IGNITE Tokamak World Model Architecture

Peter Steiner; Azarakhsh Jalalvand; Nathaniel Chen; Kouroche Bouchiat; Ricardo Shousha; SangKyeun Kim; Egemen Kolemen

physics.plasm-phOct 1, 2026

IGNITE Tokamak World Model Architecture

Authors: Peter Steiner, Azarakhsh Jalalvand, Nathaniel Chen, Kouroche Bouchiat, Ricardo Shousha, SangKyeun Kim, Egemen Kolemen

Organizations: Department of Mechanical and Aerospace Engineering Princeton University Princeton, NJ · Princeton Plasma Physics Laboratory Princeton, NJ · Princeton University, Princeton Plasma Physics Laboratory Princeton, NJ

Abstract

We introduce IGNITE, a generative world foundation model for fusion plasma behavior simulation trained in a self-supervised manner from over a decade of unlabeled experimental data at the DIII-D National Fusion Facility. The core of IGNITE is a dynamics model that can simulate DIII-D discharges from a given set of actuator trajectories. These trajectories can be supplied or generated on-the-fly from a textual prompt or from desired experimental outcomes. The model architecture consists of several spatio-temporal tokenizers that embed the different input modalities, including time-series like spatio-temporal measurement data, image sequences, and high-resolution spectrograms, each of which collected at vastly different time scales. The backbone is composed of an auto-regressive dynamics model that has the capacity to predict entire DIII-D discharges given initial latent plasma states and actuator trajectories over a theoretical infinite horizon. IGNITE paves the way towards efficient AI-driven experimental planning and world modeling for nuclear fusion.

Figures & tables

Figure 1: Main outline of the IGNITE world model.

Figure 2: IGNITE learning curves for different regularization levels and dataset sizes. Figure 2a : A subset of

1000

shots leads to overfitting, despite regularization.

Figure 3: IGNITE predicts the DIII-D discharge $190\,735$ from the held-out validation set.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Diagnostic	Measured properties	Channels	Sampling freq.	Resampled freq.	Modality type
CER	Ion temperature ( $T_{\mathrm{i}}$ ), plasma rotation ( $v_{\mathrm{tor}}$ )	$2$\times$48$	$100\text{\,}\mathrm{Hz}$	–	Slow time series
MSE	Magnetic pitch angle	$69$	$100\text{\,}\mathrm{Hz}$	–	Slow time series
TS	Electron temperature ( $T_{\mathrm{e}}$ ) and density ( $n_{\mathrm{e}}$ )	$2$\times$44$ & $2$\times$10$	$100\text{\,}\mathrm{Hz}$	–	Slow time series
Filterscopes	Line emissions	$8$	$10\text{\,}\mathrm{kHz}$	–	Fast time series
BES	Fluctuations	$16$	$1\text{\,}\mathrm{MHz}$	$500\text{\,}\mathrm{kHz}$	Spectrogram
ECE	Electron temperature	$48$	$500\text{\,}\mathrm{kHz}$	–	Spectrogram

Appendix

Table 1: Overview of the diagnostics used for this study.

Actuator	Actuated property	Channels	Sampling frequency
Electron Cyclotron Heating (ECH)	Power, toroidal angle, poloidal angle, polarization	$8$	$10\text{\,}\mathrm{kHz}$
Gas	Voltage, gas flow	$8$	$10\text{\,}\mathrm{kHz}$
Neutral Beam Injection (NBI)	Power, torque, voltage	$8$	$10\text{\,}\mathrm{kHz}$
Resonant Magnetic Perturbation (RMP)	Coil current	$8$	$10\text{\,}\mathrm{kHz}$
I-Coil	Coil current	$12$	$10\text{\,}\mathrm{kHz}$

Appendix

Table 2: Overview of the actuators used for this study.

Figure 4: Tokenizer for slow time series. The channel-time array is grouped into four channel groups and patched at 5 samples $\times$ channels; a shared linear layer projects each patch by Linear $\rightarrow 5\times$128$$ ; a time encoding and a sequence-position encoding are added to each token; the token sequence, each 128-dimensional, is processed by the transformer to produce the model output. — Figure 4: Tokenizer for slow time series. The channel-time array is grouped into four channel groups and patched at 5 samples $\times$ channels; a shared linear layer projects each patch by Linear $\rightarrow 5\times$ 128$$ ; a time encoding and a sequence-position encoding are added to each token; the token sequence, each 128-dimensional, is processed by the transformer to produce the model output.

