Snubbing DCM (nonsynchronous) Buck converter

FINAL UPDATE:

I've posted an outcome report below. I'm very satisfied with how this has come out and I learned a lot in the process. More details and before/after shots in the answer below. Thanks to @Andy and @winny for their help and encouragement :)

Just a note to help search, the board used here is colloquially known as the "Fake LM2596 eBay buck converter". The chip behaves like an LM2576 (52kHz switching frequency instead of 150kHz) although it possibly lacks the protection circuitry.

I'm working with an LM2576-based buck converter as a learning project. Basically I've made a \$0.99 eBay buck into a highly-affordable if somewhat less flexible demo board :)

The LM2576 enters discontinuous conduction mode (DCM) at light loads (<500mA or so), and when it does, it exhibits vigorous ringing on the switch node when the diode turns off, the residual energy in the inductor bouncing back-and-forth between the inductor and the diode capacitance:

The datasheet states this is nothing to worry about, but that it can be addressed with an RC snubber across the inductor if desired. I understand most folks wouldn't bother with it, but since this is a learning project and my first chance to design a real-life snubber, I want to do the best I can and learn as much as possible for future snubbing challenges.

Also, the ringing counts as EMI in my world, and it does couple to the output:

I've done a lot of research on snubbers, six distinct sources including Rudy Severns' eBook and all the application notes (and a master's thesis) I could find. Because I'm doing this as a learning project, I want to master the concepts and procedure in general, not just get a solution good enough for a particular application.

Here's the schematic (the actual diode is marked SS34, the SS3P5 is the closest available LTspice model):

And here's the board layout:

reverse side mirror-imaged for easier through-hole matching:

By following the prescribed procedure (detailed below), I get the values 1nF and 220R. These tame the ringing substantially:

UPDATE:

I installed a pot in the resistor position and discovered 430Ω provides maximal damping with a 1nF (C0G) cap.

With the optimal resistor in place, I was able to do better:

But they don't critically damp the ringing, which is what was indicated in [1] Todd at least and what I was after.

I've found on simulation that I can achieve critical damping (or something quite close) with the values 8.6nF and 350R.

So I'm wondering why the sources I found all pretty much agree, but all give me lower values than what seems to be called for.

I have a couple hypotheses:

1. Mine is a special case. Most of the sources use synchronous buck converters in their example. These have a larger capacitance and much lower inductance. The formulas don't work when the situation is reversed. (In my case, L=47uH, C=100pF.)

2. Nobody wants to snub to critical damping. In my case, the energy to be snubbed is small (single digit mW), in the more common cases, the designer is making a significant trade-off against efficiency and a couple wiggles left on the waveform is perfectly acceptable.

Can anyone help me understand what I'm seeing?

One anomaly I've noticed is that all the sources say the RC time constant of the snubber should be short compared to switching frequency but long compared to rise time of waveform to be snubbed. The values I got (1nF, 220R) have a time constant almost the same as the rise time of the ringing (~200ns).

(Update: I later realized I was measuring the rise time after installing the snubber. A close measurement of the unsnubbed circuit revealed a rise time of 116ns and the final snubber R value produced an RC of 426ns. This roughly 4x $\tau$ value seemed to be "bigger enough" :)

Prescribed procedure

1. Add capacitance across switch node until ringing frequency is reduced to 1/3 original value (some say 1/2). This determines original C is 1/8th of added value (total C is 9 times original, sqrt(9)=3).

2. Calculate characteristic impedance of new LC tank, sqrt(L/C). Use resistor of this value.

On a close second reading, I discovered the sources disagree on which values to use for the characteristic impedance calculation. [1] Todd recommends using the original $C_{parasitic}$. Another uses the new total $C_{parasitic} + C_{snubber}$. The optimal value I found is nearly right in the middle of these two (209Ω < 430Ω < 670Ω).

Resources:

• I always go for 2 and make a reality check in spice how much power will be dissipated. If your LC frequency is too close to the switching frequency, this will waste a lot of power. Forced CCM comes to mind too. – winny Jul 16 '16 at 9:47
• @winny: I get 16mW for the original values and 53mW for the critically damped case. So I do see the power cost involved now that you mention it :) – scanny Jul 16 '16 at 9:57
• Where are you measuring you voltage? Output of switch ripple always has that kind of waveform, this is why LC circuit is used to eliminate rippling that is supplied to output – Artūras Jonkus Jul 16 '16 at 10:08
• @ArtūrasJonkus: The two scope shots are on the switch node. – scanny Jul 16 '16 at 10:10
• I'm voting to close this question as off-topic because you have lied all the way through about this being a genuine TI part when it's an ebay fake. – Andy aka Jul 18 '16 at 7:39

Basically this is a non-problem. You don't need to do suppress the free-wheeling ringing due to the inductor and drain-source capacitance of the open-circuit MOSFET (more likely than the diode) in a buck converter because nothing bad happens if you leave it alone. The voltage in both polarities is never bigger than the voltage due to switching so the MOSFET cannot harmed by it.

It's totally different on a flyback converter of course but this is a non-synchronous buck converter.

Just regard it is a little bit of energy that could not make it's way to the output.

After-thought - if you were really clever you might be able to find a way of harnessing these oscillations and feed that energy back to the input capacitor. That would certainly be a step in the ecologicial right direction.

