Some follow-up Q&A to Steve's device benchmarking report
Q: Regridding tool: Are the capabilites of pdMesh that different from devedit(SILVACO) or from MESH, MDRAW(ISE) or from the TMA-tool? A: The main thing I noticed about pdMesh was its relative ease of use when you have a good command script. My work with devedit, MDRAW, and others has been very time consuming by comparison, and the grids do not tend to be nearly as efficient or smooth. For the benchmarking, the pdMesh was not used in the Atlas structures, although the currents were not really affected by the different grids. Michael: pdMesh also has the benefit that when a given grid spacing is requested in pdMesh, that is the spacing pdMesh generates. Other programs seem to take user defined grid spacings as merely a suggestion. Also, the gate oxide interface can be reproduced exactly with pdMesh. Q: Transfer characteristics: Why didn't you also compare Id/Vg curves? A: The subthreshold behavior was used to set the workfunctions in the different simulators to make sure the devices were as close to the same as I could make them. The reason this information wasn't included in my report was a space consideration mainly. My thesis will contain IdVg comparisons as well (used to show similar behavior below threshold). Q: Parameter Consistency (Drift Diffusion): Besides the comparison of default parameters, were the parameters of all the DD Lombardi models identically set in the different simulators to check whether results are comparable? A: Yes, the default parameters were the same across simulators (matching the published parameters). However, only the default parameters were used across different simulators, the suggested parameter set was only used in one of the packages. Q: Model Consitency (Drift Diffusion): Are all the different Lombardi models comparable in the different simulators? The physical way to parameterize the high field saturation within DD-simulators is to use the gradient of quasi-fermi level: mu = mu(grad Phi , ...). This is not the case in Silvaco's simulator Atlas. The field parallel to current flow is used there, which is believed to be historical ballast from PISCES. The ISE code has both possibilites, ie mu(grad Phi) OR mu(E_parallel). I don't know the situation in MEDICI. The latter parametrization is unphysical because it results in a large mobility reduction within a diode with zero or nearly zero forward bias, because E_parallel is large, though the driving force (electric + diffusion force) is small. Though the resulting currents may not be influenced too much is this case, there might be differences in other cases. A: Medici also uses the parallel field for its mobility reduction. I did not see much difference between the two methods in ISE on the 250 nm device (comparing only on-current, so this is a very limited answer, there is definitely room for more investigation, but my research is taking me elsewhere). Follow-up comment from questioner: Maybe results do not change very much. But we have the suspicion that convergence behavior is degraded by the E_parallel parametrization, especially for small currents when the direction of current flow is not well defined in some device regions. If anyone had the time and all the tools to compare convergence behavior of the simulators systematically, I would be very interested in the results. Q: Mobility model selection: Though I personally do not have experiences with alternatives, I was told that the conventional models like Lombardi are insufficient for large channel doping levels. For this case it seems to be important to not only model the dependen cy from high field, from E_eff and from doping, but also from carrier densities. Especially when the region of the universial relationship for the channel mobility is left, E_eff and doping are not sufficient to model the mobility. The screening of impurities by free carriers and thus the strength of impurity scattering depend on free carriers. I know of improved results in a research simulator of an experienced simulation reserch group after implementing the model after: S.A. Mujtaba and R.W. Dutton, Semi-Empirical Local NMOS Mobility Model for 2-D Device Simulation Incoporating Screened Minority Impurity Scattering, NUPAD Tech. Dig., vol.5, Honolulu, pp. 3-6, 1994. Maybe this model is similar to the one you mention (Darwish). But the main question is: Did you perform at least test simulations with improved models that also account for the screening of impurity scattering? At least Klaassen's model that can be selected in ISE's Dessis claims to account for this effect. A: This is the reason for comparing more than one mobility model on the two different devices. We know that all of today's mobility models breakdown for the high doping densities that will be required if we stay with the SIA roadmap. What was found in this work was that we can't just blindly extrapolate any of these models to the higher doping densities, since the models extrapolate differently. [The following is a series of comments from the questioner in chronological order that pertain to advanced DD mobility models] C: I wrote that Klaassen's model is in a sense more sophisticated than Lombardi (eg screening of impurities). But of course the pure Klaassen it is not sufficient for your simulation since it has no E_eff-dependency for inversion channels. Silvaco's Klaassen model cannot be used together with Silvaco's CVT model, because CVT overrides any other mobility specification. Maybe Silvaco's C-interpreter for mobility models can be used for a workaround. From ISE's manual I conclude that a combination of Klaassen and of an E_eff-dependent model is possible (similar to eq.(1) and (2) in Darwish). Of course a combination of Klaassen and an E_eff-dependent model is required, as proposed in Darwish et al.. ISE's simulator claims to be capable of modeling the mobility according to Darwish by just setting some parameters (see manual of DESSIS, release 5, page 16-71). In the meantime I was told that although Mujtaba,Dutton is better than Lombardi for higher doping levels, it is not satisfactory for highest dopings. Darwish et al. state something similar, so maybe Darwish et al. is the best reference for a more advanced mobility model. Q: Model Consistency (Drift Diffusion vs Energy Transport or Hydrodynamic): In my experience it is useful to check model consistency (DD vs ET) on a homogeneous level (only field current, ie small diffusion forces) before comparing device results. Sometimes even for bulk silicon DD results differ from ET results, because the high field saturation is not always being "translated" correctly into a mu(T_carrier)-relationship. Did you perform some checks for the ET model? (You mentioned the inconsistency only for the HD model). A: The HD comparison is a fairly limited comparison because I was only able to get UT-Minimos to run its HD with any regularity. Because UT-Minimos is a fairly closed simulator (in my opinion, it is very difficult to adjust parameters), I simply ran the test devices with the default settings. Q: Usability of ET models and HD models: You decided not to use the ET-option of Silvaco (convergence problems). Did you try the ET model of another simulator? Do you think a DD-simulation can still be predictive for such small devices? At least for substrate currents an examination of the need for ET models could be interesting. A: Yes, I also tried to use the ET models in Medici (similar problems) and Dessis (didn't have enough time to give it a fair shake, and decided to not include it since I had troubles with Medici, figured it would probably be the same story here). You use the word "predicitve" which I would call some question to. Personally, I don't really see a situation where TCAD is able to get ahead of actual silicon, so I am going to take the word "predictive" to mean accurate (Do I think a DD-simulation can still be "accurate" for such small devices?). In this case, from what I have seen, the mobility and parameters still have an effect of these small devices. So yes, I think it is within the realm of reason that DD can still function at that level, with a lot of effort to "calibrate" the models. Steve Bourland Nov 2, 1998 Reply from Silvaco: Nov 4, 1998. In ATLAS for mobility models combining KLA and an E_perp model you can use the Shirahata model along with Klaassen's model. These two are designed to work together. As seen in our example (mos2ex14) the results from KLA SHI and CVT are similar.
This page last updated Nov 2, 1998 by