PanoMHD: Multimodal Modelling of Plasma Dynamics towards Tokamak Control

Abstract

Modelling the dynamics of complex physical systems is a fundamental challenge, particularly where nonlinear dynamics and multi-scale interactions render traditional simulations computationally prohibitive. Nuclear fusion plasma represents a complex system where accurately predicting the plasma state, encompassing both performance and stability, is a prerequisite for active control required for sustained energy production. However, existing approaches are limited in providing a comprehensive solution as they largely focus on predicting isolated indicators such as binary stability labels. To overcome this, we present Panoramic MagnetoHydroDynamics (PanoMHD), a self-supervised multimodal framework designed to model plasma dynamics. By utilising a causal Transformer operating on tokenised representations of multimodal physical signals, PanoMHD is able to model the dynamics of high-dimensional magnetic fluctuation signals, which serve as a direct signature of plasma stability. This shifts the prediction paradigm from isolated indicators to multimodal signals. We pioneer the direct prediction of magnetic fluctuation signals for the first time, and demonstrate that this comprehensive representation enables state-of-the-art performance on KSTAR nuclear fusion plant experimental data. Our model outperforms baselines in future plasma performance prediction ($R^2=0.987$ vs. $0.957$) and surpasses dedicated classifiers in the downstream classification of distinct plasma states (L/H mode) with 97.3\% vs. 94.5\% accuracy.

Figures (9)

• Figure 1: VQ-VAE tokenisation process. The VQ-VAE tokenises the input MC cross-power $P$ and cross-phase $\delta$ spectrograms - represented in $\mathbb{R}^{F \times T \times 2}$ (where $F$, $T$ denote frequency and time dimensions) -- into a discrete latent representation $MC_{t} \in \{1, \dots, V_o\}^{d_{MC}}$. Here $V_o$ represents the vocabulary size of the codebook, and $d_{MC}$ denotes the length of the token sequence.
• Figure 2: Overview of PanoMHD. A causal Transformer predicts the next-step latent MC tokens and plasma performance metrics. The unified input sequence at each step comprises the plasma control $c_{t}$, latent MC $MC_t$, and plasma performance $p_t$. The respective token length ($d_c$, $d_{MC}$, and $d_p$) sum to a total sequence length of $d_{total}=219$. To predict the target states at $t+1$, the model processes a sequential context window of length $L=10$.
• Figure 3: Evolution of PanoMHD evaluation metrics with respect to compute budget (FLOPs). The plot tracks the test set performance for plasma parameters ($R^2$ of $\beta_N, H_{89}$), PSNR of MC cross-power, MC cross-phase, and L/H mode classification (OASIS accuracy).
• Figure 4: Zoomed-in analysis of ELM dynamics in KSTAR discharge #29655. The temporal window is restricted to $5.5\,\mathrm{s}$--$6.0\,\mathrm{s}$ to highlight the H-mode phase. (a) The $D_\alpha$ emission signal exhibits characteristic periodic spikes indicative of ELMs. (b), (c) Comparison of MC cross-power spectrograms between Ground Truth (GT) and PanoMHD prediction. The intermittent ELM events manifest as broadband vertical impulsive structures in the frequency domain, which are reproduced in the prediction. (d), (e) Comparison of MC cross-phase spectrograms between GT and PanoMHD prediction.
• Figure 5: Zoomed-in analysis of tearing instabilities in KSTAR discharge #25921. The temporal window is restricted to $3.5\,\mathrm{s}$--$5.5\,\mathrm{s}$ to highlight the onset of tearing instabilities. (a) Comparison between true (solid black) and PanoMHD-predicted (dashed lines) plasma performance parameters ($\beta_N$ and $H_{89}$). (b-e) Comparison of MC cross-power and cross-phase spectrograms between GT and PanoMHD prediction.