A machine-learning model may be used to estimate a film thickness from a spectral image captured from a semiconductor substrate during processing. Instead of using actual measurements from physical substrates to train the model, simulated images may be generated for a wide variety of predefined thickness profiles. Simulated training data may be rapidly generated by receiving a film thickness profile representing a film on a semiconductor substrate design. A light source may be simulated being reflected off of the film on the semiconductor substrate and being captured by a camera. The spectral data captured by the camera may be converted into one or more images for a wafer with the film thickness profile. The images may then be labeled with thicknesses from the film thickness profile for training a machine-learning model.