Towards a Robust Adaptive Digital Twin for Fusion Applications

The development of digital twin systems for fusion applications is key to improving prediction, analysis, and optimization of complex plasma processes. Machine learning, especially deep learning, has shown strong capabilities in modeling such nonlinear systems. However, two challenges limit deployment of deep learning-based digital twins: uncertainty quantification (UQ) and data drift. UQ is essential for trustworthy predictions, particularly in decision-support scenarios. Data-driven models are also sensitive to shifts in data distribution, such as shot-to-shot variations in fusion experiments, causing performance degradation. To address this, we are developing an uncertainty-aware, adaptive digital twin framework. Our approach uses deep learning models with Gaussian Process approximations for uncertainty estimation, combined with online learning for continuous adaptation to new data. This allows the twin to respond to evolving plasma behaviors and equipment conditions. To mitigate shot-to-shot drift, our system updates incrementally as new data arrives, improving robustness.

Our long-term vision is a self-sustaining digital twin with adaptive feedback loops to refine models and support real-time decision-making. We will share progress on the framework for modeling coil deflection at DIII-D, discuss challenges and opportunities in building an interactive digital twin, and highlight future directions—such as continual learning and human-in-the-loop feedback.