Validating the future: a guide to adaptive digital twins in GMP environments

In 2023, a VP of Quality stood before a whiteboard diagram of a new continuous manufacturing line. At the centre: a box labelled “Adaptive Digital Twin”.

“A model that updates itself during a batch?” he said. “Uncontrollable. Unvalidatable.”

Two years later, he sat across from an FDA investigator during a pre-approval inspection. The investigator looked at the twin’s real-time predictions on screen.

“Show me the validation data for the model updates,” she said.

The VP opened the Model Lifecycle Report — every parameter adaptation event, the triggering data, the verification that updates stayed within the validated design space.

The investigator reviewed it. “Good control strategy,” she said. And moved on.

The regulatory barrier most pharma teams fear rarely materialises when the documentation is in order.

The real blocker

The FDA’s Digital Health Center of Excellence actively encourages modelling and simulation. ICH Q13 explicitly supports continuous manufacturing enabled by adaptive process control. The EMA accepts digital twin data in submissions. The FDA Modernization Act 2.0 (2022) permits in silico models as alternatives to certain animal studies.

Regulatory risk is not the primary obstacle.

The real blockers are organisational: data silos, skills gaps, and the institutional inertia of a “frozen process” culture, where validation means lock-down rather than evidence generation.

The glass box requirement

The core regulatory requirement for adaptive AI in GMP is not “don’t adapt”. It is “be explainable”.

A black box — inputs in, outputs out, mechanism unknown — is not approvable. A glass box — every input documented, every adaptation logged, every control output attributable to a specific model state — is.

The compliance framework has three pillars:

• FDA Predetermined Change Control Plan (PCCP). A pre-agreed document submitted with the regulatory filing that defines what the model will learn (adaptation scope), how much it will change (bounded by the design space) and what performance triggers a new regulatory submission. The result: continuous learning without a new filing every time the model retrains.
• ICH Q13 (Continuous Manufacturing, 2022). It describes the adaptive digital twin not as an exception to regulatory expectation, but as the expected architecture for a continuous manufacturing control strategy.
• EMA AI Reflection Paper (2025). A risk-stratified approach: advisory-mode twins (medium risk) require standard validation; real-time release testing (high risk) requires rigorous model lifecycle management and ALCOA+ data integrity.

The compliance checklist

There are six documents an FDA investigator may request. (The checklist is shown as a table in the original post.)

The phased autonomy path — shadow mode → advisory → exception-based → autonomous — is both the regulatory expectation and the organisational trust-building approach. Submit Phase 1. Generate data. File a supplement for Phase 2. Don’t attempt Phase 3 as the initial regulatory position.

Model lifecycle management (MLM) is the successor to traditional CSV for systems that learn. You validate the adaptation algorithm, the guardrails and the monitoring system — not the specific parameter values at a point in time. The code doesn’t change. The validated boundaries don’t change. The model learns within them.

GAMP 5 Second Edition (2022) makes Agile development of adaptive systems tractable within GMP. The validation framework adapts to software that learns.

The regulatory path exists, is coherent, and has been walked before.

The question isn’t whether regulators will accept your adaptive twin. It’s whether your documentation will deserve acceptance.

What’s the biggest regulatory misconception you’ve encountered in digital twin conversations at your organisation?

Originally published on LinkedIn on 24 April 2026.