We discuss the methodology of physical resist model calibration for a rigorous lithography simulator under various aspects and assess the resulting predictive accuracy. The study is performed on an extensive optical proximity correction (OPC) dataset, which includes several thousands of critical dimensions (CDs) values obtained with immersion lithography for the half-pitch technology node. We address practical aspects such as speed of calibration versus size of calibration dataset, and the role of pattern selection for calibration. In particular, we show that a small subset of the dataset is sufficient to provide accurate calibration results. However, the overall predictive power can strongly be enhanced if a few critical patterns are additionally included into the calibration dataset. Also, we demonstrate a significant impact of the illumination source shape (measured versus nominal top hat) on the resulting model quality. Most importantly, it is shown that calibrated resist models based on a 3-D (topographic) mask description perform better than resist models based on a 2-D (Kirchhoff) mask approximation. Also, we show that a resist model calibrated with one-dimensional (lines and spaces) structures only can successfully predict the printing behavior of two-dimensional patterns (end-of-line structures).