A new calibration approach developed by researchers in South Korea could solve one of the most persistent problems in smart manufacturing: the gap between digital twin models and real-world production systems.

By addressing both parameter uncertainty and discrepancy, the method offers a way to improve predictive accuracy — even with limited field data — and is already being piloted in large-scale operations.

Digital twin models have become essential tools for managing increasingly complex production environments, from semiconductor fabs to CDMO facilities. These systems simulate the flow of materials across automated equipment, transporters, and storage hubs — enhancing operational visibility, productivity, and scheduling. But real-world mismatch remains a key barrier to more widespread adoption, and even small modelling inaccuracies can result in costly delays.

Researchers from Pusan National University have now introduced a Bayesian calibration framework that significantly improves prediction accuracy in digital twins of automated material handling systems (AMHSs). Led by Professor Soondo Hong of the Department of Industrial Engineering, the team’s method accounts for both real-world parameter variability and model simplifications — a common limitation of current digital twin systems.

“Our x-framework enables us to simultaneously optimize calibration parameters and compensate for discrepancy,” said Prof. Hong. “It is designed to scale across large smart factory environments, delivering reliable calibration performance with significantly less field data than conventional methods.”

The work was published in the Journal of Manufacturing Systems (Vol. 80) in June 2025.

Understanding the digital twin disconnect

Digital twins of AMHSs typically struggle with two key issues:

• Parameter uncertainty, where physical attributes such as the acceleration rate of automated vehicles are fixed in the model but vary in reality.

• Discrepancy, the broader mismatch between simplified digital logic and actual operational behaviour — which compounds over time and leads to inaccurate forecasts.

While some performance-level calibration methods address parameter uncertainty, most overlook discrepancy entirely and often require large datasets that aren’t feasible in fast-moving production environments.

The new modular Bayesian calibration framework addresses this shortfall. It uses sparse real-world data and prior knowledge to estimate uncertain parameters and correct for discrepancy — employing probabilistic modelling and Gaussian processes to generate a posterior distribution of predicted system behaviour under varying conditions.

The researchers evaluated three models:

• A field-only surrogate, which uses observational data without digital simulation

• A baseline calibrated twin that adjusts for parameter uncertainty only

• The new calibrated digital twin, which integrates both uncertainty and discrepancy

Their approach outperformed both alternatives, improving accuracy in high-complexity systems with minimal real-world data.

“Our approach enables effective calibration even with scant real-world observations, while also accounting for inherent model discrepancy,” said Professor Hong. “It provides a practical and reusable procedure validated through empirical experiments and can be customized for each facility’s characteristics.”

From research to real-world impact

The framework is already being scaled for implementation at Samsung Display, where the research team is collaborating with operations specialists to apply the system in live production environments. The goal is to reduce calibration time, support agile rescheduling, and enable self-correcting digital twins that learn and adapt as new data comes in.

The method is particularly useful in facilities where:

• Digital twins oversimplify complex operational logic

• Manual optimisation is no longer scalable

• Real-world data collection is expensive or disruptive

• Model accuracy is critical for throughput, safety, or cost-efficiency

The x-framework also supports the creation of modular, reusable calibration workflows, enabling broader application across industries where digital twins are gaining traction — including pharmaceutical manufacturing, logistics, diagnostics, and bioprocessing.

“Our research offers a pathway toward self-adaptive digital twins,” concluded Hong. “In the future, this has strong potential to become a core enabler of smart manufacturing.”