The utilization of physical laws in combination with data-driven modelling can substantially enhance prediction performance of relevant characteristics along the process-structure-property-performance chain. By combining physics-based and data-driven modelling, the advantages of both modelling approaches can compensate their respective disadvantages. On the one hand, fundamental physical relationships are exploited in physics-based modelling but inherent simplifications and assumptions can lead to significant prediction errors. The careful calibration of physical laws to the particular use-case domain can be tedious even for domain experts and the validated calibration space can be too narrow to allow for any practical application since predictions outside the calibration space often carry inadmissible prediction errors. On the other hand, complex relationships that are hidden in data can be identified and utilized via data-driven modelling without the need to explicitly formulate or calibrate any physical equations; however, the data often needs to be available in high amounts and in “good-quality” to comprehensively represent the contained relationships. In a synergetic combination, predictions generated from calibrated physical laws that show discrepancies to a target solution outside its narrow physical calibration space, can be corrected by a data-driven model that represents only those discrepancies. Through this hybrid approach of combined physics-based and data-driven models, prediction errors can be significantly reduced and the calibration space expanded, even when data is scarce. In this presentation, we demonstrate successful implementations of this hybrid modelling approach for two material processing technologies: Laser Shock Peening and Friction Surfacing. In both cases, predictions show less errors and increased generalization, especially in scarce data situations. Furthermore, physics-based feature engineering, i.e. physical normalization of inputs and outputs according to a dimensionality analysis based on the Buckingham-Pi theorem, enabled predictions in an expanded physical space beyond the one used for training.