Paper deep dive
IGNITE Tokamak World Model Architecture
Peter Steiner, Azarakhsh Jalalvand, Nathaniel Chen, Kouroche Bouchiat, Ricardo Shousha, SangKyeun Kim, Egemen Kolemen
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 94%
Last extracted: 10/5/2026, 6:21:45 AM
Summary
The paper introduces IGNITE, a generative World Foundation Model (WFM) for simulating fusion plasma behavior in the DIII-D tokamak. Trained self-supervised on over a decade of unlabeled experimental data, IGNITE utilizes spatio-temporal tokenizers to embed multi-modal inputs (time-series, images, spectrograms) at varying time scales. Its core is an auto-regressive dynamics model based on a masked generative transformer that predicts future plasma states given initial conditions and actuator trajectories, enabling efficient AI-driven experimental planning.
Entities (12)
Relation Signals (7)
IGNITE → trainedon → DIII-D
confidence 100% · trained in a self-supervised manner from over a decade of unlabeled experimental data at the DIII-D National Fusion Facility
Egemen Kolemen → affiliatedwith → Princeton Plasma Physics Laboratory
confidence 95% · Affiliation: Princeton University, Princeton Plasma Physics Laboratory
Peter Steiner → affiliatedwith → Princeton University
confidence 95% · Affiliation: Princeton University
IGNITE → usestechnique → Finite Scalar Quantization
confidence 95% · We use Finite-Scalar-Quantization (FSQ) encoders to tokenize the diagnostics
TokaMind → developedby → IBM
confidence 90% · A collaboration between Mega Amp Spherical Tokamak (MAST) tokamak and IBM has proposed the multi-modal foundation model TokaMind
IGNITE → similarto → Genie
confidence 90% · The architecture of IGNITE is adapted from the Genie architecture
IGNITE → usesarchitecture → Masked Generative Image Transformer
confidence 90% · It is a decoder-only Masked Generative Image Transformer (maskGIT) transformer
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
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.
Tags
Links
- Source: https://arxiv.org/abs/2610.02515v1
- Canonical: https://arxiv.org/abs/2610.02515v1
Trouble viewing inline? Open PDF directly →
Full Text
74,824 characters extracted from source content.
Expand or collapse full text
PhysWorldAI IGNITE Tokamak World Model Architecture Peter Steiner Azarakhsh Jalalvand Nathaniel Chen Kouroche Bouchiat Affiliation: Department of Mechanical and Aerospace Engineering Affiliation: Princeton University Affiliation: Princeton, NJ Email: peter.steiner,azarakhsh.jalalvand,nathaniel,bouchiat@princeton.edu Ricardo Shousha SangKyeun Kim Affiliation: Princeton Plasma Physics Laboratory Affiliation: Princeton, NJ Email: rshousha,skim2@pppl.gov Egemen Kolemen Affiliation: Princeton University, Princeton Plasma Physics Laboratory Affiliation: Princeton, NJ Email: ekolemen@princeton.edu Abstract We introduce IGNITE, a generative World Foundation Model (WFM)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 Generative Artificial Intelligence (AI)has rapidly become a revolutionary tool for accelerating advances across a broad range of research fields. Its emergence may be comparable to the widespread adoption of computers decades ago, which fundamentally reshaped scientific research and enabled entirely new ways of working and discovery. Recent advances in AIcan be attributed to the development of the transformer architecture [36], and the democratization of powerful modern Graphics Processing Unit (GPU)hardware and large-scale High Performance Computing (HPC)clusters. Together, with the progression of self-supervised learning, these advances have enabled billion-parameter models to be trained on vast amounts of previously underutilized unlabeled data, overcoming one of the major limitations of conventional supervised learning: the need for large, carefully labeled datasets. This led to significant breakthroughs in many fields such as image and video processing [2, 11, Wang2023videomae, 5, 12], bioinformatic [13], and audio signal processing [34, 40, 31, 21, 22]. Another area in which such models have not yet, but could lead to major breakthroughs is fusion energy [8]. Nuclear fusion offers humanity a potentially clean, sustainable, and safe energy source, and decades of research and experimentation led to the physics basis of fusion power plants such as the Affordable, Robust, Compact (ARC)reactor [17], which is expected to start operations in the mid 20302030s with the plan to deliver 400 MW400\,MW net electrical power to the grid. However, much experimentation is still required to find scalable, high-performance plasma regimes [23, 38]. Running these experiments is expensive, requiring significant investment since it depends on competitive allocation and expensive hardware. Given the long history of recorded