AI framework could enhance continuous manufacturing, say FDA researchers

As the FDA continues to advance its AI strategy, researchers from the agency’s Center for Drug Evaluation and Research (CDER) have proposed an AI-based digital twin model that enables a simplified continuous direct compression (CDC) line process.

The development could advance continuous pharmaceutical manufacturing and support regulatory assessment. In January 2025, the FDA unveiled its first guidance on utilising AI for drug development, with the aim of enhancing the credibility of AI models in product regulatory submissions.

Zhao et al. highlighted that evaluating the credibility of these models for continuous manufacturing and “accurately quantifying their impact on product quality” remains challenging.

The present study applied a two-layer Neural network predictive control (NNPC) model trained using data from the digital twin to predict system outputs, utilising a residence time distribution theory.

The researchers combined the trained neural network with an optimisation block to adjust control signals and minimise tracking error and control effort.

The AI-based NNPC model demonstrated “remarkable performance in setpoint tracking and disturbance rejection for the simulated CDC line”, outperforming to conventional proportional-integral-derivative (PID) control.

Data showed the NNPC achieved zero overshoot and superior disturbance rejection (≤1.6 percent deviation).

The findings highlight the potential of AI-based advanced process control (APC) to manage the complex, nonlinear system dynamics in continuous manufacturing. Overall, this could enhance product quality and regulatory assessment.

However, obstacles still remain, according to Zhao et al. Firstly, the industry lacks a “standardised mechanisms for neural network model verification and validation”, making regulatory submissions challenging. Specifically, the model investigated in this study would need to be assessed for its capability for real-time control.

The research was published in Computers & Chemical Engineering.