ResearchPod Summary
Digital Twins rely on surrogate models to mirror physical systems, but these models often degrade over time due to concept drift—where the relationship between inputs and outputs changes. The authors address the challenge of maintaining model fidelity in real-time, specifically focusing on three problems: determining when to update a neural network surrogate, how to update it efficiently without catastrophic forgetting, and how to statistically validate that an update actually improves performance before deployment.
The proposed framework integrates three key components into a closed-loop Digital Twin system:
The framework was evaluated using a stochastic linear system and a directed energy deposition (DED) additive manufacturing process. The results demonstrate that the system successfully detects distributional shifts with minimal delay. By combining parameter-efficient fine-tuning with rigorous statistical validation, the framework restores predictive accuracy and uncertainty quantification under both abrupt and incremental drift, providing a reliable pathway for long-term Digital Twin operation.
As Digital Twins are increasingly used for safety-critical tasks like autonomous manufacturing, their reliability cannot be static. This research provides a principled, computationally tractable method to ensure that data-driven surrogates remain trustworthy throughout their operational life cycle, bridging the gap between machine learning-based prediction and robust control.
AI-generated third-party summary by ResearchPod. Not official content or an endorsement by the paper authors or affiliated organizations.