data for nuclear fusion, there is promising opportunity to learn full experiments using AImodels. Fusion is a particularly well-suited domain for self-supervised techniques because plasma behavior is extraordinarily complex, involving nonlinear and multiscale dynamics, hence, difficult to model and predict comprehensively. Consequently, accurate prediction and control of fusion plasmas using purely model-based approaches remain major challenges. At the same time, a large number of experimental fusion devices have been built around the world, spanning different configurations, operational capabilities, and diverse sets of diagnostics. Decades of experiments on these machines have generated enormous volumes of data, much of which remains underexplored. This combination of data richness and physical complexity provides exactly the kind of environment in which large generative or foundation models, trained using self-supervised approaches, could have a revolutionary impact. In this paper, we propose Integrated and Generative Network for Interactive Tokamak Emulation (IGNITE), a new generative AImodel designed to simulate fusion plasma behavior, trained on a large dataset of over 80008000 discharges (≈ 12 h12\,h of data) collected from more than a decade of experiments. It is not only the first large-scale world model for fusion plasma, but also the first multi-modal approach, working with signals at different time scales ranging from HzHz level to MHzMHz level, and different modalities such as time-series in different sampling frequencies, spectrograms, and a tokenizer design that is extensible to other diagnostics types such as image sequences. It is also fully controllable on a frame-by-frame level by supplying actuator trajectories. Similar to many research areas, AIhas had a strong impact on nuclear fusion research in the past years due to its strength in handling high dimensional, nonlinear data [1]. Several publications like [10, 30, 39] have shown the potential of using reinforcement learning for controlling nuclear fusion devices. A key point of these publications is that the deployed AImodels were not only successful in their accuracy, but that they were also deployed on real tokamaks to actively control or design experiments. These models often rely on a subset of curated data, and work on a subset of diagnostics or target a subset of physics. To address this, works that show the potential of AImodels for enabling cross-diagnostic analysis [20] and robustness against sensor failures during nuclear fusion experiments [32] suggest the possibility of a single model that is capable of working on multiple of these tasks in parallel like predicting several profiles in parallel. A recent wave of large-scale self-supervised models for nuclear fusion have been demonstrated this move towards a developing such a single model: A collaboration between Mega Amp Spherical Tokamak (MAST)tokamak and IBM has proposed the multi-modal foundation model TokaMind [4]. It is an efficient transformer-based encoder (9.3M parameters) that maps a context window of MAST diagnostics to a target window. The context window consists of time series, profiles, and image sequences. The model was pre-trained with a masked reconstruction task of a 50 ms50\,ms window of MASTdiagnostics. A key point is that the model uses a training-free signal embedding per diagnostic, allowing for different sampling frequencies (spanning 200 Hz200\,Hz to 500 kHz500\,kHz). After pre-training, the model was fine-tuned on several downstream tasks from reconstruction to prediction. In the latter case, the prediction horizon was up to 100 ms100\,ms, meaning two windows of each 50 ms50\,ms duration. It is worth noting that the prediction task was always a window-to-window regression rather than autoregressive forecasting. In another study PanoMHD [28] is developed as a multi-modal model for predicting plasma dynamics for control purposes at Korea Superconducting Tokamak Advanced Research (KSTAR). This model with 457M parameters predicts the next step magnetic-fluctuation spectrograms together with the scalar plasma performance, based on historic magnetic fluctuations, plasma performance, actuators, and plasma shape parameters. Like TokaMind, PanoMHD is optimized to only predict the next window of 50 ms50\,ms. FusionMAE is another example of such efforts which applies non-causal masked autoencoder reconstruction to actuator and diagnostic channels of the HL-3 tokamak. This 14M parameter model is capable of restoring lost signals and imputing compressed signals ranging from plasma current IpI_p to lower hybrid resonant heating PLHP_LH [42]. While TokaMind, PanoMHD, and FusionMAE establish that self-supervised pretraining on heterogeneous tokamak diagnostics yields