Researchers from the Pusan National University, South Korea, have published a new calibration framework for digital twins. The framework promises improved prediction accuracy for digital twins of automated material handling systems (AMHSs) in semiconductor and display fabrication industries.
AMHSs typically involve complex manufacturing steps and control logic, and digital twin models have emerged as a promising solution to enhance the visibility, predictability and responsiveness of production and material handling operation systems.
However, digital twins don’t always fully reflect reality, potentially affecting production performance and may result in delays.
Digital twins of AMHSs face two major issues
Digital twins of AMHSs face two major issues: parameter uncertainty and discrepancy, according to the researchers.
Parameter uncertainty arises from real-world parameters that are difficult to measure precisely but are essential for accurate modeling. For example, the acceleration of an automated vehicle in AMHSs can vary slightly in the field but is fixed in the digital twin.
Discrepancy originates from the difference in operational logic between the real-world system and the digital twin. This is especially important since digital twins typically simplify or resemble the real processes and discrepancies accumulated over time lead to inaccurate predictions.
Most performance-level calibration frameworks overlook discrepancy and focus only on parameter uncertainty. Moreover, they often require a large amount of field data.
To address this gap, the research team, led by Professor Soondo Hong from the department of industrial engineering at Pusan National University, developed their Bayesian calibration framework.
“Our framework enables us to simultaneously optimize calibration parameters and compensate for discrepancy,” explained Professor Hong. “It is designed to scale across large smart factory environments, delivering reliable calibration performance with significantly less field data than conventional methods.”
Register for All Access: AI in Business Transformation 2025!
Enabling scalable, self-adaptive digital twins
The researchers applied modular Bayesian calibration, which uses sparse real-world data to estimate uncertain parameters while also accounting for discrepancy, for various operating scenarios. It works by combining field observations and available prior knowledge with digital twin simulation results through probabilistic models, specifically Gaussian processes, to obtain a posterior distribution of calibrated digital twin outcomes over various operating scenarios.
The researchers compared the performance of three models:
- A field-only surrogate that predicts real-world behavior directly from observed data.
- A baseline digital twin model using only calibrated parameters.
- The calibrated digital twin model accounting for both parameter uncertainty and discrepancy.
The calibrated digital twin model significantly outperformed the field-only surrogate and showed concrete improvements in prediction accuracy over the baseline digital models, the researchers found.
“Our approach enables effective calibration even with scant real-world observations, while also accounting for inherent model discrepancy,” said Professor Hong. “Importantly, it offers a practical and reusable calibration procedure validated through empirical experiments and can be customized for each facility’s characteristics.”
The developed framework is a practical and reusable approach that can be used to accurately calibrate and optimize digital twins, otherwise hindered by scale, discrepancy, complexity or the need to be flexible for widespread cross-industry application. The approach accurately predicted field system responses for large-scale systems with scarce field observations and supported rapid calibration of future production schedules in real-world systems.
“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,” Professor Hong stated.
[inlinead-1]