By combining electrical linewidth measurements and neural-network (NN) process metamodeling, lithography simulators can be calibrated in an efficient way. In this work we present a novel methodology for characterizing postexposure bake using a very large experimental data set, so that the calibrated model can be used as a truly predictive tool. The adoption of a special test reticle mask allowed us to collect more than 700000 critical dimensions CDs from 24 silicon wafers for a matrix of postexposure bake (PEB) time, and temperature conditions. The Lithographic patterns included isolated, semidense and dense lines for structures of 0.25, 0.20, 0.175, and 0.15 mu m nominal size replicated across the exposure field and across the wafer. As a result of this particular metrology, each measured CD was associated with both topological (position on the wafer and position within the field) and process information (exposure dose, FEB time, and temperature). Database management techniques were implemented in order to extract and analyze such a massive data set. Process metamodeling (PMM) was used for the calibration of a FEB model describing the joint effect of photoacid diffusion and photoacid loss, coupled with a deprotection reaction. PMM creates a NN model of the FEB original model (a "model of a model") so that the diffusion coefficient, the acid loss, and the deprotection rates can be estimated by inversion of the NN mapping. The comparisons between experimental and simulated data show excellent agreement that is maintained across the entire process space.