Global-feedback Electrostatic Amp (Watercooled!)

Hi all,

I designed a electrostatic headphone amplifier with a couple goals in mind:

• cheap parts, no specialty items, all available from Digi-Key (in stock!), including transformers
• good sound no matter the component tolerances

The end result was a high-gain voltage amp (around 60dB of gain), but it was extremely non-linear due to the huge capacitances of the main NPN BJT involved: Fairchild's KSC5027. And that's despite operating in class A!

The solution was to apply large amounts of feedback (20dB) and do some analysis to ensure stability and phase weren't a problem. End result is a 40dB amplifier that has relatively OK specifications, and sounds great (as far as I've tested, anyway, numbers forthcoming).

Quick stats (simulated):

Amplitude: +/-400V (1500Vpp stator-to-stator)

Distortion: <0.1% 0-20KHz at 300Vpp on a single output

Frequency response: 0-50KHz +/- 1dB

Open-loop gain: 69dB

Closed-loop gain: 40dB, 60 degree phase margin

The amplitude and distortion measurements check out as far as my scope goes, but I haven't had the time to hook it up to a proper analyzer yet.

Since this was the result of a school project, I've got some documentation available (and the whole shebang) at https://bitbucket.org/cyanoacry/ee91

Unfortunately, 2/4 stator boards have blown due to insulation problems (that's me being cheap and forgetful with space), so I'll be designing a proper second round of boards.

For convenience's sake, all KSC5027 output transistors and the power-supply FETs were watercooled using a Koolance setup. There's no need for it, but I wanted to play around with watercooling some day...

Pics:

Overall view of the setup (messy, I haven't cased it, I know)

Close-up view of a single stator board and the watercooling setup

Another view of the stator boards (2/4) and the power supply. The dual toroids per rail can be seen in the background.

Schematics/docs:

I'm interested in making this amp better, so please give me your feedback. It's my first design, so sorry if there are any rookie mistakes! m(_ _)m

Edit: fixed amplifier/power supply schematic to match documentation.pdf (correct version)

Originally Posted by kevin gilmore 

Q3 upside down in the final amp schematic, error in final_psu schematic, gate drive

 

Both are ok in documentation.pdf

 

Thanks for pointing that out; I've updated the top post. Both of these errors made it to my PCBs, but weren't present on the simulation diagrams. Luckily I had the foresight to check my boards before powering them up, otherwise they would've blown...

Quote:

Originally Posted by kevin gilmore 

The 2sa4686  is a 1000v rated part, and still available, much better

substitute for ksc5027.

 

ksc5026 also works and is half the capacitance of ksc5027

 

you really need the ceramic insulators on the heatsink, because

sooner or later at 400v, that setup is going to arc over.

Is there any easy source for the 2SA4686? It seems the normal places don't stock them; octopart.com doesn't list any suppliers, and Google only turns up people who have it available "on quote". 

I might using the KSC5026's on the next revision, but they'll probably have to wait until the third revision. I've yet to do the math to figure out general parts specifications (the KSC5027s here happen to contribute just the right amount of rolloff via their parasitics), so I'll toss them in and hopefully the amp doesn't oscillate. The SOA diagram also shows them having a little less current capacity at +800V (bias current is ~23mA), so as long as there's not sustained DC offset it should be OK.

At the moment they're protected via some Arctic Alumina epoxy (which is supposed to be a pure insulator, though I couldn't find breakdown voltages in the datasheet). They're held off from the heatsink via two Kapton tape "spacers": two thin slivers of tape are located on the backside of the chip, so that the package never touches the heatsink. It worked for a while, and the packages stayed relatively cool at ~40C.

This probably isn't enough though, since my second board failure (after something like 2 hours of testing) was probably caused by a heatsink-KSC5027 arc (at least there were sparks seen in that area, not sure which device it is yet).

When I get around to making a proper heatsink that I can bolt on/into, I'll be sure to use Kapton/ceramic washers.

Quote:

Originally Posted by spritzer 

It's a bit funny and also a bit sad to see something like this which is better than most of the commercial amps on the market.  Something to be proud of.  

