The capacitor in the tank is the smallest part in a QCW and the one that decides the most. It sits in series with the primary, it rings with it, and between them the two of them fix both numbers that matter: what frequency the machine runs at, and how much current it draws to do it.
The second of those is the one a QCW cares about, and it is why the capacitor here is a fraction of what a DRSSTC of the same size would carry.
What it is and why
The primary and its capacitor are a series resonant circuit. The bridge drives them, and at resonance the reactances cancel and the loop looks resistive to whatever is driving it. Two relations come out of that pair, and everything on this page is one or the other:
f = 1 / (2*pi*sqrt(L*C)) where it rings
Z = sqrt(L/C) what it costs in current to ring thereThe first is the one everybody sizes for. The second is the one that decides whether the machine survives the way a QCW is driven.
They are not independent choices, and that is the whole subject. Any pair of L and C with the same product gives the same frequency, so the frequency does not pin either one down. What the ratio pins down is the impedance, and the impedance sets the current for a given drive. Halve the capacitance and double the inductance and the machine still rings where it did, drawing rather less to do it.
A DRSSTC has no strong reason to care. It fires for a few hundred microseconds and the heat has nowhere to go in that time. A QCW holds the same loop on for tens of milliseconds, and over that window current is the only thing that matters, because everything that gets hot gets hot as its square.
What you decide
How small, and what small is buying you
The band is single digit nanofarads to about fifteen, against the tens a DRSSTC runs, and it is worked with the published builds at both ends on the primary page. Rather than repeat it here, the question this page answers is the one underneath it: why the small end.
Take this site's own coil. 12.2 nF against a 16.96 µH primary rings at 349.9 kHz and presents sqrt(L/C) = 37.3 Ω. Now keep the frequency and walk the ratio the other way, to David Knierim's larger machine at 64.8 nF against 5.6 µH. Same class of frequency, 264.2 kHz, and sqrt(L/C) = 9.3 Ω, four times lower [derived, both]. His measured peak primary current is 345 A against this coil's 268 to 322 A.
That is the trade in one line. The capacitance is not a target, it is a lever on impedance, and impedance is a lever on current.
What voltage it actually sees
At resonance the capacitor's reactance equals the tank impedance, so the voltage across it is the current times that impedance. On this site's coil:
V_C = I * Z = 268 to 322 A * 37.3 ohm -> 10.0 to 12.0 kV [derived]Ten to twelve kilovolts across a part that costs a few pounds, sitting in a machine whose bus never goes above a few hundred volts. This is the number that kills tank capacitors, and it is nowhere on the bus, which is why it surprises people. The same figure is why the feedback ring goes where the potential is low rather than at the capacitor, which is worked on the current transformer page.
And on a ramp, it arrives last
Here is the part that belongs to QCWs and to nothing else.
A DRSSTC bang is square. The bus is at full voltage from the first cycle, so the capacitor gets its worst voltage immediately and its heating immediately, and both stop together a few hundred microseconds later.
A QCW ramp is not square. The bus starts near nothing and climbs: on Gao Guangyan's machine, "at the beginning of the pulse, the bus voltage is quite low, on the order of 30 to 40V" (his QCW page). So the tank current climbs with it, and the voltage across the capacitor climbs with the current.
Which means the two stresses arrive in the wrong order. The capacitor spends the whole ramp accumulating heat at rising current, and then, at the end of the ramp, when it is as hot as it is going to get, it is handed its highest voltage. Dielectric strength does not improve with temperature. A part chosen on its cold rating is being asked for its rated voltage at the one moment it is least able to give it.
What it carries, and for how long
The current is lower than a DRSSTC's. The duration is not, and that is the whole of the thermal problem.
DRSSTC bang about 300 us "typically no more than about 300 us a note"
QCW ramp 14 to 25 ms this site's own worked ramps
ratio 47 to 83 times [derived]The site rounds that to fifty times the ontime, and it is a fair round number for a span that wide. What it does not buy is a fiftyfold rise in heating, and the reason is everything above: the impedance was raised precisely so the current would fall, and heating goes as the square of current. Work the multiple on your own two machines rather than taking either number for it.
