Modelling the effect of base component properties and processing conditions on mixture products using probabilistic, knowledge-guided neural networks

Development of materials by mixing different base components is a widespread methodology to create materials with improved properties compared to those of its base components. However, efficient determination of the properties of mixture-based materials during design remains challenging without prior knowledge of the underlying physical phenomena. In this work a new data-based methodology is proposed involving the use of probabilistic, knowledge-guided artificial neural networks to jointly model the properties of the base components, the proportions in which they are mixed, and the processing conditions used during manufacture to predict properties of final products. The method proposed does not involve any assumptions in terms of ideal mixing rules of the base components, and allows for estimation of aleatoric uncertainty in the prediction. Additionally, an extension is presented that incorporates expert knowledge into the model by the implementation of monotonicity constraints between certain inputs and outputs. The methodology is illustrated with a case study involving the formulation of drug products using direct compression. The model is used to predict pharmaceutical tablets’ quality attributes (mass variation, tensile strength, disintegration time, friability and ejection force), showing that the method is able to predict properties of the final product overcoming gaps currently present in previous modelling approaches. [Display omitted]