 

The 4686A would have been better with it's far lower output capacitance.  The isolated version of the 5027 sold by Mouser would also have helped with the potential shorts. 

 

Thanks for the compliment. :) My goal here was to make something a little like an upgraded eXStatA, now that the KSC5042 is out of production. 

I'll try the FJPF5027 on the next revision, should make heatsinking a little easier.

My plan for now is to fix the PCBs, upload some gerbers and the BOM for those people who want to take it into their hands, and then build one with proper casing. 

In the end, I'd like to make this a design that one can throw "any working part" into and get tolerable results. With HV BJTs going out of fashion (goodbye CRTs!), I'm not sure the KSC5027 even will be around much longer, so it'd be nice to specify exactly what _about_ the transistors is necessary (say, output capacitance of X for stability). The strong loop gain should, barring oscillation, bring even marginal parts to good listening quality.

The ultimate end goal (and what prompted me to design this amplifier), is to integrate this with an additional feedback loop that takes advantage of capacitive sensing on the diaphragm (consider using your Stax as a huge condenser microphone), which should offer considerable benefits with regards to damping. I'm also not sure if this will actually work, but with active diaphragm feedback, it could turn any pair of open-backed stats into noise-cancelling ones, too. This will probably involve me homebrewing some headphones to change the diaphragm characteristics, so we'll see what happens.

Here's the open-loop frequency response (simulated). Not great, by any means. Distortion's also terrible.

And by closing the loop, we get something that's razor flat out until 200-300KHz, and distortion becomes the limiting factor. We've also got flat phase response all the way out there too!

Originally Posted by spritzer 

 

If you want people to build the amp in any quantity then I'd advise either to use the 2.5" onboard heatsinks or set the boards up to use angle brackets as it makes assembly and servicing much easier. 

Yeah, a large problem is definitely the cooling issue. It was nearly impossible to solder together the magnet "bond wires" I used for the TO-220s in this case, cool idea, terrible execution. I'm working on a second set of PCBs right now that use onboard heatsinks; I'll probably use the Aavid-Thermalloys (Digi-key HS380-ND or variant), since they're pretty well-available and dissipate lots of heat.

I just got an email about the diyAudio chassis(es?) that look pretty cool, any thoughts on the universal mounting standard?

http://www.diyaudio.com/forums/images/diy/store/board-documentation/universal/universal-mounting-specification-v2.1.pdf

http://www.diyaudio.com/forums/site-announcements/209285-diyaudio-store-usa-soon-stock-chassis-requesting-your-feedback.html

Quote:

Originally Posted by jcx 

the Wilson Current Mirror is only worthwhile with matching Q - to have one of the Q be a much lower hfe, larger area, radically different doping profile with 100s of V Vce means the Ib matching condition is not even close

http://en.wikipedia.org/wiki/Wilson_current_mirror

 

I hadn't realized the 3rd transistor also needed to have an Hfe match for a perfect mirror. It's really not to make the mirror better, but rather I use it because it at least allows me to match the bottom two transistors. In this case, both halves of Q2 are a matched pair (at least NXP's BC857BS is indicated as such). I trust a little bit more in using a bottom matched pair and a mismatched cascode versus trying to match Vbe on two different high-voltage devices. The latter seems like a poor idea, especially due to the power dissipation and temperature imbalance that'll be involved. I kind of think of it as a current mirror feeding a common-base amplifier as a tunnel device, though that may not be exactly true.

The few simulations I've done also tell me that distortion goes way down versus using two KSC5027s/FZT560s in the 2-transistor arrangement.

The buffering's necessary to prevent loading issues (before I buffered, the input divider took too much current in order to give good bandwidth into the op-amp summer). There shouldn't be much phase shift since the op-amps are rated out to 50MHz (well, so maybe a couple of degrees of phase shift), and they're internally compensated to be unity-gain stable.

I actually believe it's common practive to reduce parasitic capacitance by eliminating the ground plane around the op-amp (more information at http://www.support.wdv.com/Electronics/Fab/PCB%20Layout/HighSpeedPCBLayout.pdf, pg 70). I applied local bypassing with a ceramic underneath the DIP footprint, and a tantalum a little while away, seems to work OK so far.

