Image Diagnostics with CNNs
Applying convolutional networks to fast cameras, tomography, and other imaging diagnostics.
Imaging in fusion
Fast cameras, bolometer and soft X-ray tomography, and spectroscopic imaging produce image or image-like data of the plasma. These reveal edge filaments, radiation patterns, impurity influxes, and instabilities. Convolutional neural networks (CNNs), designed for spatial data, are a natural tool for analyzing them.
What CNNs do here
- Classify events visible in camera frames, such as edge-localized modes
- Detect and localize impurity influx or hot spots
- Reconstruct emissivity from tomographic projections
- Track features across frames for velocity and evolution
Why convolution fits
Convolutional layers share weights across the image, detecting the same pattern anywhere it appears, and build up from local features to larger structures. This makes them data-efficient for spatial patterns compared to treating each pixel as an independent input.
Tomographic reconstruction
Recovering the 2D emissivity from a few lines of sight is an ill-posed inverse problem. Learned reconstruction, trained on simulated or phantom data, can be faster and can incorporate priors, but must be validated against classical reconstruction and against synthetic ground truth, because a network can hallucinate plausible-looking structure that is not real.
Cautions
Camera data vary with viewing geometry, lighting, and wall conditions, so a model trained on one setup may not transfer. Labels for events are expert-provided and noisy. As with all imaging ML, a confident-looking output must be checked, especially in reconstruction, where the danger is invented detail rather than obvious error.