Lithography process modeling is critical for effective model-based optical proximity correction (OPC) or verification. Physics based full resist and etching model can provide very accurate prediction of the resist profile, but its speed forbids the use in practical production OPC and verification applications. Simplified models have therefore been developed. These models collapse some complicated but less crucial physics into "parameters" which are tuned to best fit the real measurement data. However, as the feature patterns vary, the aerial image around the patterns can experience a wide range of intensity distribution patterns. It is difficult to use a single set of "parameters" to fit into all these profiles. As compromises are made, accuracy suffers. The properties that contribute to such variations are primarily pattern shapes, dimensions, and in the case of phase-shift masks, phase-interaction. One way to improve the model accuracy is to build multiple "local" models such that each model contains a set of parameters that are optimized for the given pattern. As we perform simulation, we identify the pattern and then pick the model that is best suited for the given pattern. In this paper, we demonstrate how it is difficult for a single model to fit a set of data with large varying patterns. Then we show how multiple model methodology can be applied to improve model accuracy. As we apply the models, there will be "gray" areas where the pattern is not clearly identified to belong to the class for which a model is available. We explain how such situation should be coped with, and how the simulation responds to model "switching".