When I visited the electronica fair in Germany last year, two things caught my attention that made up a perfect match: The spiral tesla coil by Daniel Eindhoven, displayed at the Elektor booth, and Texas Instruments new Gallium-Nitride switches, advertised for fast switching power converters.
Here is the result of what I created since I saw this in November:A pcb tesla coil that can generate 5cm arcs into the air, powered from my 5V 3A USB C laptop jack. The tesla coil enumerates as a USB MIDI device and can be used from any synthesizer software like LMMS or Ableton.
Disclaimer: This is not an easy project. You have to solder eg. fine-pitch QFN. I might sell pre-assembled kits in the future.
Latest video with three coils controlled via midi:
Known bugs (V1.5)
Missing pulldown for gatedriver input of the flyback converter.
This can cause problems when in DFU mode as the mosfet might switch on permanently, shorting VUSB
Missing input decoupling capacitors underneath the LMG5200.
Adding additional high-quality 100pF and 100nF ceramic capacitors on the bottom of the pcb improves performance drastically
Unknown problem causes GaN module to fail when touching the arcs with the finger or a screwdriver.
The reason for this is unknown so far. Maybe EMI, maybe bad layout. Just don't touch the arcs for too long for now. Discharges to air work reliably.
Ideas for V1.6 (or V2.0?)
4 layer design to improve the ground return loop
connect the button to both BOOT0 and a GPIO so no additional jumper for initial flashing is required
primary current feedback for overcurrent protection and feedback with less propagation delay for better zero current switching, using phase lead compensation
The PWM for the
switching mosfet is generated by the STM32. The STM32 thous has full
control of the output voltage, allowing things like slowly ramping up
output voltage to not overload USB, software-configurable bridge
voltage and potentially QCW-DRSSTC operation.
shoot-through condition happens when both switches in a half-bridge
design are conducting simultaneously. This usually doesn’t happen
intentionally, it rather happens as a side effect during switching:
If you switch on one fet in the exact same moment you switch off the
other, due to time delay mismatch it can happen that for a very short
period of time, both switches are conducting, usually causing death
to both fets. To counteract this unfortunate event, you can add a
small amount of time where both fets are off, “break before make”,
or, in EE slang, dead-time.
Even though the
LMG5200 already integrate an impressive amount of circuitry, one
thing missing is dead-time logic. Texas Instruments arguments that
this is an advantage as it gives the designer the freedom of
controlling the two individual GaN devices in the half-bridge as they
wish. However, it also gives you the freedom to kill both GaN Fets if
you turn them on simultaneously.
I didn’t realize
that in my first attempt, and adding dead-time logic was the
milestone that lead to the very first sparks from this design. The
circuit is pretty easy: Two triple-input AND gates are feed with the
feedback signal, one inverted, the other not. Also, enable from the
STM32 is applied to both AND gates. The opto-coupler was added to separate the STM32 from the electronics for the tesla coil: As you will learn later, the GND for everything related to driving the LMG5200 is on a separate ground plane, only conneted to power ground through the LMG5200 itself. This is again because the fast switching of the LMG5200 can induce voltage spikes that can lead to unwanted behavior. The opto-coupler is there to connect these to domains without introducing noise to this separated ground plane.
The third input is the inverted
output of each other AND gate: Only one of them can output high at a
time, as they are blocking each other. As the signal has to pass
through the inverter first to enable the output of the other AND
gate, the delay time of the inverting gate is added to the signal as
dead-time where both AND gates outputs are low.
Is is very important
for the tesla coil to operate at it’s own resonant frequency.
However, this is not a fix value, it heavily changes during
operation: Both the amount of metal in the surrending as well as the
very arc created by the tesla coil forms an additional capacitance
that decreases the resonance frequency. The easiest way to follow
these changes is to operate the coil in a self-resonanting mode: The
output of the coil is sensed, amplified and feed back to the driving
circuit as positive feedback. This causes an undamped oscillation at
the resonant frequency.
There are multiple
known attempts to recover this feedback signal. The biggest
difference is either it is recovered from the primary resonance
circuit or the secondary coil. The easiest way is to sense...
As explained earlier, the coil design has an usually high resonance frequency. With traditional Si-MOSFETs it is very hard to build a efficient fullbridge at this frequency. This is mostly because of the high gate charge and switching losses of mosfets the size required for this application.
It is possible to get sparks even at these frequency, using class-E topology. However, in this mode of operation, the mosfet is in linear mode and has very high power losses (http://www.richieburnett.co.uk/hfsstc.html
Daniel Eidhoven solved this issue by using traditional bipolar transistors in emitter follower configuration. This allows fast switching and small delay. However, to drive these power transistors, you need another totem-pole driver that outputs a signal with the full output voltage swing. Daniel used a mosfet driver IC for this. With the device he uses he is limited to 32V bridge voltage. Also, bjts are not known for their great efficiency...
