Over the past year we have built on a previously developed mechanistic model of granuleswelling and disintegration behaviour, with the aim to create a practical, fast tool for predicting and designing wet-granulated product performance. This work integrates physics with machine learning so that routine formulation inputs can be turned into reliable performance curves in seconds.
Core Models
• Mechanistic Performance Models (Single-granule Swelling → PBM Disintegration). Coupled single-granule swelling with population balance disintegration model to predict: Rp(t) (granule radius) and F(t) (released-mass/ particles-released fraction).
• Physics-Informed Neural Network (PINN). A neural network implementation of the same physics that preserves mechanistic meaning while enabling efficient learning across formulations.
• ANN-Based Parameter Mapper. A supervised neural network that ingests standard formulation descriptors (e.g., L/S ratio, %SSG, filler type, %HPMC, initial porosity, skeletal density) and predicts the mechanistic parameters required by the Mechanistic model.
The workflow (Figure A) has been implemented and delivers rapid, physics-based forward prediction by converting formulation descriptors into mechanistic parameters, then full disintegration performance curves. Following further experimental validation, it will be ready to incorporate inverse modelling. Once incorporated, this inverse model will optimize formulation inputs and granulation conditions to give a desired released fraction at specific times.
modelling. Once incorporated, this inverse model will optimize formulation inputs and granulation conditions to give a desired released fraction at specific times.