The biggest layout restriction was probably the voltage stand-off distances; I tried to ensure that traces weren't too close given their voltage differential.

The quad package was mainly chosen for price and simplicity (routing less power pins gave me more room for signals).

Op-amp rolling will be hard with this circuit, since there's minimum bandwidth and gain requirements that most audio op-amps will fail. I'm not sure that different op-amps will even sound that much different here, as long as they don't cause instability. They're not acting as active gain devices, nor are they outputting much current, so it should be OK.

It's hard to make good analog circuits using MOSFETs, I believe. Mainly because most modern MOSFETs are specifically NOT RATED for the linear regime. They'll probably blow up (literally), since they're meant for switching purposes where power dissipation is minimal. Thermal runaway due to temperature variances in the die will happen, unless you run parts very conservatively. Unfortunately I think that places most FETs out of the picture for this design.

More details: http://sound.westhost.com/articles/hexfet.htm and http://www.irf.com/technical-info/appnotes/an-1155.pdf. The failure mode illustration in the last link is particularly showing.

One thing that makes me feel uneasy is the use of devices that don't have published SOA graphs (whether BJT or FET), because that usually means the manufacturer only intends use in switching applications. They might blow up, they might not, who knows?

Ultimately we are at the whim of the manufacturers. High voltage devices are out of fashion, so I'm not sure how much longer DIY stat amps are going to be around, given the dwindling device selection.

Quote:

Originally Posted by jcx 

I disagree with Gilmore's practice of using device limitations for loop gain compensation: semiconductor junction parasitic C, ft roll off are nonlinear with device operating point

use faster parts, make the loop gain compensation explicit with quality C and gain design flexibility

even ceramic NP0/C0G are way better caps than a semiconductor junction

 

Sorry, these are the fastest devices that I could find at Digi-Key given the voltage requirements (I didn't realize the existance of the 2SC4686 at Mouser, but it's still somewhat EOL, so I'm not sure how I feel about using them).

I'd like to use a good 'ol dominant pole capacitor, but I don't see a way to do it without putting the capacitor would in the signal path (if not, putting it in the compensation network leads to serious high-frequency peaking).

Quote:

Originally Posted by analogsurviver 

Very intersting contribution. Subscribed. Any real data/meausurements with real world load, ie. electrostatic headphones attached ? The graphs seem to show performance of the amp without any load.  Particular of interest is power bandwidth and rise time(s) for given capacitive load (s).

 

I tried to be fairly conservative in design, so the simulation runs are done with a 160pF capacitive load attached. The 500 ohm output resistors are to minimize the phase effect of the capacitive load, so it should be stable given most (if not all) electrostat headphones. Can't speak for ESLs though.

I haven't been able to get any more real-world data because the high-side power supply seems to have arced over (oops). When I get the new run of boards, I'll let you all know.

Thank you for the support and feedback so far. I hope I can make this a little more accessible for people who want a slightly more advanced DIY design.

edit: didn't realize post sounded so angry, edited to be happier, my bad

Quote:

Originally Posted by analogsurviver 

Well, I am known to be pretty sarcastic about the use of tubes/valves in audio - by calling such designs HSDs ( Hollow State Devices ).  But if there is any place in audio circuitry where tube is better choice than anything else (dwindigly (un)available) for high voltage OTL amps driving electrostatics,  I have not heard of it yet. 10-12 KV peak to peak circuits for driving ES loudspeakers are possible - perhaps looking around which tube type that would be suitable, is still available in reasonable quantities and is still under the radar by audio community ( and rebrand/markup artists that cater to this community needs ) is perhaps in order. Generally, ESL amps use whatever devices used to be used for horizontal deflection in cathode tubes of  (colour) TVs. 

 

For this extraordinary application, tubes are still the premium choice - by far.

I really, really wanted to make a hybrid version of this amp using EL34s as output followers, but the analysis didn't bode that well. (SS won in distortion measurements in all frequencies by about an order of magnitude) Tubes introduce a lot of phase shift (their poles just stack and stack), so it's hard to build a feedback circuit around them. For the moment, I focused on lower overall distortion, so solid-state was the way to go. Maybe in the future though, I'll build the hybrid version and see how it sounds.