Figure 5: Tokenizer for fast time series. A $50\text{\,}\mathrm{ms}$ window of 8 channels sampled at $10\text{\,}\mathrm{kHz}$ gives $500$ samples per channel; a Conv1D front end reduces each channel to 5 frames, patched at $5\times 8$ over frames and channels. A shared Linear $\rightarrow 8\times$256$$ projects each patch, a $256$ -wide time encoding is added, and the token sequence is processed by the transformer to produce the model output. — Figure 5: Tokenizer for fast time series. A $50\text{\,}\mathrm{ms}$ window of 8 channels sampled at $10\text{\,}\mathrm{kHz}$ gives $500$ samples per channel; a Conv1D front end reduces each channel to 5 frames, patched at $5\times 8$ over frames and channels. A shared Linear $\rightarrow 8\times$ 256$$ projects each patch, a $256$ -wide time encoding is added, and the token sequence is processed by the transformer to produce the model output.

Figure 6: Tokenizer for spectrograms. SEach of the $C$ channels contributes a time-frequency map of 98 frames, cut into $16\times 16$ patches, giving $192$ patches per channel. Patches at the same grid position are concatenated across channels and an MLP reduces them to $$192$\times$256$$ . A frequency encoding ( $$32$\times$256$$ ) and a time encoding ( $6\times$256$$ ) form a $32\times 6$ table whose entries are 256-dimensional vectors, added to the token at that grid position, before the transformer produces the model output. — Figure 6: Tokenizer for spectrograms. SEach of the $C$ channels contributes a time-frequency map of 98 frames, cut into $16\times 16$ patches, giving $192$ patches per channel. Patches at the same grid position are concatenated across channels and an MLP reduces them to $192$\times$256$ . A frequency encoding ( $32$\times$256$ ) and a time encoding ( $6\times$ 256$$ ) form a $32\times 6$ table whose entries are 256-dimensional vectors, added to the token at that grid position, before the transformer produces the model output.

World Model Learning

World model learning builds predictive models of how environments change under actions to support planning, simulation, and decision-making.

See all 399 papers

Decomposition-Guided Diffusion Language Models for Inertial Confinement Fusion Prediction
Sep 7, 2026Xiang Zhang, Varchas Gopalaswamy, Rahman Ejaz +2Time Series Forecasting Physics-Informed ML
Purdue University · University of Rochester
TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics
Feb 16, 2026Tobia Boschi, Andrea Loreti, Nicola C. Amorisco +13Multimodal Foundation Models Multimodal Pretraining
IBM Research Dublin, Ireland · UK Atomic Energy Authority Abingdon, UK · STFC Hartree Centre Daresbury, UK
World Machine: Towards Generative World Modeling for Time-Series
May 21, 2026Elton Cardoso do Nascimento, Alexandre da Silva Simões, Esther Luna Colombini +2World Model Learning Latent World Models
Universidade Estadual de Campinas (UNICAMP), {FEEC, IC} · Universidade Estadual Paulista (UNESP), Instituto de Ciˆencia e Tecnologia, Sorocaba

Sep 7, 2026J/K move · Enter open · S save

Decomposition-Guided Diffusion Language Models for Inertial Confinement Fusion Prediction

Xiang Zhang, Varchas Gopalaswamy, Rahman Ejaz, Riccardo Betti+1

Purdue University · University of Rochester

Time Series Forecasting Physics-Informed ML Diffusion Language Models

Open paper AskPDF arXiv

Inertial confinement fusion (ICF) is a leading pathway toward clean energy, but each shot at the National Ignition Facility costs on the order of one million dollars, making accurate AI surrogates a high-value target. We study exogenous-driven ICF waveform prediction, where a 512-step neutron-rate diagnostic must be inferred directly from a laser pulse and target design parameters, with no historical response observed. The regime stresses standard time-series predictors with temporal sparsity (picosecond peak in a nanosecond window), input-output scale mismatch (under 300 real shots), and peak sensitivity (picosecond timing). We propose ICF-DLM, to our knowledge the first LM-based ICF predictor, combining (i) a physics-typed decomposition into yield

Y_{DT}

, peak timing

t_{\mathrm{peak}}

, and local waveform

w_{\mathrm{local}}

; (ii) bidirectional denoising that defers commitment to peak location; and (iii) a physics-driven PPO reward re-injecting metric structure across numeric tokens. On ICFBench (50K simulations + 232 experimental shots), ICF-DLM cuts peak-timing error from 11.6 to 9.2 steps over a matched autoregressive LLaMA-3-8B and outperforms classical sequence models and LLM-based time-series predictors. Beyond ICF, the recipe shows potential to address science domains with low data and sparse events.

Abstract

World Model Learning

Explore similar work

Decomposition-Guided Diffusion Language Models for Inertial Confinement Fusion Prediction

TokaMind: A Multi-Modal Transformer Foundation Model for Tokamak Plasma Dynamics

World Machine: Towards Generative World Modeling for Time-Series