Data-efficient prediction in tableting using word embeddings and empirically-guided neural networks

The development of robust oral tablet formulations remains time-consuming, often limited by scarce data and the difficulty of incorporating categorical formulation variables into predictive models. Traditional regression methods are interpretable but struggle with nonlinear interactions, whereas modern machine learning approaches offer higher predictive power at the expense of transparency. In this study, we present a neural network framework that employs word embedding layers to represent categorical formulation factors, such as active pharmaceutical ingredients (APIs), as trainable semantic vectors. These embeddings are integrated with empirically-guided output functions and a deep ensemble strategy to predict tablet quality attributes, including tensile strength and density as well as ejection force, and dosing height, based solely on formulation composition, compression pressure, and tablet weight. The model achieved predictive accuracy comparable to or exceeding classical regression while reliably avoiding physically implausible outputs. Analysis of the learned embedding vectors revealed meaningful clustering of APIs, enabling transfer learning across materials and robust predictions even for APIs with few or no training data. Furthermore, information gain analysis demonstrated that low-concentration formulations can substantially enhance predictive accuracy, supporting more material-efficient experimental designs. These results highlight embedding-based, empirically-guided neural networks as explainable and practical tools that could accelerate pharmaceutical formulation development in the future.