Predicting curcumin release kinetics from nanocarriers using a physics-informed machine learning framework.

BACKGROUND

Curcumin nanoformulations display highly heterogeneous release behaviors, yet the field lacks an integrated framework capable of predicting kinetics across diverse material classes. However, the diffusion parameters governing these systems have remained previously unresolved due to inconsistent experimental conditions and limited datasets.

AIM: Here we establish a physics-informed machine-learning framework that predicts curcumin release kinetics with unprecedented physical consistency across heterogeneous nanocarriers.

METHODS

A curated dataset of 75 formulations was encoded using 13 physicochemical descriptors, and release kinetics were parameterized using a two-component diffusion model. A Physics-Informed Neural Network incorporating explicit monotonicity and non-negativity constraints was trained and benchmarked against Random Forest, XGBoost, and a multilayer neural network. External validation used 31 independent formulations.

RESULTS

The model demonstrated 0.92 R², 2.15 MSE, and 98 % physical consistency, exceeding baseline methods by up to 1.6-fold. External validation achieved 0.885 R² and mean absolute errors < 0.08 across release-time distributions. Design-space mapping uncovered previously inaccessible parameter regions associated with rapid, balanced, and sustained-release profiles. Application-specific performance analyses showed 0.87-0.91 R², mean absolute error 2.10-4.30 h, and physical consistency 94-98 %.

CONCLUSIONS

This framework resolves key kinetic parameters across diverse nanocarriers with high fidelity and provides a new foundation for data-driven formulation engineering. The approach enables predictive optimization of curcumin delivery systems and opens the door to automated design of controlled-release nanomedicines.