To overcome these issues, GaN and SiC switches require some special gate drivers (eg. with a output voltage of -6 to 10V. The negative voltage helps to bias the gate against mentioned voltage spikes during turn-off.) located as close as possible to the switch. Texas Instruments overcame this issue by developing a device with two internal GaN switches in half-bridge configuration that includes all the required low- and highside gate drivers. You can buy these devices since ~6 months for $9 each.
We can separate the design in five functional blocks:
Flyback boost converter
Feedback signal recovery
GaN full bridge
(three selected design revisions. From left to right: V1.0, 2018-11-22 - no dead time, bad power layout; V1.3, 2018.12.22 - dead time, first arcs, GaNs randomly die after 1-20min; V1.5, 2019-12-04 - good performance, good reliability - still room for optimization ;)
In total, it took 6 pcb and design revisions to get the current performance and reliability. This is mostly due to the very challenging requirements that come with the use of GaN switches. Some pcb revision could have been avoided if I would have followed the recommendations in the datasheet in the first place. Usually, you can interpret design guidelines and maximum ratings as "recommendations" - this is definitely not the case when it comes to GaN and SiC design. When TI writes about placing components "as close as possible" like they do on page 10, 13, 14, 15 and 16 of the datasheet - 8 times in total! - they really, really mean it like that. "Close" does not mean "nearby". "Close" does not mean "right next to it". "Close", in terms of low parasitic inductance, means: on the opposite of the board, right under the according pads, connected with as many vias as you can stick in there.
I had to learn this the hard way - I might have killed 10 LMG5200 modules because I though I knew better. But more about layout considerations, later. Let's start with the easy stuff:
(yes, it looks like USB DN and DP are shorted - the schematic symbol is for a USBLC6-2SC6, but I'm using a USBLC6-4SC6. I'll fix this soon but no worries, it's just graphical)
There is nothing too special about this. I used the STM32F072 for this project, one of my favorite parts from the STM32 series. The STM32F072 is cheap, has a proper ARM M0 core, and supports crystalless USB operation by using clock recovery / syncing from the USB frames. So you get a decent Micro with USB and lot's of STMs powerful timer peripherals for under $1 that requires minimal external components. I used the QFN version because I find it much easier to solder than QFP.
Basically, a tesla coil consists of an air-coupled transformer with a very high turn ratio. Low AC input voltage is transformed to high AC voltage, that is, in case of the dual resonance system, additionally amplified by electrical resonance effects. This high voltage is concentrated at a small tip, the breakout point, where the local field strength it high enough to ionize the air and create the arc.
We can separate the design in two parts, the passive transformer itself, and the electronics driving it. Let's start with the transformer first, as this explains some of the design choices for the driving part.
Most of you probably already have seen a tesla coil
(cc by Johannes Müllers)
These devices are usually smaller than 1m in height, but almost always bigger than 10cm. This has a very good reason: As the transformer has no core other than air, and there is some limit to the number of turns the coil can has (usually in the range 200-1000), the secondary resonance frequency (the frequency of the LC oscillator defined by the inductance of the secundary windings and the capacitance of then top electrode to ground) is only defined by geometry. For a 30cm coil, this would be in the range of 200-300kHz.
Here lies the first problem: Building bigger tesla coils is actually easier than a very small one. For DRSSTCs, the power electronics driving the transformer is usually based on IGBT switches, sometimes MOSFETs. But even with very modern Si-based switches, the switching losses increase massively at a frequency >1MHz. For reference, most motor controllers use 15-25kHz. Your laptops power supply no more than 200kHz. My coil design is at 2.6Mhz.
Usually, you use isolated magnet wire around a plastic tube for the secondary coil. This has the obvious advantage that as the voltage increases towards the top of the coil, also the distance towards the sensitive power electronics and primary windings on the bottom increases. However, manufacturing these coils is a lot of hand work. You can't just buy them in arbitrary quantities.
PCBs, however, have become super cheap. You can get 10pcs 100x100mm dual layer pcbs for $5 at your Chinese manufacturer of choice. With 5mil tracing and 5mil spacing as required by most manufacturers, this gives you 190 turns on a 100x100mm pcb.
(left: bottom side with single primary winding, right: secondary winding with 190 turns)
There is nothing special really to these coils. It's really just a huge spiral going from the outside to the inside. I placed a circle of white silkscreen in the center the increase isolation a bit. The primary winding is a single turn. I also tried two and three turns (yes, I had to order new pcbs to test this, but it's just $5 so why not?) but as the secondary turns are already so low (190), you want to get maximum transformation ration. The primary winding is connected with pads to a XT60 high current connector, and the secundary winding has a pad in the center and a plated mounting hole on the low voltage side to connect the base point to the electronics for signal recovery and grounding.
Unfortunately, I was not able to create the gerbers for this in KiCAD as it interpolates circles with very bad resolution. I tried a svg to gerber converter but that also failed. I ended up booting my Windows VM and just used Altium Designer to create the files. Luckily, I did not have to change these files a lot, and you can find the gerbers on GitHub.
I ordered my coils at jlcpcb. They rejected the design because it failed the audit. Technically the design is perfectly within spec, however it is somewhat relatable: You're requesting...