuit to test how board geometry, short current paths and several kinds of filtering affect noise across a wide frequency range. I am planning a number of these types of builds.


Tags: electronics powerelectronics EMI Blender resinprinting prototyping measurement 

Experimental power electronics build following a thermal failure. Project period: 2026 rebuild in progress.

The first version of this circuit failed thermally. That is an efficient way to discover that a design still contains several unanswered questions, although not one I recommend as a general project-management method.

I am rebuilding it as a controlled experiment in high-frequency power electronics and electromagnetic interference. The circuit is intended to make switching noise, then give me something repeatable against which to test board geometry, short current paths, shielding and several complementary kinds of filtering. It is not yet a finished low-noise power supply, and it is definitely not evidence that every idea in the design works. The point of the rebuild is to find out which ones do.

The most useful part of the project so far has not been a particular capacitor or MOSFET. It has been treating the whole assembly as a mechanical object before drilling a single hole.

Building the board in Blender first

The board starts as a 70 mm by 50 mm stack built around FR-4, thin copper sheet and Kapton insulation. Instead of measuring the components one at a time and hoping the holes met their pins, I modelled the board and the physical components at their actual dimensions in Blender.

That model became a resin-printed drilling template. Every component lead, mounting point and feed-through has a matching hole. I drill the corner holes first and use PEEK grub screws to clamp the printed template to the copper and FR-4 stack. The template then guides a high-speed drill through the remaining positions.

It sounds like an elaborate answer to perfboard, it's in fact aligning to a 2.54mm grid which is covered entirely on both sides by 0.2mm copper ground and power planes - not your average circuit, in fact having discussed in detail with various AIs the techniques I've intuitively reach are akin to home made military radar. In practice it worked extremely well. The components fit the planned locations, the holes remain registered through the layers, and awkward groups of wires can be prepared before the parts are fitted. It also lets me inspect clearances and current paths while changing a digital model is still cheap.

The copper arrives with a protective coating. After drilling, I remove that coating and apply Kapton to the inner faces. After covering the outer faces in kapton tape, At each intended connection ( drill point ) I cut a small window in the Kapton, expose fresh copper and tin only that patch with leaded solder paste. This produces neat connection points without asking solder to adhere to an oxidised sheet or leaving large areas of bare copper.

The same approach works underneath the active devices. Short lengths of enamelled copper wire pass through the drilled holes and meet the trimmed component pins on the upper side. The small decoupling capacitors can sit close to the plane, with a Kapton bridge insulating the connection that must cross it. The technique is simple once the geometry is right. Getting the geometry right is the interesting bit.

What the copper planes should do

The large copper faces are intended as closely coupled power and return planes. A broad return path placed close to the outgoing path reduces loop area. That should lower loop inductance and give high-frequency current a shorter route than a long wire around the board.

I originally described the stack as an “induction-cancelling giant decoupler”. That is a memorable description, but it gets ahead of the evidence. The plane pair does form a capacitor, although its value depends completely on the layer order and dielectric thickness. But it's is in fact true, just like twisting wires together or opposing positive negative flat magnet wire, they cancel one anothers electromagnetic field drastically reducing parasitic inductance along with vias (small holes) allow extremely short connections - something incredibly important as your approach 100khz and above circuitry, GAN devices require exact fractional mm connections, else they simply fail instantly because the frequency is such that a bump in the road is hit so fast it might as well have been a 100m high ramp as far as parasitics go.

For scale, two 70 mm by 50 mm 0.2mm copper plates separated only by 20 micrometres of Kapton would be about 5.3 nF if the Kapton had a relative permittivity of 3.4. If the electric field instead crosses 1.8 mm of FR-4 (which it does) plus two Kapton layers, a rough series-dielectric estimate is about 70 pF. Those two answers differ by nearly two orders of magnitude. Until the final stack and dielectric properties are fixed, a precise capacitance would be exceptionally hard to calculate. 

A switching circuit built for observation

The planned source is a 12 V current-limited bench supply. A small NE555 timer module provides adjustable frequency and duty cycle so I can approach each test condition gradually. A TC4420 gate driver then drives a Texas Instruments CSD18533KCS MOSFET. Earlier experiments used an IRFZ44N, which is one reason some of the concept images and notes do not agree with the intended rebuild.