transferable representations, they are formulated as conditioned predictors or reconstructors over a single short window. FusionMAE compresses a 10 ms10\,ms slice into an embedding and imputes missing channels, with no temporal-dynamics model, its authors noting that capturing dynamics would require an autoregressive reformulation. TokaMind maps a context window to a target window (≤100 ms≤$100\,ms$). PanoMHD predicts the next 50 ms50\,ms window of magnetic-fluctuation spectrograms from a short history. FusionMAE predicts 10ms given a consistent input sample-grid of 1 kHz1\,kHz. IGNITEbuilds on these ideas with methods from state-of-the-art spectrogram [22], video [2, 14, 37], and time-series models [21, 22] as well as world models like Genie [5]. Similar to Genie, we use spatiotemporal transformers [41] in our dynamic model. We use Finite-Scalar-Quantization (FSQ)encoders [27] to tokenize the diagnostics and actuators, and these tokens are passed to a dynamic model that autoregressively predicts the tokens for the next frame. 2 The IGNITE world model IGNITEis a generative multi-modal actuator-conditioned world model of the DIII-D tokamak. It takes a sequence of 2020 windows of 50 ms50\,ms duration of heterogeneous diagnostics plus seven actuator groups and autoregressively forecasts the diagnostics in the subsequent windows. Each modality is tokenized independently, and we use discrete FSQcodecs for all modalities. The world model thus learns dynamics in a fixed latent space. The architecture of IGNITE is adapted from the Genie architecture [5] and thus based on the transformer architecture [36]. We also use the memory-efficient spatiotemporal transformer architecture from [41] to balance the model capacity with computational constraints. Figure 1: Main outline of the IGNITE world model. As shown in Figure 1, IGNITE contains two key components: 1. modality-specific tokenizers that convert raw measurements into discrete tokens, 2. a dynamics model that, given an actuator trajectory and past frame tokens, predicts the tokens of the next frame. The model is trained in two phases: First, the tokenizers that are used for the dynamics model are trained for encoding and decoding diagnostics. The actuator embeddings and the dynamics model are then trained together for autoregressive rollout. 2.1 Per-modality tokenization Typical world models like [5, 29, 15, 25] work with uniformly sampled video data or very few modalities that have a common sampling frequency. To accurately simulate a DIII-D discharge, however, a caveat is to deal with sampling frequencies that differ in the order of magnitudes, spanning from 100 Hz100\,Hz to 1 MHz1\,MHz. Furthermore, the model has to handle different modalities like multi-channel time-series with low and high sampling frequencies, spectrograms, and further modalities like image sequences. As a result, modality-specific tokenizers are designed to deal with the unique properties of each modality. In general, we compress each diagnostic frame-based into a discrete token space using FSQencoders [27]. Each tokenizer takes diagnostics from a window of 50 ms50\,ms and encodes them into carefully tuned FSQencoders. The hyper-parameters and the loss functions for each specific modality are described further in this section. In this work, the modality specific encoders were crucial to process the diagnostics data of wildly different scales. The discrete tokens allow us to reduce the dimensionality and enable a high quality spectrogram generation even in high-frequency ranges. An overview of the exact number of channels, measured properties, and sampling frequencies of all diagnostics can be found in the appended Tables 1 and 2. Slow time series encoder Several diagnostics are an array of probes in the DIII-D device which provide spatially rich information about the measurements in different locations of plasma. Thomson Scattering (TS)provides electron density and temperature at the core and tangential areas of the tokamak cross section [6]. From Charge Exchange Recombination (CER), we use ion temperature and plasma rotation [7, 16] of the tangential subsystem. We also use measurements of the magnetic pitch angles as captured by the Motional Stark Effect (MSE)diagnostic [19, 18]. In total we have seven different categories of slow measurements, each supplied to an individual tokenizer. Each slow time series has been resampled to 100 Hz100\,Hz. Since IGNITE works with windows of 50 ms50\,ms duration, each window contains the measurements of all channels for five consecutive samples, i.e., a matrix of number of channels by 55. We refer to Table 1 for the number of channel as it depends on the diagnostic. Each tokenizer maps the value of a window to four tokens, each having a dimension of 128128. First, the