|
Lithography simulation tools eliminate costly and time consuming experiments allowing new processes to be developed quickly. There are excellent simulation programs that allow sophisticated modeling of the optics in current and future lithography tools. In many instances, the weak point in lithography simulations is the relatively poor capability to model resists. Sophisticated and accurate models have been developed for many technologically important i-line resists. However the models for 248nm chemically amplified resist are not as mature, and there are many resist of interest for which there are no reliable models. Even when they do exist, these full resist models are computationally expensive and not suitable for some applications such as model based optical proximity corrections. When useful models do not exist, lithographers use the aerial imaging portions of the lithography simulation tools and apply the simplest of resist models, the so-called constant threshold model. While this allows the critical dimensions to be approximated for high contrast resist, it fails to capture important aspects of most resist processes. Empirically trained resists models have come to be used where more accurate lithography simulations are required, but full resist models either do not exist or are to slow to be useful. This paper explores the use of a class of empirically trained models known as variable threshold resist models. This type of model stats with an aerial image calculation and uses a function to locally vary the threshold used to predict CDs. This type of model may be quickly trained for a specific resist process and potentially applied for a wide range of numerical aperture and partial coherence settings. We show how multiple dose and focus data can be used to train a model that includes input parameters extracted from the aerial image as well as pattern factors and exposure dose. The data present suggests that models trained with one set of optical conditions are useful at other optical settings. We also explore different approaches to validate the models and demonstrate some consider the effect of across wafer variation on the training data.
|
