Abstract: Model-based plasma scenario development lies at the heart of the design and operation of future fusion powerplants. Including turbulent transport in integrated models is essential for delivering a successful roadmap towards operation of ITER and the design of DEMO-class devices. Given the highly iterative nature of integrated models, fast machine-learning-based surrogates of turbulent transport are fundamental to fulfil the pressing need for faster simulations opening up pulse design, optimization, and flight simulator applications. A significant bottleneck is the generation of suitably large training datasets covering a large volume in parameter space, which can be prohibitively expensive to obtain for higher fidelity codes. In this work, we propose ADEPT (Active Deep Ensembles for Plasma Turbulence), a physics-informed, two-stage Active Learning strategy to ease this challenge. Active Learning queries a given model by means of an acquisition function that identifies regions where additional data would improve the surrogate model. We provide a benchmark study using available data from the literature for the QuaLiKiz quasilinear transport model. We demonstrate quantitatively that the physics-informed nature of the proposed workflow reduces the need to perform simulations in stable regions of the parameter space, resulting in significantly improved data efficiency. We show an up to a factor of 20 reduction in training dataset size needed to achieve the same performance as random sampling. We then validate the surrogates on multichannel integrated modelling of ITG-dominated JET scenarios and demonstrate that they recover the performance of QuaLiKiz to better than 10%. This matches the performance obtained in previous work, but with two orders of magnitude fewer training data points.
Introduction
To design and operate future fusion power plants like ITER, scientists must accurately predict "turbulent transport"—the chaotic movement of heat and particles out of the plasma core. While high-fidelity gyrokinetic codes can simulate this turbulence, they are computationally too slow for real-time control or large-scale design optimization.
To bridge this gap, researchers build "surrogate models" (fast neural network approximations). However, these surrogates typically require millions of simulation data points to train, a requirement that is prohibitively expensive for the most accurate codes. This paper introduces ADEPT (Active Deep Ensembles for Plasma Turbulence), a framework that uses "smart sampling" to build accurate surrogates using up to 100 times less data than previous methods.
The ADEPT Framework
The core of ADEPT is a two-stage, physics-informed Active Learning (AL) strategy. Rather than training on a random dataset, the model iteratively "queries" the simulation code to identify regions in the 15-dimensional parameter space where it is most uncertain.
As illustrated in the workflow above, the system utilizes several key components:
- Two-Stage Architecture: ADEPT employs a Classifier to determine if the plasma is stable or unstable and a Regressor to predict fluxes only when the plasma is unstable. This architecture mimics the physical "critical threshold" behavior of plasma turbulence.
- Deep Ensembles for Uncertainty: To guide its learning, ADEPT uses an ensemble of neural networks to measure epistemic uncertainty—the uncertainty caused by a lack of training data. This uncertainty acts as a compass, pointing the model toward the most informative data points.
- Active Learning Loop: The model identifies high-uncertainty regions, runs the expensive physics model for those specific points, and appends the results to the training set to refine itself.
Data and Methodology
The study utilized the JET-Exp-15D dataset, which contains input-output mappings from the QuaLiKiz quasilinear transport model. The inputs consist of 15 dimensions representing plasma parameters (such as temperature gradients and magnetic shear) derived from over 2,000 experimental discharges at the JET tokamak. The model was tasked with predicting multiple transport channels simultaneously, including ion and electron heat fluxes.
Key Results and Impact
- Massive Data Reduction: ADEPT achieved a 20x reduction in data compared to random sampling and a 100x reduction compared to previous state-of-the-art surrogate models (requiring 200,000 points vs. 20,000,000).
- Physics-Informed Efficiency: In typical random datasets, roughly 75% of points are "stable" (zero flux). ADEPT’s classifier learns this boundary quickly, allowing the regressor to focus 80% of its effort on the complex, non-zero flux regions.
- High Accuracy and Validation: Despite the smaller dataset, the surrogates recovered the performance of the original QuaLiKiz model to better than 10% in integrated JINTRAC simulations. The model was successfully validated on distinct JET discharge scenarios (Carbon and Beryllium walls), proving its generalizability.
Conclusion and Future Work
ADEPT demonstrates that physics-informed active learning can drastically lower the barrier to creating high-fidelity surrogates for fusion research. While this proof-of-concept focused on Ion Temperature Gradient (ITG) turbulence, future work will expand the framework to include other types, such as Trapped Electron Modes (TEM) and Electron Temperature Gradient (ETG) turbulence. Further refinements to the acquisition functions may reduce the computational requirements even further, bringing us closer to real-time high-fidelity plasma control.
Summary updated