The challenges of traditional machine learning (ML) for real-time systems are primarily related to the issue of "dataset shift" in digital twins and the need for frequent retraining. In dynamic reactor/plant environments, ML models trained on specific operating conditions (specific datasets) may not generalize well to new, varied operational reactor conditions. This leads to a decrease in model performance and reliability when faced with scenarios (operating conditions) not covered during training. Although regular retraining of the ML model (with new operating conditions) can resolve this challenge, the requirement for constant retraining to adapt to new conditions is impractical and resource-intensive, posing a significant challenge in real-time applications. It can be especially challenging for real-time monitoring, which needs to be solved for complex nuclear plant operation and maintenance. ML-based prediction algorithms that need extensive retraining for new reactor operational conditions definitely prohibit real-time inference for digital twins across varying operational scenarios.
Real-time solutions to solve monitoring of nuclear digital twin
This challenge has been addressed in a recent study published in Nature Portfolio’s Scientific Reports. In the study, Ph.D. student Kazuma Kobayashi and Assistant Professor Dr. Syed Bahauddin Alam of the Nuclear, Plasma, and Radiological Engineering Department unveiled the potential of the deep neural operator (DeepONet) in revolutionizing digital twin-enabling technologies for nuclear energy systems.
In this study, DeepONet is trained with possible operational conditions that relax the requirement of continuous retraining for new conditions, making it suitable for online and real-time prediction components for DT. After thorough training, DeepONet can find solutions across the entire domain for new parameters up to ~1,000 times faster than conventional methods like nonlinear finite element analysis. This means what used to take a long time for a single simulation can now be done in mere fractions of a second. While a heavy ion transport code (PHITS) simulation took about 30 seconds, DeepONet performed the task in just 0.02 seconds. DeepONet stands out as a robust surrogate modeling method, capable of making accurate predictions in complex and dynamic nuclear reactor environments in real-time. This remarkable speed makes DeepONet an effective modeling method for digital twin systems, enabling real-time predictions of the intricate behaviors and spatial distributions of neutron flux in nuclear systems based on data from sensors installed on physical assets.
Their contribution marks an advancement in digital twin technology for nuclear systems, offering a scalable, efficient solution for real-time monitoring and predictive analysis. The study was built on the development of George E. Karniadakis, a professor of applied mathematics, and his research team at Brown University for inventing the DeepONet method (Source: Lu, L., Jin, P., Pang, G., Zhang, Z. & Karniadakis, G. E. Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators. Nat. Mach. Intell. 3, 218–229 (2021)).
Dr. Alam’s MARTIANS (Machine Learning and ARTIficial Intelligence for Advancing Nuclear Systems) lab is dedicated to developing digital twin-enabling technologies to implement real-time and online prediction algorithms for nuclear systems with a specific focus on explainable AI (XAI)-based prediction, diagnostic/prognostic, uncertainty quantification, natural language processing, etc.