Manuel Borja

Department of Data Analysis and Mathematical Modelling, Ghent University, Coupure links 653, Gent, 9000, Belgium

This study introduces a neural network–based method for the estimation of tablet quality attributes using raw material properties, blend ratios, and processing conditions as inputs. Attention is given to model interpretability and to how to extract insights from the model, as well as to how to combine traditional data-driven neural network approaches with expert knowledge (in the form of some pre-established monotonicity rules) to create a hybrid model that is more trustworthy and reliable. This model was applied to a dataset of direct compression experiments, and its performance was benchmarked against existing methods that are purely data-driven. Additionally, it is demonstrated how this method can be applied to assist formulation and process development in a practical setting, with the aim of reducing associated costs and timelines.