Introduction
In nuclear fusion research, understanding the internal state of plasma—its "profile"—is critical for maintaining stability and achieving control. However, the interior of a tokamak cannot be observed directly. Instead, researchers rely on diagnostics like soft X-rays (SXR), which provide "line-integral" measurements. These measurements are essentially the sum of emissivity along various lines of sight across the plasma volume.
Traditionally, reconstructing a 2D image from these 1D sums requires complex inversion algorithms, such as Fourier-Bessel or Gaussian process tomography. While accurate, these methods are computationally expensive and far too slow for real-time reactor control. Standard deep learning models offer the necessary speed but are often "physics-blind," potentially producing results that violate the underlying geometric and physical principles of the diagnostic system.
The ONION Architecture
The authors propose the Onion model, a novel physics-informed deep learning architecture designed to bridge the gap between speed and physical fidelity. The model is named for its layered approach to integrating information through several key innovations:
- Physical Information (PI) Side-Chain: The model incorporates a simplified VGG-style network to process the "response matrix," which contains the geometric path lengths of the diagnostic lines of sight.
- Element-wise Multiplication Fusion: Rather than simple concatenation, the model uses element-wise multiplication to fuse physical features with the backbone network (e.g., VGG or ResNet). This approach better captures the multiplicative relationship between device geometry and measured signals.
- Physics-Informed Loss Function (PILF): The loss function penalizes the model not just for pixel-wise differences from the target, but also for physical inconsistency. It performs a "back-projection" to calculate what the measurements should have been based on the predicted 2D profile, ensuring they match the actual inputs.
- Softplus Activation: By replacing standard ReLU activation with Softplus in the final layers, the model ensures smooth, continuous, and strictly positive emissivity profiles, avoiding unphysical "jagged" edges.
Datasets and Validation
The model was validated using four distinct datasets covering both synthetic and experimental conditions from the EAST and HL-2A tokamaks. Synthetic datasets provided a perfectly known "ground truth," while experimental datasets tested the model's robustness against real-world noise and the inherent errors of traditional inversion codes used for labeling.
Key Results and Insights
The Onion model demonstrates significant improvements over physics-blind deep learning approaches:
- Accuracy and Consistency: The inclusion of physical information reduced average relative error by approximately 52% on synthetic data. Crucially, the PILF reduced back-projection error by up to 85.4% in experimental cases, suggesting the model can sometimes surpass the quality of its own training labels by enforcing physical consistency.
- Real-Time Performance: Running on an NVIDIA GPU, the model performs reconstructions in just 1–9 milliseconds. This is over 700 times faster than traditional inversion codes, making it a viable candidate for real-time feedback loops in fusion reactors.
- Smoothness: The transition to Softplus activation was essential for reflecting the continuous nature of physical emissivity, preventing the "stepped" artifacts common in standard neural network outputs.
Future Directions
While highly effective, the current model does not yet account for measurement uncertainty or complex Magnetohydrodynamic (MHD) phenomena like magnetic islands. Future work will focus on integrating more diverse physical knowledge and exploring "Mixture of Experts" (MOE) architectures to handle a wider variety of experimental discharge scenarios.
Summary updated