samples in each window are patchified into non-overlapping patches, each carrying spatio-temporal information. The patches are then flattened to a window and linearly projected to 128128. Two learned positional embeddings for spatial and temporal encoding are added. Finally, a transformer with four layers and four heads is used to learn spatio-temporal dependencies in the tokens. The resulting tokens of size 4×1284×$128$ are used as continuous features handed to the FSQquantizer. Fast time series encoder While the diagnostics mentioned in the previous paragraph provide spatio-temporal information in slower time pace, other signals like DαD_α filterscopes capture significantly more temporal information. These signals are sampled at 10 kHz10\,kHz and capture fast transient events like Edge Localized Modes (ELM). These occur as spikes in DαD_α signals [9]. In this paper, we use the eight filterscope signals that are most commonly used for DIII-D experiments. Each window of 50 ms50\,ms contains now eight channels each having 500500 samples. Similar to slow time series, we patchify the signals first, here only in the temporal direction as the amount of spatial information is limited for DαD_α filterscopes. The patches of all eight channels belonging to the same temporal position are concatenated and represent one token. So, one token carries all the information of eight channels in a window of 10 ms10\,ms duration. We found that it is important to have a much more expressive latent feature space for fast time series. As a consequence, a solution would be to linearly project the tokens to 256256 dimensions and then add a learned per-patch positional embedding for learning temporal dependencies. Finally, a transformer with six layers and eight heads is used to learn spatio-temporal dependencies in the tokens. The resulting tokens of size 4×2564×$256$ are used as continuous features handed to the FSQquantizer. Spectrogram encoder The DIII-D tokamak has several high-resolution diagnostics that are typically analyzed in the spectral domain. A few examples include Beam Emission Spectroscopy (BES)[26] for fluctuation analysis, Bi-color CO2 Interferometry (CO2)[35] for density measurements, Electron Cyclotron Emission (ECE)radiometer [3] for electron temperature analysis, and Magnetics High Resolution (MHR)[33] for equilibrium reconstruction, plasma shape control and mode analysis. In this paper we resampled all diagnostics to 500 kHz500\,kHz before model training to simplify the data unification and because most of the mode activities can be captured in a frequency range of 0 kHz to 250 kHz0\,kHz250\,kHz. We computed spectral features using a window size of 10241024 samples, a hop size of 256256 and a Hann window. And we disregard the constant DC offset bin in order to get a spectrogram height of 512512 bins. We use a logarithmic amplitude log10(⋅+1×10−5) _10(·+$1×10^-5$) and standardize the resulting spectrograms per frequency bin. Each window of 50 ms50\,ms contains 9898 frames. We patchify the spectrograms in non-overlapping patches of 16×16$16$×$16$ over time and frequency, and we stack the patches of all channels. This leads to differently sized feature vectors that are tokenized subsequently, but to consistently 192192 tokens per spectrogram modality. The small patch size is a compromise between limiting the number of tokens required for spectrograms and frequency resolution, required for an accurate mode reconstruction. After a linear projection to 256256 dimension of all tokens, we add trainable positional encodings for time and frequency to the tokens, and use again a transformer with six layers and eight heads to learn spatio-temporal relationships between the tokens. The resulting tokens are again handed to a FSQquantizer. Output heads The transformer backbone outputs the predicted token sequence, which contains all tokens for diagnostics in the same order as the input. To decode the diagnostics at the next time step, we thus split the tokens and pass all tokens belonging to a particular diagnostic to a separate decoder that is responsible for reconstructing the actual diagnostic from the predicted token. The decoders are designed symmetrically to the encoders. For slow time series we thus use a linear projection layer, and all convolutional operations are reversed by using transposed convolutions. Since the autoregressive roll-out takes place solely in the token space, it is not necessary to decode every time step. Instead, the tokens cycle through the backbone during autoregressive roll-out. Decoding is only necessary to compare against the ground truth. During autoregressive model training, we thus use the decoders only for comparing a particular roll-out step with the ground truth when we want to