A South Korean research group says it has completed roughly 80% of the work on a hybrid digital twin that couples mechanistic modeling of Chinese Hamster Ovary (CHO) cells with explainable artificial intelligence, with the remaining 20% focused on linking the models to process control and future robotics.

The project runs at the Bioprocess Digital Twin Lab at Sungkyunkwan University, led by professor and lab head Dong-Yup Lee, PhD. His team has built mathematical models of mammalian CHO cells and uses them to predict how the cells behave under different bioreactor conditions — the core workload question for any CMC or upstream process group watching the effort.

"We've been focused on combining this mechanistic model with the data-driven AI model to create a hybridized model," Lee explained in an interview with GEN – Genetic Engineering and Biotechnology News.

What separates this from a standard simulator?

According to Lee, the distinction lies in continuous connectivity. A conventional simulator or standalone model runs offline. His lab's twin stays linked to a multi-sensory modeling system that collects data directly from the bioreactor in real time, feeding the hybrid model as the process unfolds.

That architecture, not the underlying equations alone, is what the team is betting on for closed-loop operation.

Why add AI to a mechanistic model?

Lee is blunt about the limits of each approach on its own. Predictions from the mechanistic model alone, he says, do not always reach the accuracy or adaptability that real-time bioreactor operation demands. Data-driven AI fills that gap with more adaptive control.

But AI brings its own constraint, and Lee frames it as an interpretability problem:

"One of the limitations of AI is, if you have lots of data, it's good for prediction, but it can't explain why. So, we use what is called explainable (XAI)."

Explainable AI, in this setup, reports which input conditions most strongly affect process outputs rather than returning a bare prediction. That interpretability gives process engineers multiple options for improving and controlling bioprocess performance — a practical advantage for teams that must justify control decisions to regulators or internal quality functions.

For R&D managers, the implication is concrete: the hybrid approach targets the two failure modes that typically kill model-based control in GMP environments — mechanistic models that drift from reality, and black-box predictors that nobody can defend in a deviation investigation.

Where the project stands — and what remains

Lee puts the hybridized model itself at about 80% complete. The hardest part so far, he says, has been integrating the mechanistic model with XAI in a way that connects three functions end to end: data collection, prediction and forecasting, and direct process control.

"We're probably at 80% on [developing] the hybridized model," he said. "The remaining 20% is working out how these models can be linked to and interact with the control system and our future robotics."

That last 20% is non-trivial. Model-to-control integration and robotics interfaces are typically where academic digital twin projects stall before reaching pilot scale, so the self-reported figure should be read as progress on the modeling stack, not on a demonstrated autonomous lab.

The group's stated trajectory moves through three stages:

• Mechanistic CHO cell models predicting behavior under varying bioreactor conditions
• A hybrid layer adding data-driven AI plus XAI for adaptive, interpretable real-time control
• Autonomous process operation with minimal human intervention, at academic scale and larger

What it means for portfolio decisions

No timeline, funding figure, or industrial partner is disclosed in the source material, and the performance claims — the 80% completion estimate in particular — are the lab's own, not independently benchmarked. Sample sizes, validation datasets, and prediction accuracy figures are not reported.

For now, the work matters to two audiences: upstream process groups evaluating digital twins as alternatives to purely empirical scale-up, and automation teams scoping the modeling depth required before autonomous biomanufacturing becomes credible.

Lee says he is open to industry collaborations, specifically around digital twins, advanced bioprocess monitoring, and autonomous biomanufacturing.

The team's next research phase will concentrate on connecting the predictive twin to control systems and robotics, with the stated aim of moving beyond prediction toward increasingly autonomous process operation.