PaperlayerView exact revision on arXiv

Source-linked AI summary

IGNITE Tokamak World Model Architecture

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

arXiv:2610.02515v1physics.plasm-phcs.AIcs.LG

TL;DR

Fusion plasma behavior is complex and difficult to predict, and prior multimodal fusion models have focused on short windows rather than autoregressive dynamics. IGNITE uses heterogeneous diagnostics and actuator trajectories to model DIII-D discharges, and the authors report that generalization requires substantially more data than a 1,000-shot training set. They identify physics-based evaluation and the impact of bad or unavailable inputs as future work.

  • Problem

    Fusion plasma behavior is nonlinear and multiscale, while prior multimodal fusion models predict or reconstruct short windows rather than autoregressively modeling discharge dynamics.

    Source: §1 · p. 2

  • Method

    IGNITE tokenizes heterogeneous DIII-D diagnostics into discrete codes and autoregressively predicts discharge windows conditioned on actuator trajectories.

    Source: §2 · p. 3

  • Results

    With the complete training dataset, validation performance could exceed that of the 1,000-shot model even at an early training stage, indicating that generalization requires substantially more data.

    Source: §4.1 · p. 7

  • Takeaways & Limitations

    The authors conclude that sufficiently large training data are crucial for generalization in autoregressive prediction across diverse signal modalities.

    Source: §5 · p. 8

    Physics-based evaluation remains future work, and the impact of bad or unavailable inputs during training and inference still needs study.

    Source: §5 · p. 8

Abstract

from arXiv · show

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.

1 Introduction

Fusion plasma behavior is complex and difficult to predict, while experiments are costly and decades of data remain underexplored. IGNITE addresses this setting with a multimodal, actuator-conditioned model; unlike prior fusion models focused on short windows, it autoregressively simulates discharges.

Motivation

  • Fusion plasma behavior is nonlinear, multiscale, and difficult to predict, despite decades of underexplored experimental data across diverse devices and diagnostics.Experiments are expensive and further work is needed to find scalable, high-performance plasma regimes.Source: §1 · p. 2

Contribution

  • IGNITE is a multimodal generative model trained on over 8,000 discharges (about 12 hours) from more than a decade of experiments, with frame-by-frame actuator control.It handles signals from Hz to MHz and time series, spectrograms, and a tokenizer design extensible to image sequences.Source: §1 · p. 2

Related work

  • Prior reinforcement-learning studies demonstrated AI models deployed on real tokamaks to control devices or design experiments [20].The cited studies showed the potential of reinforcement learning for fusion-device control.Source: §1 · p. 2
  • TokaMind [23], PanoMHD [24], and FusionMAE use heterogeneous diagnostics, but their prediction or reconstruction tasks operate over short windows rather than autoregressive dynamics.FusionMAE imputes missing channels without a temporal-dynamics model; TokaMind predicts up to 100 ms and PanoMHD predicts the next 50 ms window.Source: §1 · pp. 2–3

Approach

  • IGNITE uses spatiotemporal transformers and FSQ encoders to tokenize diagnostics and actuators, then autoregressively predicts tokens for the next frame.Its design draws on spectrogram, video, and time-series models, as well as world models such as Genie [5].Source: §1 · p. 3

2 The IGNITE world model

IGNITE tokenizes heterogeneous DIII-D diagnostics and actuator signals into discrete codes, then uses an autoregressive transformer to predict successive discharge windows. Its modality-specific tokenizers accommodate signals with widely different sampling rates and forms.

2 The IGNITE world model

  • IGNITE takes 20 windows of 50 ms of heterogeneous diagnostics and seven actuator groups, then autoregressively forecasts diagnostics in subsequent windows.Each modality is independently tokenized with discrete FSQ codecs, so the dynamics model learns in a fixed latent space.Source: §2 · p. 3

2.1 Per-modality tokenization

  • Modality-specific tokenizers address diagnostics ranging from 100 Hz to 1 MHz, including slow and fast time series, spectrograms, and image sequences.Each 50 ms diagnostic frame is compressed into discrete FSQ tokens to reduce dimensionality while supporting high-frequency spectrogram generation.Source: §2.1 · p. 4
  • Slow time series are resampled to 100 Hz and encoded from five samples per 50 ms window into four 128-dimensional tokens.The encoder patchifies measurements, adds spatial and temporal positional embeddings, and uses a transformer to learn spatiotemporal dependencies.Source: §2.1 · p. 4
  • Fast Dα signals are sampled at 10 kHz and encoded in four 256-dimensional tokens per 50 ms window; spectrograms use 192 tokens per modality.For spectrograms, 16 × 16 patches balance token count against the frequency resolution needed for mode reconstruction.Source: §2.1 · pp. 4–5

2.2 Shared Transformer Backbone

  • The dynamics model mixes modalities with within-frame cross-attention, then applies causal temporal attention to predict the next 50 ms window for iterative full-discharge simulation.Predicted codes are decoded back into the original modalities for comparison with measurements; the backbone has 1024 dimensions and 48 transformer layers.Source: §2.2 · p. 5

2.3 Training methods

  • Training first pretrains modality codecs, then trains the dynamics model with maskGIT masked-token prediction over 100 consecutive frames and per-modality cross-entropy losses.At inference, the first 20 real frames seed autoregressive generation, with each predicted frame appended to the input sequence.Source: §2.3 · pp. 5–6

3 Dataset

The dataset comprises selected DIII-D discharges collected from 2018 through 2026, with preprocessing and splitting performed at the discharge level. The remaining shot data are normalized and divided into 50 ms training chunks.

