Smarter Digital Twins With Bayesian Calibration Framework
Digital twins have become indispensable in advanced manufacturing, especially in the semiconductor and display industries. These virtual replicas of automated material handling systems (AMHSs) offer visibility, predictability, and responsiveness across complex networks of transporters, machines, and storage hubs. Yet, despite their promise, they often fall short of mirroring real-world operations with precision. The culprits? Parameter uncertainty and operational discrepancy.
Parameter uncertainty refers to those subtle but crucial variables that resist precise measurement. Consider an automated vehicle’s acceleration in a factory setting. In the real world, it fluctuates slightly; in the digital twin, it remains fixed. Discrepancy, on the other hand, stems from the inherent differences between simplified digital models and the multifaceted logic of actual systems. Over time, these small deviations compound, leading to inaccurate predictions and potential production delays.
Traditionally, calibration frameworks have focused on fine-tuning parameters while largely ignoring discrepancy. This narrow approach often demands extensive field data, which can be costly and time-consuming to collect.
A Bayesian Leap Forward
Recognising these shortcomings, a research team led by Professor Soondo Hong from Pusan National University’s Department of Industrial Engineering has introduced a new Bayesian calibration framework. The innovation lies in its ability to simultaneously address both parameter uncertainty and discrepancy—delivering more accurate and reliable digital twins.
“Our framework enables us to simultaneously optimise calibration parameters and compensate for discrepancy,” says Prof. Hong. “It’s designed to scale across large smart factory environments, delivering reliable calibration performance with significantly less field data than conventional methods.”
The study, available online from 8 May 2025 and featured in Volume 80 of the Journal of Manufacturing Systems on 1 June 2025, demonstrates the framework’s capability to thrive under real-world constraints.
How the Framework Works
The team employed modular Bayesian calibration across varied operating scenarios. Bayesian methods integrate sparse real-world data with prior knowledge and simulation results, using probabilistic models—such as Gaussian processes—to produce posterior distributions of calibrated outcomes.
To gauge performance, the researchers compared three models:
- Field-only surrogate – predicts real-world behaviour solely from observed data.
- Baseline digital twin – uses only calibrated parameters without discrepancy compensation.
- Calibrated digital twin – accounts for both parameter uncertainty and discrepancy.
The calibrated digital twin consistently outperformed the other two, delivering marked improvements in prediction accuracy, even when working with minimal field observations.
“Our approach enables effective calibration even with scant real-world observations, while also accounting for inherent model discrepancy,” Prof. Hong notes. “It offers a practical and reusable calibration procedure validated through empirical experiments, and can be customised for each facility’s characteristics.”
Real-World Impact and Industry Applications
The potential of this framework extends far beyond semiconductor fabs. Its design makes it adaptable for large-scale, high-complexity production environments, where manual optimisation is challenging and discrepancies between model and reality are inevitable.
At Samsung Display, for instance, the research team has collaborated closely with operational staff to customise the framework for the company’s specific production complexities. By doing so, they’ve enabled more accurate predictions of system responses and streamlined the calibration of future production schedules.
Industries that stand to benefit include:
- Automotive manufacturing – where robotic assembly lines require precision timing and coordination.
- Aerospace – with intricate supply chains and high-precision assembly.
- Pharmaceutical production – where environmental variables can significantly impact output.
The framework’s versatility and low data requirement mean it could become a core enabler of cross-industry digital twin adoption.
Why This Matters for Smart Manufacturing
In an era where factories are increasingly interconnected and automated, the ability to adapt and calibrate digital twins efficiently is critical. Traditional calibration approaches that demand exhaustive field data are no longer practical. This Bayesian method offers a leaner, smarter alternative.
By addressing both uncertainty and discrepancy, manufacturers can:
- Improve production accuracy
- Reduce downtime
- Enhance predictive maintenance
- Support real-time decision-making
Prof. Hong believes this is just the beginning: “Our research offers a pathway toward self-adaptive digital twins, and in the future, has strong potential to become a core enabler of smart manufacturing.”
Pusan National University
Pusan National University, established in 1946, is South Korea’s top national university for research and educational excellence. With campuses in Busan, Yangsan, Miryang, and Ami, it serves around 30,000 students, supported by 1,200 professors and 750 staff. The university is committed to the values of truth, freedom, and service, and spans 14 colleges and 103 departments.
Prof. Soondo Hong’s work sits at the intersection of simulation, optimisation, and material handling operations. His team has developed decision-support tools that blend theoretical rigour with practical application, collaborating with global industry leaders to bring cutting-edge digital twin systems into real-world manufacturing environments. The group has also filed international patents to safeguard and expand the framework’s industrial reach.
Paving the Way for Next-Generation Digital Twins
The introduction of this Bayesian calibration framework marks a pivotal moment in the evolution of digital twin technology. By marrying mathematical precision with industrial pragmatism, it offers manufacturers a tool that’s not only technically advanced but also operationally feasible.
As industries continue to push toward fully integrated smart factories, solutions like this could become the linchpin for achieving the next level of efficiency, adaptability, and predictive accuracy.
