Musings about a 600A QCW coil

Stuck at home thinking about things at 6 in the morning => Terrible Things...
Goodchild's and Ward's coils both run at about 300 KHz, and neither pushes over 200A of primary current.
200A is real wimpy for a DRSSTC! QCW's are apparently forced to run at high frequencies - the long, sword-like sparks fail to appear when the resonant frequency is < 300 KHz. No brick will do 300 KHz happily, so we're limited to minibricks and the like, which can't cope with more than a couple hundred amps.
Ideally, we'd like to push more amps; it gives us more room to play with (and one of my goals with this project was to figure out whether the bus-voltage ramping trick scales well with coil size and primary current).  We'd also like to not use minibricks - as gorgeous as they are, $25 a die for the exact same silicon found in a $4 IGBT is not really sane or affordable.
Well, it turns out Fairchild is pushing a new IGBT. It's essentially the beloved 40N60A4D, only twice as fast and significantly cheaper (only $2.40 from Arrow!). It can't handle as much current as a minibrick can (a bit of staring at the transient thermal impedance charts tells us that if we can hold the case at 25C [via water-cooling or some other means] we should be able to push 120A for 10 mS), but a trick pioneered by Steve Ward's "Gigantor" coil helps us here - instead of paralleling multiple devices per leg, we can instead use multiple single H-bridges, synchronized to the same secondary, to drive the resonator.  Five bridges gives us 600A to play with, and for only $50. Even nicer, we can hope that if and when it goes bang, it only takes out one bridge instead of all six.
Finally, a trick used by Steve Conner's OLTC II lets us conveniently connect the coil to the bus modulator - 10  parallel coax cables will carry 1000+ pulsed amps without adding very much stray inductance.

Gate Drivers!

I'm really, really, really bored right now, so I've decided to document the art of turning on your bricks hard and fast. Amazingly enough, there is very little out there about high power gate drives - high-side drivers don't go over 4A or so, and the low-side gate drive IC's become distressingly expensive past 9A.
First of all, NEVER, EVER, EVER drive your gates directly off the outputs of your logic IC's. Contrary to the ideal model, MOS gates are not ideal high-impedance loads; they behave like capacitors. So, depending on stupid you were when you decided to drive a gate with a logic chip, one of two things will happen:

1) You used no gate resistor at all. In this case, the gate capacitor will present itself as a short when uncharged, try to draw too much current from the chip, and the chip will let out its magic smoke.
2) You used something like a 1K resistor on the output of the chip. The RC circuit formed by the resistor and capacitor will have a high time constant, causing the transistor to stay in linear mode for too long. The switching losses will eventually heat the transistor to the point of failure, the magic smoke will come out, and you will be very, very sad.

How much current exactly do we need? Well, the datasheet reveals all secrets:

The parameter we care about here is "gate charge". The datasheet says that it is 7.2 microcoulombs, which means that to raise the gate from -15V (more on negative gate drive later) to 15V, we need to put 7.2 uC of charge into the gate capacitor. Suppose we're switching at 60 kHz (typical for a large Tesla coil). This means each cycle is about 15 uS long. A Tesla coil switches twice per cycle, so we actually care about the length of the half-cycle (7.5 uS). We are going to, somewhat arbitrarily, say that we want to turn on in 150 nS. Putting 7.2 uC into a capacitor in 150 nS requires 7.2e-6/1.5e-7 = 50A!
Why the arbitrary 150 nS? Once again, the datasheet says hi:

An IGBT or MOSFET's switching times are determined by four distinct periods: turn-on delay, rise time, turn-off delay, and fall time. In an IGBT, the turn-off delay dominates, and cannot ever be significantly shortened by the user. The "rise time" and "fall time" are largely determined by the gate drive, and are the periods during which the gate capacitance is being charged. In hard-switched circuits, these periods determine the losses. In a DRSSTC, the turn-off delay determines how far from zero current the transistor switches, and therefore how much power is dissipated in the transistor.
We can get the gate drive current out of either a single IC, or out of a discrete gate driver made from a handful of parts.