For the part itself the figure that matters is the RMS current through it, which at a peak of 268 to 322 A of sine is 190 to 228 A while the coil is actually ringing [derived, peak over root two]. The ramp's envelope means the RMS across the whole window is below that, since most of the ramp is spent below the final current. What it is not is intermittent. There is no gap in a ramp for the dielectric to give heat back.
What sort of part
Polypropylene film, in a series and parallel array, which is what MMC stands for and why it is one. Two published QCWs, both quoted from their builders' own specs:
Gao Guangyan's QCW 2 lists "5.875nF MKP 3.2kVAC/8kVDC MMC" (his page). Note which rating he prints first. An AC rating on a film capacitor is the useful one for this job, because the tank is the one place in the machine where the voltage genuinely alternates at the ring frequency, and the DC rating alone says nothing about what the dielectric does when asked to reverse a few hundred thousand times a second.
Knierim's larger machine goes the other way for the reasons above, and pays for it with parts: 64.8 nF built as 3s72p of 2.7 nF 1600 Vdc polypropylene, split into two 32.4 nF halves, one for each of his two interleaved primary windings. Check the arithmetic on that array, because it is the arithmetic of every MMC: three in series makes 0.9 nF at 4800 V, and seventy two of those in parallel makes 64.8 nF [derived]. Two hundred and sixteen capacitors to hold a tank that one part could hold the capacitance of, because no one part holds the voltage.
The numbers
C L f sqrt(L/C) I peak
this site's coil 12.2 nF 16.96 uH 349.9 kHz 37.3 ohm 268-322 A
Gao QCW 1 and 1.5 12.8 nF - - - 100-160 A
Gao QCW 2 5.875 nF - - - -
Knierim, larger 64.8 nF 5.6 uH 264.2 kHz 9.3 ohm 345 AEverything in the impedance column and the frequency column is sqrt(L/C) and 1/(2*pi*sqrt(LC)) on the figures beside it [derived]. The gaps are gaps: those builders publish a capacitance without the primary inductance to go with it, and a number that is not on the record does not go in a table.
What will get you
Sizing it from a DRSSTC calculator. They exist, they are good, and they are built around a machine that fires for 300 µs. Type a QCW frequency in and you will get tens of nanofarads, which rings correctly and draws two to four times the current, which you then hold for fifty times as long.
Reading 64.8 nF as permission. It is on the record, it works, and it works because the bridge under it is eight IGBTs in pairs per switch. The capacitance did not make that machine possible. The bridge did, and the capacitance is what made the bridge necessary.
Testing on a short ramp. Covered in the warning above and repeated here because it is the one that passes every bench test and fails in the room.
Buying the voltage rating in DC only. The tank is AC at a few hundred kilohertz. A part rated 8 kVDC and 3.2 kVAC is telling you two different things and only one of them is about this job.
What goes wrong
It gets hot and nothing else looks wrong. The frequency is right, the arc is right, the bridge is cool, and the tank is warm after a run and hotter after the next. That is dielectric loss with no gap to shed it in, and the fix is not a bigger fan. It is less current, which means more impedance, which means less capacitance and more turns.
It fails at the end of the ramp, repeatedly. The two stresses arriving in the wrong order. If the failures cluster at the top of the envelope rather than scattering through it, that is the signature.
One can in a string goes and the rest look fine. A series string divides the voltage between its members, so losing one hands its share to the survivors and the string is now over its rating. MMC failures are rarely a single part by the time you find them.
Where next
The band itself, both ends held by published builds, is on the primary on a QCW, which is also where the impedance argument continues into the winding.
What the same capacitor does on a machine that is not ramped, where the current is higher and lasts a few hundred microseconds, is the MMC is sized by current. That is where the array arithmetic lives: the ripple ratings and their conditions, and why smaller parts carry more current per nanofarad.
Whether the part in that position resonates with the primary at all, which on an SSTC it deliberately does not, is the DC blocking capacitor, where the value alone decides which of the two machines you have built.
Why the frequency this all resonates at does not stay where you put it is the frequency slides down the ramp.
And the bus bank behind it, which is a different capacitor with a different job and is sized by arithmetic that actually is easy, is capacitance is the easy half.