Abstract
Nuclear fusion research relies on a suite of diagnostics to monitor plasma behavior, yet these sensors often face a fundamental trade-off between temporal and spatial resolution. This paper presents Diag2Diag, a multimodal machine learning framework designed to overcome these limitations. By leveraging high-speed data from auxiliary sensors and exploiting intrinsic physical correlations, Diag2Diag generates "super-resolution" versions of slow diagnostics. Applied to the DIII-D tokamak, the framework successfully increased the temporal resolution of Thomson Scattering measurements from 0.2 kHz to 1 MHz. This breakthrough allows for the observation of fast-scale instabilities, such as Edge Localized Modes, and provides critical experimental validation for long-standing physics theories.
Introduction
In the quest for clean energy via nuclear fusion, monitoring the "burning" plasma inside a tokamak is a massive challenge. Scientists use various "diagnostics" (sensors), but they all have trade-offs. For instance, Thomson Scattering (TS) provides excellent spatial detail of electron density and temperature, but it is slow—sampling only every 5 milliseconds (200 Hz).
Unfortunately, many critical plasma events, such as Edge Localized Modes (ELMs)—violent instabilities that can erode the reactor walls—happen on sub-millisecond timescales. Because the TS laser cannot fire fast enough without overheating, these events often fall between the data points. Traditional super-resolution techniques usually interpolate between points, but in fusion, if a diagnostic misses an event entirely, there is no baseline to interpolate. Diag2Diag addresses this by using other high-speed sensors to "fill in" the missing physics.
Methods
The authors introduce Diag2Diag, a multimodal deep learning approach. Instead of looking at a single sensor's history to guess what happened next, it learns the intrinsic physical correlations between different types of sensors.
- Spectrogram Mapping: The researchers first used a Convolutional Neural Network (CNN) to map high-frequency data from one diagnostic (Electron Cyclotron Emission, or ECE) to another (Interferometer). This proved that AI could reconstruct complex features like "frequency chirping" by looking at a different physical modality.
- Time-Series Super-Resolution: They built a Multilayer Perceptron (MLP) that takes inputs from five high-speed diagnostics (sampling up to 2 MHz) to predict the state of the slow TS diagnostic.
- Physics-Preserving Constraints: By including first and second temporal derivatives of the input signals, the model captures the evolution of the plasma, ensuring the generated data obeys underlying laws of plasma physics, such as momentum balances.
Datasets and Validation
The study utilized a massive dataset from the DIII-D tokamak, consisting of over 4,000 discharges recorded between 2017 and 2022. Diagnostics included CER, Interferometer, ECE, MSE, and Thomson Scattering. For the time-series model, they used over 135,000 training samples and nearly 19,000 test samples. To validate the results, they used a special "bunch mode" operation of the TS laser—where it fires in rapid bursts for a very short time—as a ground-truth benchmark to verify the machine learning model's accuracy.
Key Results and Physics Discovery
The model successfully reconstructed ELM crashes that occurred entirely between the actual pulses of the TS laser. It didn't just smooth the data; it "saw" the crash because the other high-speed sensors felt the perturbation. The time-series MLP achieved a high $R^2$ score of 0.92, indicating a very strong match between synthetic data and real-world measurements.
Beyond mere reconstruction, the super-resolution data allowed researchers to see "profile flattening" at the plasma edge caused by Resonant Magnetic Perturbations (RMPs). This provides experimental evidence for the existence of "magnetic islands," a key theoretical mechanism for stabilizing plasmas that was previously too blurry to confirm.
Conclusion
Diag2Diag effectively increased the temporal resolution of Thomson Scattering from 0.2 kHz to 1 MHz. While the current model is robust, its performance on entirely new reactor designs (like ITER) would require further transfer learning. However, the authors suggest this "Diag2Diag" philosophy could be applied to other complex fields where sensors have mismatched resolutions, such as medical imaging (MRI vs. CT), astrophysics, or fluid dynamics.
Summary updated