Gate Drive IC's:


These are monolithic IC's designed to simplify the task of driving gates. They have logic-level, high-impedance inputs and output stages capable of sourcing and sinking brief pulsed currents. Gate drive IC's are available in inverting and non-inverting flavors, and often come with an ENABLE pin that can be pulled high or low to turn the output on or off. They range in from tiny 2A drivers, all the way to the amazing IXDD630, which can source and sink 30A of peak current and dissipate over 100W.
Gate drivers are easy to use, but have their problems. For one thing, high-power gate drivers are hard to come by and costly - the IXDD630 is $8 a piece and is, as far as I know, a unique chip. Secondly, they are often very delicate. While able to source and sink tremendous peak currents for their size, the common ones in a DIP-8 package are thermally limited to a couple hundred milliamps average. This makes them of little use when it comes to driving large transistors at high frequencies.
As an aside, my preferred gate drive IC are the UCC37322/37321 from TI. 9A is enough to drive a full-bridge of 60A IGBT's, and they won't burn up doing it at 100KHz. And you can sample them from TI!

Discrete Gate Drivers:


To get around the problems of pre-made gate drivers, we can use a gate drive made of discrete parts, like so:

This is an inverting gate driver using a half-bridge consisting of a PFET and an NFET. The gate of the PFET is brought up to +15V as the NFET turns off , which turns the PFET on. This driver will switch the largest IGBTs, and is extremely robust - the FETs themselves can handle tens of amps continuously with proper heatsinking. It suffers from shoot-through if the NFET gate voltage is too high (i.e. both MOSFETS turn on at the same time, shorting the supply). This can be solved by placing an LM317 or other adjustable regulator on the supply of the IC driving the NFET, or by using a P-N pair known not to shoot through (I use the IRFZ24N/9Z24N).

The magic high-side isolated gate drive

There's one important point we've failed to mention: in most configurations (half/full bridge, buck converters, etc.) there is transistor on the "high-side" - its source is not connected to ground, but rather, to the load. When the low-side switch turns off, the source of the high-side switch is at Vcc with respect to ground, which means that the gate needs to be driven to Vcc+15V above ground. Furthermore, the input signal, which is ground-referenced and usually comes from the logic circuit, needs to be isolated from the high voltages that are often present. We can achieve this isolation in several ways:

1) Bootstrapped high-side gate drives:

Here, a capacitor servers as an isolated power supply, and is charged to Vcc+15V through a diode. Internal level shifter circuitry provides the necessary isolation. This circuit is simple and easy to use, but has several caveats. It cannot operate at very low or very high duty cycles, and gate drivers with internal level shifters are available only up to 4A.

2) Gate drive transformers:

A small ferrite transformer provides isolation. The input signal is first converted to an AC signal (by means of a DC blocking cap, or via a push-pull pair of gate drivers) and fed into the transformer. The secondary of the transformer is floating, and can be safely connected to the gate and source of the high side. GDT's are ideal when driving at 50% duty cycle, e.g. in a H-bridge inverter, since then a single GDT with four secondaries can provide drive signals for all four transistors. They provide a very simple way of getting the magic isolated gate drive, but suffer from several serious problems:

a) Loss of signal quality. Unless a GDT is very carefully designed, the square wave you put in will not be a perfect square wave when it comes out. This can reach the point where the output waveform becomes complete garbage, especially at high frequencies.

b) Poor efficiency. This doesn't manifest itself with small transistors, but driving a H-bridge of bricks using a GDT is rather cumbersome. Because of the low efficiency of the magnetic coupling, a gate drive that would normally draw only a few watts can draw over 50 watts, resulting in a hot GDT and hot gate drivers.

c) Inability to pass DC signals. The GDT will saturate if the input signal has a nonzero average component. This makes it impossible to pass signals of greater than 50% duty cycle.

Gate drive transformers are nearly ideal for inverters, and not really usable for DC-DC converters.

3) Discrete, floating gate drivers:

This circuit runs a standard low-side gate drive off of a discrete floating power supply, either a small iron isolation transformer or a isolated switching power supply. The low side driver can be the discrete driver mentioned above, or your favorite gate drive IC. To couple an isolated signal to the driver, one can use an optocoupler or an intermediate GDT.

