"Deep learning model boosts plasma predictions in nuclear fusion by 1,000 times."
"The proposed FPL-net can solve the FPL equation in a single step, achieving results 1,000 times faster than previous methods with an error margin of just one-hundred-thousandth, demonstrating exceptional accuracy."
Is this a key advance towards the establishment of fusion-powered energy generation?
"In this work, we resolve the stability problem by presenting FPL-net, a surrogate FPL collision operator based on deep learning methods, for stable, long rollout simulation of anisotropic electron plasma relaxation. We applied the model structure of U-Net, which is a convolutional neural network, as the backbone of FPL-net [36]. Originally proposed for image segmentation in the biomedical field, U-Net consists of an encoder that captures the context of an input image and a decoder that performs upsampling for precise localization and high resolution. With these characteristics, U-Net demonstrates high performance with a small model size and has been applied in various fields [37]. We chose to utilize U-Net due to its encoder–decoder architecture that maintains the same input and output sizes, preserves high-resolution local information, and achieves strong performance even with a relatively small model size.
FPL-net focuses on fast and stable relaxation simulation. During training, we ensure stable predictions in the presence of perturbations in the input by using the model's output as the subsequent input and computing the output for two future time steps. We implement loss functions for density, momentum, and energy conservation, which are fundamental characteristics of general collision operators, to train the model and ensure that it learns physical laws. As a result, when simulating up to 200 time steps, the time-integrated relative error remained at the level of 10?5. While sustaining such a level of accuracy, our approach demonstrated over 1000 times faster speed compared to traditional numerical analysis algorithms in the tested environment."