Dataset

  • The dataset contains 8,600 DIII-D discharges collected from 2018 to 2026, split randomly into 90% training, 5% validation, and 5% test sets.Disrupted shots were excluded when the planned-versus-actual plasma-current correlation fell below 0.8.Source: §3 · p. 6

Preprocessing

  • Each diagnostic was normalized by subtracting its global channel mean and dividing by its standard deviation; missing samples were zero-masked and excluded from loss.The binary mask was used to exclude missing data from the training loss.Source: §3 · p. 6
  • After skipping each shot’s first second because of transient effects, the remaining data were divided into non-overlapping 50 ms training chunks.The initial phase was excluded from the chunks used for model training.Source: §3 · p. 6

4 Results

On a held-out DIII-D discharge, IGNITE predicted modality trajectories from a one-second seed and actuator trajectory; learning curves show that generalization required substantially more data than memorizing a few shots.

4 Results

  • On 16 shots, IGNITE memorized training data after up to 80 epochs at 1 × 10−3, versus Genie’s 1 × 10−4.The 16 shots came from a single experiment.Source: §4 · p. 7

4.1 Learning curves

  • On 1,000 shots, dropout slowed training–validation divergence, but validation loss still rose after 12–15 epochs.Training loss was unaffected by regularization.Source: §4.1 · p. 7, Figure 2 · p. 7
  • With the complete training dataset, validation performance could exceed the 1,000-shot model early, showing that generalization required substantially more data.The minimum validation loss was hardly affected by regularization; the team retained dropout 0.3 and weight decay 0.1.Source: §4.1 · p. 7, Figure 2 · p. 7

4.2 Example validation output

  • From a one-second seed and given actuator trajectory, IGNITE predicted the full trajectory of multiple modalities on held-out discharge 190 735.The spectrograms retained fine-grained mode-structure errors, especially at higher ECE frequencies.Source: §4.2 · p. 7, Figure 3 · p. 8

5 Conclusions and future work

IGNITE is a world model trained on historical DIII-D data that autoregressively predicts signals across modalities. The authors identify data-quality and modality challenges and call for physics-based evaluation and study of bad or unavailable inputs.

5 Conclusions and future work

  • IGNITE is a world model trained on historical DIII-D data that autoregressively predicts signals across a broad range of modalities.The authors report that a sufficiently large training dataset is crucial for generalization.Source: §5 · p. 8
  • IGNITE’s 14 modalities span four groups and different sampling rates; noisy, malfunctioning, or historically changing diagnostics complicate the data.Channels were added over time, making historical data difficult to use directly.Source: §5 · p. 8
  • Future work includes physics-based evaluation and studying the impact of bad or unavailable inputs during training and inference.The authors identify these studies as necessary for developing IGNITE as a tool for experiment design and actuator-defect simulation.Source: §5 · p. 8

NeurIPS Paper Checklist

The checklist documents IGNITE’s scope, reproducibility conditions, and evaluation, including unquantified uncertainty and possible unexpected diagnostic outputs. It also records the authors’ responses and disclosures regarding LLMs and AI-assisted implementation and figure editing.

Main claims

  • The authors say IGNITE predicts discharge evolution from an initial state and actuator trajectory, evaluated with objective nRMSE measures and learning curves.They also discuss generalization and overfitting limitations.Source: NeurIPS Paper Checklist · p. 12

Limitations

  • The authors limit the model’s scope to DIII-D and do not claim transfer to other machines or physically validated diagnostic evolutions under changed actuator trajectories.They identify the latter as future work.Source: NeurIPS Paper Checklist · p. 13

Experimental result reproducibility

  • The authors state that hyperparameters, implementation details, optimizers, processing scripts, and experiment-selection details are provided; code is public, while DIII-D data require access procedures.They say data are available upon request under DIII-D guidelines and that their tools support preprocessing and result reproduction.Source: NeurIPS Paper Checklist · pp. 14–15

Experiment statistical significance

  • The authors report objective measurements but no uncertainty quantification, citing limited compute that prevented parallel training of several-billion-parameter models.They also say the hardware and resource use, training time, and inference time are described.Source: NeurIPS Paper Checklist · pp. 15–16

Societal impacts and safeguards

  • The authors say IGNITE uses no private data, could generate unexpected diagnostic trajectories, and relies on internal DIII-D publication review to identify unreasonable outputs.They say the training dataset cannot be released without DIII-D user agreements and that model checkpoints will be available upon reasonable request.Source: NeurIPS Paper Checklist · pp. 16–17

Assets and human subjects

  • The authors say they own the code, cite model-development foundations, document IGNITE’s training, validation, and limitations, and place its license in the FAITH repository; the research involved no human subjects.Source: NeurIPS Paper Checklist · pp. 17–18

Declaration of LLM usage

  • The checklist asks whether the paper describes LLM use when it is an important, original, or non-standard component of the core methods.It says declaration is not required for LLM use limited to writing, editing, or formatting that does not affect core methodology, rigor, or originality.Source: NeurIPS Paper Checklist · p. 18
  • The authors based model design on Genie [5] and say they made every development decision themselves.They used Claude Code to assist implementation, then reviewed the code, double-checked results, and used unit tests during development.Source: NeurIPS Paper Checklist · p. 19
  • AI tools were used to make the main figure more concise after the authors prepared its initial version.Source: NeurIPS Paper Checklist · p. 19

A Utilized diagnostics and actuators

Tables 1 and 2 summarize the diagnostics and actuators used to train IGNITE; the diagnostics were grouped and resampled.

B Details on the modality-specific tokenizers

The modality-specific tokenizers convert slow and fast time series and spectrograms into tokens for transformer processing, using different patching and encoding schemes.

Loading 2610.02515v1…