Methods and systems for robust overlay error measurement based on a trained measurement model are described herein. The measurement model is trained from raw scatterometry data collected from Design of Experiments (DOE) wafers by a scatterometry based overlay metrology system. Each measurement site includes one or more metrology targets fabricated with programmed overlay variations and known process variations. Each measurement site is measured with known metrology system variations. In this manner, the measurement model is trained to separate actual overlay from process variations and metrology system variations which affect the overlay measurement. As a result, an estimate of actual overlay by the trained measurement model is robust to process variations and metrology system variations. The measurement model is trained based on scatterometry data collected from the same metrology system used to perform measurements. Thus, the measurement model is not sensitive to systematic errors, aysmmetries, etc.