The device data sheet has this circuit for reducing output ripple by up to 10 times: -

The LC network will rely on the ESR of the capacitor added being as low as possible and quite probably if you are using "any old" capacitor in the Cout position (ref diagram above) then its ESR and ESL will be poor. TI are not recommending a snubber to reduce ripple!

• Except for EMC. – winny Jul 16 '16 at 10:28
• The LM2576 uses a BJT (Darlington) for the switching, although I take your point. The ringing might not hurt the switch, but it does couple to the output (2.5MHz, about 15mV p-p). In my case, it's my first chance to design a real-live snubber, so I'm keen to make the best job of it I can and learn as much as possible from the exercise. I understand the question might be academic, but, well, I'm academizing at the moment :) – scanny Jul 16 '16 at 10:32
• @winny - don't you think that the fast rising edges of the switcher cause much more EMI than the relatively slow moving ringing? – Andy aka Jul 16 '16 at 10:45
• @scanny - a MOSFET that clamps the back emf instead of a diode will do the trick but then you have a synchronous buck converter. Adding a passive snubber can work but it also significantly lowers the efficiency of the buck conversion - each time the "transistor" hard switches on it dumps a finite amount of energy into the resistive part of the snubber thus you degrade the efficiency. – Andy aka Jul 16 '16 at 10:48
• @scanny if you get output ripple problems due to the ringing then you probably need to consider how the hundred or so pF parasitic capacitance is managing to to shift the output voltage around to the tune of 15 mVp-p. This would be my objective. Maybe the output capacitor is not that good or maybe your scope probes are giving a false reading (very common in switchers). – Andy aka Jul 16 '16 at 10:51

OUTCOME REPORT:

Okay, I think I've sucked all the learning juice out of this exercise, thanks to @AndyAka and @winny for their help and encouragement :)

Here's before the snubber, switch-node in yellow, output in blue:

And here it is after:

The values I used were 1nF and 470R for anyone just looking for a recipe to try :)

My key takeaway interpretations are:

• Snubbing is to improve waveforms, not make them perfect. It's always a trade-off and getting to "tidy waveform" is going to cost more power/efficiency than it's worth.

• Snubbing is not an exact science; the analytics just get untenably complex. Be satisfied with the calculations getting you close, then simulate or solder in a variable resistor to get the rest of the way.

• Decide on a capacitor value, then tune in the resistor, letting the calculations guide your starting point. The resistor matching the characteristic impedance of the tank is what absorbs the energy, which is what stops the wiggles. So this step is where you achieve the optimal waveform. Some overshoot may remain; you're going to want to learn to live with that :)

• If you really can't live with the overshoot, you're going to need a bigger capacitor. The downside is that power dissipation (and consequent decrease in efficiency) scales proportional to the capacitor size. If you do change the capacitor, you'll need to tune in the R value again, but it probably won't change drastically.

• Do take the time to measure the rise time of the ringing waveform at the start. When comparing it to the RC ($\tau$) of the snubber, $\tau > 3t_{rise}$ is about minimum, 10x may be the upper limit. Mine was 4x and that worked fine.

Thanks again to @Andy and @winny, this was a really useful "lab" and I learned a lot :)

If there are any additional details that might be helpful let me know and I'll add them.

• You are welcome. Also, you seem to have single handedly learned my second best EMC trick up my sleeve from 8 years of EMC compliance work so you have a bright future cut out for you :-) – winny Jul 17 '16 at 18:28
• I would just like to add that in my opinion the snubbing you have done does nothing to improve the EMI on the output - this is dominated by the main pulse being turned into a ramp-up and ramp-down - it has roughly the same rise and fall time as the minute-in-comparison wiggle due to inductor resonance with what is almost certainly a MOSFET (given that you were deceitful about the un-TI part used). Please scanny, don't play tricks like this again - I feel you went too far and I feel a little mugged by this and I have to down vote your question and this answer.... – Andy aka Jul 18 '16 at 7:22
• .... to show my concern at you deliberately hiding the true facts. – Andy aka Jul 18 '16 at 7:22
• @Andyaka - I am so sorry about that, I had no intent to deceive, as far as I know the innards are actually an LM2576, I've found no difference in behavior, although of course you're right there's no real way to know. On page 14 of the LM25*96* datasheet (as this part is marked) mentions specifically [During DCM] "a small amount of energy can circulate between the inductor and the switch/diode parasitic capacitance", which is why I focused on the diode. But I guess I don't see the reason why it matters where the capacitance is coming from, the snubbing works the same. Am I missing something? – scanny Jul 18 '16 at 7:42
• How can you say you had no intent to deceive when you clearly state in the amendment to your question that it runs at a vastly different speed. I'm not talking about this anymore. – Andy aka Jul 18 '16 at 7:46

Snubber starts working after the loop is optimal thereby correcting inductors low quality at the purity level.
Consider improving the current-phase peak within the capacitive loop (adding or substracting a couple microhenries to your equation results) before softening the output harmonics. thus driving a specific load capacitance for an instance.

Find any capacitor with the closest to zero ESR. Recoil using thiner copper or special purity copper-alloys, find a more specific ferrite.

• Stacking of words without mutual meaning? Errr, can you clear that up? Perhaps with periods at end of sentences and capital letter at the start of the same? – winny Mar 29 '17 at 19:36