calculate a loss. 2.2 Shared Transformer Backbone We adapt the dynamics model from Genie to be used in IGNITE. It is a decoder-only Masked Generative Image Transformer (maskGIT)transformer that processes FSQcodes. The backbone is a factorized space-time transformer as in [41]. Since the dynamics model operates purely in the discrete code space, it is effectively a “language model over plasma-state tokens”. At each time step, it receives the tokens of the different modalities. IGNITE works with a context of 2020 time windows. The dynamics model thus receives tokens of 2020 frames. Since the tokens have different dimensions, as described in Section 2.1, the first step is to embed them to the model dimension of the dynamics model (10241024) with a linear projection. The actuator signals of each window are passed separately to the backbone. In each window of 50 ms50\,ms, we first of all calculate the mean value per channel and per actuator signal to obtain a vector of 8888 dimensions. This is again linearly embedded to the model dimension, and then Each of the transformer blocks first calculates cross-attention across all tokens within its own frame for cross-modality mixing, followed by causal temporal self-attention. In this case, each token position attends itself at earlier frames, and this allows information to move through time. Importantly, information can only be passed forward. The output of the dynamic model are then the codes for the next window of 50 ms50\,ms. To simulate a full DIII-D discharge, the output is then appended to the context, and the prediction process repeats until the end of a discharge, simulating the discharge in an autoregressive manner. The predicted codes are finally decoded back into the original modalities (time-series, spectrograms, camera frames) via the frozen decoders, for a comparison against the actual measurements. The dynamics model has 10241024 dimensions and 4848 transformer layers, which is a trade-off between training time and accuracy. 2.3 Training methods IGNITE has a two-stage training process: First, we pre-train the codecs for each modality. Afterwards, the codecs are fixed, and the dynamics model is trained to autoregressively predict the full flat-top evolution of a discharge. The training stages are describe subsequently. Tokenizer pre-training The first step of the whole IGNITE training pipeline is to train the codebooks for the modalities. This is considered to be a pre-training step, during which the diagnostic encoders and decoders are trained in order to represent diagnostics by a FSQ-based codebook. For this purpose, each window of 50 ms50\,ms duration is treated independently. Dynamics model training The dynamics model closely follows the dynamics model in Genie [5]. After encoding each modality into integer tokens, each modality m∈1,…,Mm∈1,…,M (M=15M=15) contributes a specific amount of tokens per frame from its codebook of size VmV_m. We use the maskGITapproach for training the dynamics model. A training sample consists of 100100 consecutive frames, each of which being encoded into integer codes. In each window, a randomly selected amount of tokens is masked out, with the masking probability randomize over windows. In this way, some windows have to be predicted completely, whereas other windows are nearly complete. The actuator states in each window are projected to the target dimension (10241024), and added to the corresponding token window. Three more embeddings for modality, positional and temporal encoding are added to each modality. Afterwards, the different modalities are concatenated to form the input to the transformer. The backbone now consists of 2424 spatio-temporal transformer blocks. Each block has a pre-normalization layer and a feed-forward network. As in Genie, we use causal masking in the temporal attention, ensuring that only historical information can attend to a given time frame. The training objective is masked-token prediction. The loss function is the cross-entropy between the class of the target token codec and the predicted codec in the masked sections. The masking and hence the cross entropy is done independently per diagnostic. As a consequence, the final loss is the mean of per-modality losses. We use the AdamW optimizer [24] for weight optimization with weight decay, β=(0.9,β2)β=($0.9$, _2), and a warmup-cosine schedule, adapted from Genie. Since we are parallelizing over multiple ranks (eight GPUnodes, each having eight GPUs), the gradients are averaged across ranks by explicit all-reduce. The dataset suffers from a lot of missing diagnostics. In the case of absent diagnostics during the training process, these are still passed through the model, but the contribution of these diagnostics to the loss function is zero weighted. This is necessary to still allow a modality to receive a gradient signal, which would lead to detection of unused parameters. During inference, the first 2020 real frames are rollout seeds. Afterwards, the dynamics model generates each next frame from all-masked in K confidence-ordered unmasking steps, keeping a cosine-decaying masked fraction, sampling the most confident tokens per modality. The predicted codes are appended to the input sequence, and the process repeats. 