PaMMA-Net: Plasmas magnetic measurement evolution based on data-driven incremental accumulative prediction

TL;DR

PaMMA-Net tackles the challenge of long-sequence tokamak magnetic measurement evolution by introducing an incremental accumulative prediction framework that forecasts future measurements from observed signals and known future inputs. The method combines state fusion embeddings, variable-separated projections, and a causal-attention, decoder-based architecture to model multi-modal signals, with an increment-first loss formulation that improves convergence. Physically consistent data augmentation and ablations demonstrate substantial performance gains over strong baselines on the EAST dataset, achieving high similarity and low mean absolute error while maintaining efficiency. The approach supports integration with equilibrium reconstruction for plasma shape evolution and offers a robust, offline-friendly tool for controller-assisted tokamak operations.

Abstract

An accurate evolution model is crucial for effective control and in-depth study of fusion plasmas. Evolution methods based on physical models often encounter challenges such as insufficient robustness or excessive computational costs. Given the proven strong fitting capabilities of deep learning methods across various fields, including plasma research, this paper introduces a deep learning-based magnetic measurement evolution method named PaMMA-Net (Plasma Magnetic Measurements Incremental Accumulative Prediction Network). This network is capable of evolving magnetic measurements in tokamak discharge experiments over extended periods or, in conjunction with equilibrium reconstruction algorithms, evolving macroscopic parameters such as plasma shape. Leveraging a incremental prediction approach and data augmentation techniques tailored for magnetic measurements, PaMMA-Net achieves superior evolution results compared to existing studies. The tests conducted on real experimental data from EAST validate the high generalization capability of the proposed method.

Figures (8)

• Figure 1: Workflow and model architecture of PaMMA-Net. The left half presents the process of collecting magnetic measurements and control commands from tokamak discharge experiments to form a dataset, as well as the subsequent steps of preprocessing and data augmentation applied to this dataset for training purposes.. The right half displays detailed architecture of PaMMA-Net, including state fusion prediction, variable separation projection, and incremental accumulative prediction.
• Figure 2: (a) Distribution of $BP$ channel 1 within the dataset, (b) Distribution of the increment of $BP$ channel 1 within the dataset.
• Figure 3: The principle and efficacy of the physically consistent data augmentation.
• Figure 4: Learning behaviors of PaMMA-Net with and without incremental design.
• Figure 5: The evolution similarity metrics and its distribution of all discharges within the test set are presented. The left half depicts each discharge as an individual data point, while the right half illustrates the distribution using a horizontal histogram. Two discharges exhibiting moderate levels of evolution similarity have been selected and will be analyzed subsequently.