Why Kirchhoff's Current Law is not satisfied at my circuit?

I'm sorry, but that's a ridiculous goal. Your model won't have anywhere near that accuracy - and if it's not modelling leakage, what is the point?

You can get it lower with conservative, reltol=1e-5, gmin=1e-15 (500fA spike - the flat parts are about 12fA), but I cannot see any practical example with made up models where you really would expect this level of accuracy. That's not what circuit simulators are designed for.

BTW, you could also (with this simple circuit) set the gmin to 0, and then the flat parts are rather spiky due to numerical accuracy, but peak at 13aA (and is much lower than that the rest of the time). I would still question the meaning of such a simulation though.

Setting reltol to 1e-9 is not going to be a good idea, because you'd be running out of numerical resolution - given that you only have about 14-15 digits in a double precision floating point number, you're expecting to resolve the error by using 9 of those digits, plus there needs to be accurate derivatives computed and so on - that's asking for trouble. It's not just about simulation time - it's just that you're asking for accuracy when the model itself is probably a much coarser fit than that.

Andrew.

Thank you for your valuable reply, Andrew.

Leakage current of my transistor (IGZO TFT) is very low. (1e-21A/um).

And, I am simulating a situation where the charge stored in the capacitor is drained by the leakage current.

Therefore, I need very low error value.

But, you said that 13aA error value is impossible goal.

Why it is impossible?

It is because of the limitations of the simulator?

Or is it because of the limitations of the model (RPI-aTFT model)?

Or is it just because of the time it takes to simulate?

Please tell me.

Thank you 

yysunj

The issue is that you have a very high dynamic range - the normal currents in the circuit (which I assume might be of the order of 1uA) are 1e15 times higher than your leakage current is supposedly. If your model is really modelling that level of leakage current, it's going to be very difficult to resolve. Simulators use numerical analysis techniques to iterate towards a solution - and have to determine when they are close enough (within an acceptable error tolerance). That is usually (simplifying it for here - for more details you might want to look at The Designer's Guide to SPICE and Spectre by Ken Kundert or Steady-State Methods for Simulating Analog and Microwave Circuits by Kundert et al.) of the form |Fn(v(k))| < reltol*Fnmax + iabstol where Fn(...) are the currents flowing into a node, and Fnmax is the absolute value of the largest current. You can set iabstol to a very small number, but it does mean that it will spend a long time trying to resolve the differences in tiny currents (iabstol defaults to 1pA in spectre, and is normally about a millionth of a typical current in the circuit to give reasonable accuracy/performance). reltol is a measure of the normal dynamic range (worse-case) in the circuit, but you can't set it above around 1e9 because you'll just run out of numerical headroom. Added to that, there's gmin which is added to provide an artificial conductance to ground to ensure that nodes are not floating - effectively floating nodes (when there's no leakage path) lead to the simulator having to solve an ill-conditioned problem (in essence the matrices become impossible to solve as there are an infinite number of solutions that meet the conditions). That too helps with convergence but degrades modelling of leakage (especially if the gmin resistance is 1e-12 Siemens, which leads to pA scale currents in circuit operating around 1V). 

So you could set iabstol very small (e.g. 1e-30) and you could set gmin=0 and for this circuit that will simulate very quickly. However, I see a pretty noisy solution in the flat parts (getting a very accurate answer in the switching parts will be difficult because the dynamic range is higher there). You could also try relref=pointlocal but that will probably be very slow indeed and fail to converge.

This is really beyond the normal accuracy criteria used for circuit simulation. I can't spend more time trying to coax the simulator into an extremely atypical usage case.

Overall, I doubt the models you're using are good enough too.

Andrew.

Thank you for your exact reply!

yysunj