3 Dataset The IGNITE foundation model was trained on a large dataset of experimental data of the DIII-D tokamak, which has one of the most sophisticated sets of diagnostics in the world. The subset of diagnostics used for the experiments in this paper is summarized in Table 1, and the actuators are summarized in Table 2. For preparing the original dataset, we collected data from all DIII-D discharges between 20182018 and 20262026. Based on the plasma current, we then filtered out successful runs by eliminating disrupted shots where the correlation between the planned plasma current and the actual plasma current falls below a correlation of 0.80.8. As a result, we get 86008600 discharges. For model training and evaluation, we randomly split the discharges in 90 %90\,\% training, 5 %5\,\% validation, and 5 %5\,\% test data. Based on the training dataset, we calculated coefficients for normalization. For each diagnostic, we subtracted the global channel mean and divided by the standard deviation. Since we were dealing with a substantial amount of missing data, we masked these with zeros and excluded these samples from the training loss via the binary mask Mk(d)M^(d)_k. We skipped the first second of each shot, because this initial phase is dominated by a lot of transient effects. We divided the remaining parts of each discharge into non-overlapping chunks of 50 ms50\,ms, which were used for model training. 4 Results During model development, we started with overfitting the IGNITE architecture at different scaling levels based on 1616 shots that belonged to a single experiment. We observed that we needed a significantly higher learning rate of 1×10−31×10^-3 compared to Genie (1×10−41×10^-4), and that we had to train the model for up to 8080 epochs in order to memorize the training set. 4.1 Learning curves Next, we increased the dataset size to 10001000 training shots, randomly drawn from the entire training dataset, and trained IGNITE models with weight decay and dropout. The learning curves of training these models for 5050 epochs are visualized in Figure 2a, and they show that IGNITE tends to overfit on the training dataset strongly. Dropout regularization can slow down the divergence between training and validation dataset. However, the validation loss still started to increase between 1212 and 1515 epochs. The training loss was not affected by regularization. (a) Figure 2: IGNITE learning curves for different regularization levels and dataset sizes. Figure 2a: A subset of 10001000 shots leads to overfitting, despite regularization. Next, we used the complete training dataset to investigate the impact of dataset size. Following Figure 2a, the minimal validation loss was hardly affected by regularization. Hence, we decided to keep dropout at 0.30.3 and weight decay at 0.10.1 to still enable light regularization. Due to the model size, it becomes clear that, despite the model being capable of overfitting to very few shots, generalization requires significantly more data. Even in an early training stage, the model performance on the validation dataset can now exceed the model trained on 10001000 shots. Figure 3: IGNITE predicts the DIII-D discharge 190 735190\,735 from the held-out validation set. 4.2 Example validation output This evaluation is based on a representative example from the validation set, namely DIII-D 190 735190\,735. It was not used for any kind of training process, only for monitoring the validation loss across the training progress. Since we did not use early stopping based on validation, we claim that we have not used this shot for any model optimization. From Figure 3 can be observed that IGNITE can predict the full shot trajectory of different modalities given the first second of the shot (2020 frames) as an initial seed plus a given actuator trajectory. The spectrograms still suffer from fine-grained mode structures. Especially for ECE, it becomes clear that the model still has difficulties predicting mode structures in higher frequency ranges. But overall, the prediction accuracy is already at a satisfying level. 