Discrete gate drivers work for everything. For DC-DC converters using bricks, it is the only choice. For Tesla coils and transformer drivers, it is arguable whether a discrete driver is superior to a GDT - the former offers more control and is more efficient, but the latter has fewer parts to break.

Getting the most out of your bricks:

There are a few things you can do to help push your transistors to their limits:

1) Negative gate drive: when you turn the transistor off, don't pull the gate to 0V. Pull it to -10V or -15V; this helps the transistor turn off more quickly, and also prevents dI/dt induced turn-on. Powerex IGBT's are in fact spec'd for -10V drive.

2) Smaller gate resistors: this only really works when your trying to push the operating frequencies. Using out-of-spec gate resistors increases ringing on the gates, but can also improve switching times. Start from something absurdly small (perhaps 0 if your driver can handle it) and go up until you find the ringing acceptable. Note that this is not an acceptable technique in hard-switched circuits, as faster switching means more dI/dt, and a transistor can only handle so much; at the very least, LdI/dt induced voltage will break down the transistor.

3) Out-of-spec gate voltages: applies only to soft-switching, where most of your losses are conduction losses. Raising your gate voltages to 20, 24, or even 30V can help open up the channel more and reduce forward voltage drop. Often more importantly, 15V gate drive might not be enough to keep the IGBT in saturation when running out-of-spec currents. Unfortunately, it also slowly degrades your gates over time, so be careful! This is more relevant to small, cheap IGBTs - it is usually not necessary to drive a brick's gate beyond the spec'd 20V, and isn't really worth the risk (especially since DRSSTC's no longer try to push 2kA+ on 300A devices!).

QCW DRSSTC Bus Modulator

The big new hit in the coiling community is the "QCW" topology - essentially a DRSSTC with an arbitrary waveform generator on the bus. QCW's run at long pulse lengths (typically 10 ms) and relatively low peak currents (only a few hundred amps, compared to the 1000+ amps in large DRSSTC's). By ramping up the bus voltage slowly, QCW's can maintain low topload voltages while producing long spark lengths, thereby avoiding the flashover problems associated with smaller coils. For reference, Steve Ward's coil did ~60" of spark off a 9" secondary. The current theory is that the gradually increasing output voltage allows the spark to follow its own ion channel better, leading to longer sparks.
The core of the QCW DRSSTC, and the part that sets it apart from classic DRSSTC's, is that your usual fixed bus voltage (doubler/variac/boost converter/whatever) is replaced by a regulated step-down converter. The converter modulates the pulsed output from a capacitor bank on its DC bus into a ramp, which then powers the bridge in the DRSSTC. Rough schematic of the power end:

Nothing special - just a synchronous buck/class-D amplifier. Q1 and Q2, which are IGBT modules, chop up the DC input, and L1 and C3 form a low-pass filter which turns the square wave back into a DC voltage. Q2 is essential because it allows the output capacitor to be discharged rapidly. C1, C2, R1, R2, D1, D2 form RCD snubbers on the bricks to keep them nice and cool.
Bricks need beefy gate drives, especially important here since we are hard switching:

The optocoupler is a 2.5A gate drive which drives the P-N half-bridge, which drives the gate. The P-N bridge is necessary to charge the 200nF+ gate capacitance of the brick as quickly as possible.
The controller is just a microcontroller (currently an Mbed, but soon to be an STM32F4). The built-in ADC is used to compare the output voltage to two thresholds, a lower bound and an upper bound. If the output is lower than the lower bound, the micro turns on the high side, and if its higher than the upper bound, it turns on the low side (hysteresis control/delta modulation/bang-bang control/cycle-by-cycle limiting). It remains to be seen whether the naive software-only implementation can track the Tesla coil load quickly enough regulate properly...
And finally, pictures of all this in real life:

EDIT 3/23/2012: damn I fail at designing RCD snubbers. Schematic updated.

Good morning!

Apparently I have a blog now.

Yay documentation?