In this paper, we propose a new model construction method utilizing application specific physical information and present its application to SRAM yield calculation. The physical information is extracted as statistical distributions from past simulation results including not only similar type of SRAM's but also completely different type of SRAM's. Experimental results show our method achieves 700x speed up over non modeling method and more than 10x speed up over the conventional modeling method. It requires only 5.3 samples to model fifth order full cross term polynomial with 21 coefficients and free from over-fitting and singular matrix problem. This modeling method can be a general approach to create models with application specific physical information.