5 Conclusions and future work We introduced IGNITE – a world model for simulating DIII-D discharges that has been trained on historical experimental data from the DIII-D nuclear fusion device. We showed that a sufficiently large training dataset is crucial for generalization. IGNITE can autoregressively predict signals across a broad variety of modalities. During model development, we faced several challenges that make our task different to the original Genie architecture. Genie had a training dataset of more than 200 000 h200\,000\,h of video clips. After filtering, they ended up with 6.8×1066.8×10^6 video clips with 16 s16\,s duration. For IGNITE, we used 86008600 sequences of 6 s6\,s duration for model training. While Genie is a unimodal world model, ingesting only video data at a constant frame rate, IGNITE ingests 1414 different modalities in 44 different groups, each having a different sampling rate. So, we are dealing with a multi-rate problem. While video data is recorded in a clean environment, DIII-D experimental data is collected in a dynamical environment that is impacted by external factors. Diagnostics and actuators can be malfunctioning or not working as programmed. The data quality can suffer from insufficient calibration or noise. And the diagnostic system was growing historically. Channels were added over time, making it difficult to simply use historical data. Future work includes a physics-based evaluation of this model. We want this to become a useful tool for DIII-D researchers to design experiments, to simulate actuator defects that can happen during experiments. This requires us to study the impact of bad/unavailable inputs during training and inference. Acknowledgments and Disclosure of Funding The authors gratefully acknowledge the collaboration of the DIII-D Team supported by the U.S. Department of Energy, Office of Science, Office of Fusion Energy Sciences, using the National Fusion Facility, a Department of Energy (DOE)Office of Science user facility, under Award No. DE-FC02-04ER54698, DE-AC02-09CH11466, DE-SC0024527, DE-SC0026429. Additional support was provided by the Princeton Laboratory for Artificial Intelligence under Award No. 2025-97. This research used resources of the Oak Ridge Leadership Computing Facility at the Oak Ridge National Laboratory, which is supported by the Advanced Scientific Computing Research programs in the Office of Science of the U.S. Department of Energy under Contract No. DE-AC05-00OR22725. Disclaimer: This report was prepared as an account of work sponsored by an agency of the United States Government. Neither the United States Government nor any agency thereof, nor any of their employees, makes any warranty, express or implied, or assumes any legal liability or responsibility for the accuracy, completeness, or usefulness of any information, apparatus, product, or process disclosed, or represents that its use would not infringe privately owned rights. Reference herein to any specific commercial product, process, or service by trade name, trademark, manufacturer, or otherwise does not necessarily constitute or imply its endorsement, recommendation, or favoring by the United States Government or any agency thereof. The views and opinions of authors expressed herein do not necessarily state or reflect those of the United States Government or any agency thereof. References [1] R. Anirudh, R. Archibald, M. S. Asif, M. M. Becker, S. Benkadda, P. Bremer, R. H. S. Budé, C. S. Chang, L. Chen, R. M. Churchill, J. Citrin, J. A. Gaffney, A. Gainaru, W. Gekelman, T. Gibbs, S. Hamaguchi, C. Hill, K. Humbird, S. Jalas, S. Kawaguchi, G. Kim, M. Kirchen, S. Klasky, J. L. Kline, K. Krushelnick, B. Kustowski, G. Lapenta, W. Li, T. Ma, N. J. Mason, A. Mesbah, C. Michoski, T. Munson, I. Murakami, H. N. Najm, K. E. J. Olofsson, S. Park, J. L. Peterson, M. Probst, D. Pugmire, B. Sammuli, K. Sawlani, A. Scheinker, D. P. Schissel, R. J. Shalloo, J. Shinagawa, J. Seong, B. K. Spears, J. Tennyson, J. Thiagarajan, C. M. Ticoş, J. Trieschmann, J. v. Dijk, B. V. Essen, P. Ventzek, H. Wang, J. T. L. Wang, Z. Wang, K. Wende, X. Xu, H. Yamada, T. Yokoyama, and X. Zhang (2023) 2022 review of data-driven plasma science. IEEE Transactions on Plasma Science 51 (7), p. 1750–1838. External Links: Document, ISSN 1939-9375 Cited by: §1. [2] A. Arnab, M. Dehghani, G. Heigold, C. Sun, M. Lučić, and C. Schmid (2021) ViViT: a video vision transformer. