Using shape constraints to incorporate extrapolation capabilities into the regression models of a hybrid approach for predicting the polymer melt flow through twin-screw extruder kneading blocks

• Expert knowledge is included into symbolic regression based on genetic programming. • Shape constraints were formulated with the expert knowledge. • The shape constraints were assessed with twin-screw extrusion modelling data. • The inter- and extrapolation performance of the regression models was improved. • More stable models and no points of discontinuity with shape constraints. The fully intermeshing co-rotating twin-screw extruder is in common use throughout various industries, for instance, in polymer, pharmaceutical, and food applications. The excellent mixing capabilities of this extruder type are attributed to its kneading blocks. Knowledge of the conveying mechanism of these elements is essential for designing extruder screw configurations. The flow behavior of twin-screw extruders is commonly approximated using analytical modeling approaches that are based on a substantially simplified geometry. We, however, avoid this by building on a hybrid approach that combines analytical, numerical, and data-based modeling methods. We demonstrate that incorporating expert knowledge of physical boundaries and correlations enhances the extrapolation capabilities of our regression models. To this end, the example of dimensionless drag-flow capacity was used, and shape constraints were included when training the regression models based on genetic programming. The results confirm that this knowledge integration allows a more accurate prediction of phenomena outside the trained parameter space.