Monte Carlo simulation: there are two peaks

That's an impossible question to answer without more info - probably seeing the testcase would make it easier (so customer support would be the right avenue).

My first impression is that your measurement has only discrete levels - the measurement on the left has values which are exactly 161, 164 or 167. The measurement on the right has values which are only 138 and 141. These look remarkably exact and you can see all the vertical circles are exactly in line (these are the individual results from each measurement). So I wouldn't ever expect a gaussian response when you have a small number of discrete outputs which are possible (it might be Poisson depending on the test, but even that doesn't sound right in this case). So what is your measurement computing?

Regards,

Andrew.

Hello Andrew,

Thanks for your answering!

I'm simulating a temperature monitoring schematic. In above simulation curves, the left curve is hysteresis high of overtempt (overtempt is output signal), and the right curve is hysteresis low of overtempt.

The schematic diagram is below:

From the schematic, you can see that, a resistors divided reference voltage compares with a negative temperature coefficient voltage, then following after two invertors to get overtempt.

You said my measurement has only discrete levels. I think it's because of the invertors and comparator.

But why the points on right side curve distribute like a two peak?

Or for the sampling method, I selected 'random', do you think 'low-discrepancy sequence' will be better?

Nelson.

Thanks!

Hi Nelson,

First of all, I doubt that changing from random to low-discrepancy sequence would affect this significantly. That will just mean that the the variation of the statistical parameters affecting the devices would be smoother for smaller numbers of samples, but I doubt it would make a massive difference to the distribution of your output measurements given that you have a reasonable number of samples already. I'd generally recommend LDS anyway over random, but the issue here is that what you're measuring appears to be discrete.

It's still not clear what you're measuring. If it was the output of the inverter, then I'd expect that to be one of two values (1 or 0 - scaled by the voltage), but presumably you're measuring something else. When you say you're measuring the "hysteresis high of overtempt" it's not clear what you mean by that (I certainly don't know what you mean). What is the expression you're using? It can't just be the output voltage because the y values are too high.

If the output is some kind of discrete measurement I wouldn't necessarily expect the circuit response to be gaussian. You don't really have "two peaks" you just have two values in the result measurements. The density estimator is meaningless in this case - you can see the y values (the red circles) are all lined up for two different values - so it's not two "peaks". If you looked at the Detail or Detail Transpose view it would be clear that all your measurements have one of two values.

Regards,

Andrew.

Hi Andrew,

My explanation for above schematic is not very clear.

I do DC simulation in Monte Carlo sample, and sweep temperature from -40 to 175C, and from 175C to -40C.

The node voltage of diode in schematic changes following the temperature value.

If the temperature is increasing from -40 to 175C, and it gets to high enough like 164C, the comparator can detect that the voltage of diode is higher than resistor divided voltage, then the comparator will output high, this temperature (164 C) is "hysteresis high of overtemp";

If the temperature is decreasing from 175 to -40C, and the temp is lower than like 138C, the comparator can detect that the voltage of diode is lower than resistor divided voltage, then the comparator will output low, this temperature (138C) is "hysteresis low of overtemp".

The X axis in above curve means "temperature". And you can think that "hysteresis high or low" is a threshold , temp is higher than the threshold, the output of this schematic is high, else it low.

My doubt now is that the threshold should be very normal distribution, just like threshold voltage of a NMOS or PMOS which does not have just two or three discrete value in Monte Carlo simulation.  But the values in above right curve are discrete 138C and 145C, and are not dispersive from 138C to 145C.

Regards,

Nelson.

What was your step size for the temperature in the dc sweep? 

Andrew.

Hi Andrew,

My simulation step size is 3, and sweep from from 0 to 230C.

Nelson

Hi Nelson,

Perhaps the variation is small enough that it only covers a few degrees, and so a step size of 3 is far too large. Perhaps you could sweep in 0.1 degree steps and sweep from (say) 120℃ to 180℃? (just to cut down simulating over an unnecessarily wide range) to see if that improves things. It's hardly surprising that you've only got discrete peaks at 3° intervals if you've only got a step size of 3!

Andrew.

Hi Nelson,

Perhaps the variation is small enough that it only covers a few degrees, and so a step size of 3 is far too large. Perhaps you could sweep in 0.1 degree steps and sweep from (say) 120℃ to 180℃? (just to cut down simulating over an unnecessarily wide range) to see if that improves things. It's hardly surprising that you've only got discrete peaks at 3° intervals if you've only got a step size of 3!

Andrew.

Hi Andrew,

You reminded me that I ignored to check the step size in simulator and discrete step in curves! Thank you so much!

I am simulating with much smaller step size,  and I think it will be work. Thanks.

Nelson