Enhancing Detail Recovery in ICF Radiographs: A Transformer-based Approach with ViXReg
Keywords: Vision transformers, Fine-tuning, Regression, Nuclear fusion, ICF, Foundation model, Scaling
TL;DR: Developing ViXReg, a transformer-based framework for advanced image analysis in nuclear fusion research.
Abstract: We introduce ViXReg, a framework that adapts Vision Transformers (such as Google ViT, Swin, BEiT) to tackle image analysis challenges in Inertial Confinement Fusion (ICF) radiography. ViXReg introduces a novel approach by adapting Vision Transformers, originally designed for pixel-level classification, to handle complex image regression tasks. This transformation, which is not commonly explored in current methods, enables precise reconstruction of asymmetric double-shell structures, essential for diagnosing nuclear fusion dynamics and identifying instabilities in high-energy-density plasmas. Our investigation explores architectural adaptations, including nonlinear and linear mappings, and advanced fine-tuning strategies like multi-scale pre-training and knowledge distillation, enhancing model scalability and generalization across diverse data distributions. Evaluating 60,000 synthetic ICF radiographs and 115 radiographs captured from 6 ICF experimental shots, we further craft domain learning techniques with weakly pseudo-labeled data, enabling ViXReg to transfer robust representations effectively to experimental dataset. The results from the above tasks demonstrate considerable advancements in using transformers as backbone architectures for fusion imagery, effectively capturing the subtle double-shell structures identified in plasma physics. Additionally, fine-tuning the pre-trained ViXReg model accelerates training convergence and enhances the accuracy of double-shell reconstructions, surpassing the performance of traditional convolutional neural networks and generative adversarial models. These findings demonstrate ViXReg's potential as a candidate for a foundation model component in scientific modeling for nuclear fusion research.
Submission Number: 74
Loading