The timer is there because this is an experiment, not a regulated converter. A predictable external drive makes it easier to compare one physical change with the next. The final transformer, current limit, clamp and snubber values have not been selected. Supply voltage alone is nowhere near enough information to choose them. Primary inductance, duty cycle, leakage inductance, peak current and the measured drain waveform all matter.

This is highly important. A flyback circuit supplied from 12 V can still create dangerous secondary voltages and stored energy. I will not publish a build recipe until the topology, isolation intent and protective parts have been checked as a complete system. The bench supply is a source limit, not a substitute for a primary-current limit.

One small construction detail is already worth documenting. I wanted a compact gate resistor with several parallel elements, so I arranged five 100 ohm 2510 SMD resistors as the faces of an open pentagonal tube and joined them with copper buses at either end (Not fiddly... at all). Black tack holds the parts while one end is soldered; the assembly is then turned over for the other bus. The effective resistance should be about 20 ohms if the five parts share current evenly. Its real parasitic inductance and pulse behaviour still need measurement, but it is a useful example of turning a circuit element into a piece of geometry, because not only does it dissipate heat being the inner and outer surface of a tube unlike a metal film resistor, but each opposing resistor actually cancels it's own inductance via opposition with the rest.

Filtering more than one frequency decade

No single capacitor behaves like its label over every frequency. A large polymer capacitor stores plenty of energy but carries more inductance than a small surface-mount ceramic. An MLCC can respond quickly, but its useful capacitance may fall with DC bias and its mounting path can dominate at high frequency. Combining parts can also create anti-resonance instead of the smooth broadband response shown in a tidy sketch.

The provisional network therefore uses several technologies: bulk polymer electrolytics near the power entry, smaller polymer parts closer to the active devices, film capacitors and groups of MLCCs at the pins. The driver supply also uses feed-through capacitors, including TDK and Murata parts named in the design notes, so the conductor passes through the filter element rather than arriving along a long branch, it also uses very small value resistors on some capacitance  to distribute power delivery across the need for instant fast power, through to much slower smoother power, and creates a broadband capacitance network that filters noise at multiple ranges without stressful power spike being thrown all the way back from the PSU.

The power lead is intended to include a Murata DST9 feed-through between XT30U connectors. On the board, the input should meet the return plane close to the MOSFET source. I have called this a star ground in notes, but a solid plane does not have a single magical point. The useful claim is narrower: keeping the high-current switching loop local should reduce the amount of shared return path seen by the timer and driver.

The resistor values sketched beside the bulk capacitors are hypotheses about damping, not a finished impedance design. Before treating the network as broadband filtering I need an impedance or Bode measurement, including the fixture and a reference run. Otherwise I am merely arranging an impressive number of capacitors.

Enclosure and cooling

The eventual assembly is intended to sit in a rectangular, conductive-lined printed enclosure connected to the circuit reference at a deliberate point. A shield is only as useful as its seams, apertures and grounding. The fan opening in particular is also an RF opening, so the enclosure cannot be described as a Faraday cage merely because copper appears somewhere inside it, but it's highly effective.

The MOSFET will use an electrically insulating thermal interface to a heatsink. A 30 mm high-static-pressure fan is planned above it, exhausting through side vents. I want the cooling system to remain quiet, but I have not measured its sound level, so what I will say is that you can run one up hanging on it's power cable, and notice it's actually hard to know if it is spinning save the fact that it's exhaust drives it to move.

Thermal measurements matter more than the description. I want device, heatsink, transformer and enclosure temperatures logged at repeatable operating points. The earlier failure makes that test a requirement, not a decorative graph at the end.

What I need to prove

The rebuild gives me a useful sequence of tests:

  1. Verify the gate waveform and peak primary current at low duty cycle.
  2. Measure the drain overshoot before selecting the final clamp and snubber.
  3. Record temperature rise under bounded operating conditions.
  4. Compare conducted-noise measurements with and without each filter stage.
  5. Measure the plane stack and complete network instead of relying on nominal component values.
  6. Repeat the tests after the shielding and cooling are fitted.

The result may be that one of my clever constructions adds little beyond making assembly harder. That is allowed. A useful experiment needs the possibility of being wrong.

For now, the project demonstrates a fabrication method I will use again: model the physical circuit, print the tooling, register the complete stack, and make the current paths part of the design rather than an afterthought. The electrical claims come after the measurements.