CompactAI: A hybrid intelligent system for surrogate stress–strain prediction in pharmaceutical tablet compaction

Accurate prediction of powder mechanical behavior is essential for pharmaceutical tablet manufacturing, yet conventional experimental campaigns and discrete element method (DEM) simulations remain costly and computationally intensive. This limits rapid formulation screening, parameter exploration, and simulation-assisted decision support. We present CompactAI, a structured deep-learning surrogate framework for DEM-based stress–strain prediction in pharmaceutical tablet compaction. CompactAI predicts loading and unloading responses under uniaxial compression from four physically interpretable DEM descriptors: loading stiffness, unloading stiffness, adhesion stiffness, and plasticity depth. The framework combines neural coefficient prediction, equation-based curve reconstruction, and collapse classification for catastrophic unloading cases associated with severe plastic deformation. We evaluate CompactAI on CompactAI-26, a DEM-generated dataset of 4796 tablet-compaction simulations, using held-out testing, five seeded train–validation–test splits, baseline comparisons, objective ablations, and curve-level error analysis. CompactAI achieves 0.1023 nMAE for loading reconstruction and 0.1949 nMAE for non-collapse unloading reconstruction on the original held-out test set, while producing end-to-end predictions in approximately 0.593 ms, compared with DEM runtimes of roughly 20 min to 2 h per simulation. Across the five seeded splits, CompactAI outperforms mean-curve, nearest-neighbor, and PCA–ridge baselines for loading reconstruction. Unloading reconstruction remains more challenging, with nearest-neighbor retrieval serving as a competitive local-interpolation baseline. These results validate one CompactAI surrogate instance for a single DEM configuration and illustrate a lightweight framework for instantiating configuration-specific surrogates from DEM-generated tablet-compaction datasets. CompactAI-26 is provided as a benchmark dataset for future work on predictive tablet-compaction simulation.