Predicting the tabletability of binary mixtures from individual powder compaction behavior.

Successful development of direct-compression formulations can be hindered by poor compactibility of active pharmaceutical ingredients active pharmaceutical ingredients (APIs), necessitating rational formulation strategies. This work presents a neural network model trained on a large tableting dataset comprising over 200 formulations prepared from 33 powders, including 17 APIs, to predict tablet tensile strength across the full compaction pressure range directly from material properties and formulation composition for binary mixtures. The predictive accuracy of the neural network model was compared to a mixing rule based on tabletability parameters. The neural network outperformed the power-law mixing rule for binary mixtures, demonstrating particular strength for APIs with poor compaction behavior and for mixtures where the mixing-rule approach could not be applied due to the inability to produce intact compacts from pure components. The model also exhibits permutation invariance and only requires 3-5 g of material for the compaction characterization for a new powder, necessary for making model predictions.