Download as doc, pdf, or txt
Download as doc, pdf, or txt
You are on page 1of 3

Criteria for assessment of a model 1.

Does it cover the desired range (typically process window +/- 50% for all parameters)? a. This is determined based on the data that the model was built on and is the main purpose for the DOE. However, if the process changes (and well discuss what is meant by that below), then the model may have to be redone. 2. Does it give sufficient model quality feedback? a. For a normal model, this is done by used of the Model Quality metric, where a higher number is a worse fit for the model. The exact cutoffs are determined separately for each model. b. Further, model quality is typically used for screening (though some other internal parameters may also be used). If the model is over-screening (to be determined by the team), then this is another indication that the model needs to be adjusted. 3. Is the defined error acceptable? a. This is different for each application. In general, we want a TMU (total measurement uncertainty) that is comparable to the reference error. b. Reference error: we take both the 3 sigma error in the actual measurement AND the 3 sigma variation within the macro (as this is what the MBIR sees) and use that as the reference error. This is important because the within macro variation has a big impact on the quality of spectra (and in turn the quality of the model) c. TMU: TMU is the standard metrology method for comparing two methods when there is error in the reference value. It is superior in many ways to the R^2 method. It returns a 3 sigma error value in the units of the measurement. What triggers the need for a new iteration of a mode? 1. In general, if the above criteria are not met, then a new iteration of a model is needed. a. If the DOE does not meet the full desired range either because of inability to hit the full range when the wafers are created or some of the spectra may be unusable. b. If the process changes in a way so that the spectra upon which the model was built no longer represents what the spectra looks like.

i. E.g. Shifting the process center, increasing the size of the BOX, etc. will cause this c. Also, if the model is built on spectra that represents a highly variable process (for example, if the variation within the macro of the trenches is high), then making them less variable will fundamentally create different looking spectra. This will require a new iteration. i. To be more specific, spectra with a lot of variation are inherently flat (there are no real fringes). It is difficult to build a good model on this and the error in the model will match the variation within the macro ii. Spectra with a small amount of variation are inherently fringe-y and it is relatively easy to build a good model. Again, the error in the model will match the variation within the macro. d. If the model is screening too many data points (that is, the model quality is too high on too many spectra), then this is a signal that a new model needs to be built. Best wafer practices for building an initial model (or a substantial update) 1. A wide DOE set of wafers is needed covering the process range +/- 50% in all important variables (not just those to be reported, but all of those that will see variation should be generated and spectra taken on the tool. This will include the POR steps 2. These wafers will also need to get reference measurements (xSEM or AFM typically) in as many die as possible matching the MBIR measured die. For AFM, we run through the AFM step. For xSEM, the cross-sections must be done on the measured macros. Close macros are not good enough and this data cannot be used. 3. As many different products as possible should be used. This helps account for things like pattern factor, closeness of external features, and other variables that have been known to affect whether or not the model works properly. In the past we have had issues with building models on only one product and then having them not work on another product.

Updating an existing model It is important to see that the model building process is an iterative process, developing along with the technology. The reasons for this include that the process changes in significant ways during the technology development and this changes the

shape of the spectra that are gathered on the tool. This includes not only conscience process changes (like thicker BOX), but also changes that affect the within macro variability of the features. Further, because the process starts out as inherently immature and highly variable, the early models will have significantly worse quality than later models. 1. A broad range of data should be collected for updating an existing model. If only a tweak is needed, as determined by the team, then fewer spectra are necessary. However, if it is a major update, then it might be necessary to have a broader span of wafers done (perhaps another DOE). 2. This, also, should be done on as many different products as possible.

You might also like