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), Vol. , p. 6816–6826. External Links: Document, ISSN 2380-7504 Cited by: §1, §1. [3] M. E. Austin and J. Lohr (2003) Electron cyclotron emission radiometer upgrade on the diii-d tokamak. Review of Scientific Instruments 74 (3), p. 1457–1459. External Links: ISSN 0034-6748, Document, Link, https://pubs.aip.org/aip/rsi/article-pdf/74/3/1457/19084026/1457_1_online.pdf Cited by: §2.1. [4] T. Boschi, A. Loreti, N. C. Amorisco, R. H. Ordonez-Hurtado, C. Rousseau, G. K. Holt, E. Székely, A. Whittle, S. Jackson, A. Agnello, S. Pamela, A. Pascale, R. Akers, J. B. Moreno, V. Alexandrov, and M. Zayats (2026) TokaMind: a multi-modal transformer foundation model for tokamak plasma dynamics. External Links: 2602.15084, Link Cited by: §1. [5] J. Bruce, M. Dennis, A. Edwards, J. Parker-Holder, Y. Shi, E. Hughes, M. Lai, A. Mavalankar, R. Steigerwald, C. Apps, Y. Aytar, S. Bechtle, F. Behbahani, S. Chan, N. Heess, L. Gonzalez, S. Osindero, S. Ozair, S. Reed, J. Zhang, K. Zolna, J. Clune, N. de Freitas, S. Singh, and T. Rocktäschel (2024) Genie: generative interactive environments. External Links: 2402.15391, Link Cited by: NeurIPS Paper Checklist, §1, §1, §2.1, §2.3, §2. [6] T. N. Carlstrom, F. Glass, X. Du, A. G. McLean, D. Taussig, and R. L. Boivin (2018) Thomson scattering measurements on diii-d using in-vessel laser mirrors and lenses to diagnose a new divertor location. Review of Scientific Instruments 89 (10), p. 10C111. External Links: ISSN 0034-6748, Document, Link Cited by: §2.1. [7] C. Chrystal, S. R. Haskey, K. H. Burrell, B. A. Grierson, and C. S. Collins (2021) Charge exchange recombination spectroscopy measurements of diii-d poloidal rotation with poloidal asymmetry in angular rotation. Review of Scientific Instruments 92 (4), p. 043518. External Links: ISSN 0034-6748, Document, https://pubs.aip.org/aip/rsi/article-pdf/doi/10.1063/5.0043087/14030718/043518_1_online.pdf Cited by: §2.1. [8] R. M. Churchill (2025) AI foundation models for experimental fusion tasks. Frontiers in Physics 12. External Links: Link, Document, ISSN 2296-424X Cited by: §1. [9] R. J. Colchin, D. L. Hillis, R. Maingi, C. C. Klepper, and N. H. Brooks (2003) The filterscope. Review of Scientific Instruments 74 (3), p. 2068–2070. External Links: ISSN 0034-6748, Document, Link Cited by: §2.1. [10] J. Degrave, F. Felici, J. Buchli, M. Neunert, B. Tracey, F. Carpanese, T. Ewalds, R. Hafner, A. Abdolmaleki, D. de las Casas, C. Donner, L. Fritz, C. Galperti, A. Huber, J. Keeling, M. Tsimpoukelli, J. Kay, A. Merle, J. Moret, S. Noury, F. Pesamosca, D. Pfau, O. Sauter, C. Sommariva, S. Coda, B. Duval, A. Fasoli, P. Kohli, K. Kavukcuoglu, D. Hassabis, and M. Riedmiller (2022) Magnetic control of tokamak plasmas through deep reinforcement learning. Nature 602 (7897), p. 414–419. External Links: ISSN 1476-4687, Document Cited by: §1. [11] M. Dehghani, J. Djolonga, B. Mustafa, P. Padlewski, J. Heek, J. Gilmer, A. P. Steiner, M. Caron, R. Geirhos, I. Alabdulmohsin, R. Jenatton, L. Beyer, M. Tschannen, A. Arnab, X. Wang, C. Riquelme Ruiz, M. Minderer, J. Puigcerver, U. Evci, M. Kumar, S. V. Steenkiste, G. F. Elsayed, A. Mahendran, F. Yu, A. Oliver, F. Huot, J. Bastings, M. Collier, A. A. Gritsenko, V. Birodkar, C. N. Vasconcelos, Y. Tay, T. Mensink, A. Kolesnikov, F. Pavetic, D. Tran, T. Kipf, M. Lucic, X. Zhai, D. Keysers, J. J. Harmsen, and N. Houlsby (2023) Scaling vision transformers to 22 billion parameters. In Proceedings of the 40th International Conference on Machine Learning, A. Krause, E. Brunskill, K. Cho, B. Engelhardt, S. Sabato, and J. Scarlett (Eds.), Proceedings of Machine Learning Research, Vol. 202, p. 7480–7512. External Links: Link Cited by: §1. [12] A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby (2021) An image is worth 16x16 words: transformers for image recognition at scale. In International Conference on Learning Representations, External Links: Link Cited by: §1. [13] F. Guo, R. Guan, Y. Li, Q. Liu, X. Wang, C. Yang, and J. Wang (2025) Foundation models in bioinformatics. National Science Review 12 (4), p. nwaf028. External Links: ISSN 2095-5138, Document, Link, https://academic.oup.com/nsr/article-pdf/12/4/nwaf028/61626536/nwaf028.pdf Cited by: §1. [14] A. Gupta, S. Tian, Y. Zhang, J. Wu, R. Martín-Martín, and L. Fei-Fei (2023) MaskViT: masked visual pre-training for video prediction. In The Eleventh International Conference on Learning Representations, External Links: Link Cited by: §1. [15] D. Ha and J. Schmidhuber (2018) Recurrent world models facilitate policy evolution. In Advances in Neural Information Processing Systems, S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, and R. Garnett (Eds.), Vol. 31, p. . External Links: Link Cited by: §2.1. [16] S. R. Haskey, B. A. Grierson, L. Stagner, C. Chrystal, A. Bortolon, and F. M. Laggner (2022) Details of the neutral energy distribution and ionization source using spectrally resolved balmer-alpha measurements on diii-d. Review of Scientific Instruments 93 (10), p. 103504. External Links: ISSN 0034-6748, Document, https://pubs.aip.org/aip/rsi/article-pdf/doi/10.1063/5.0101854/19816848/103504_1_online.pdf Cited by: §2.1. [17] J. C. Hillesheim, A. J. Creely, T. H. Eich, N. T. Howard, N. Leuthold, R. Sweeney, A. LeViness, A. O. Nelson, L. Nichols, R. A. Tinguely, and et al. (2026) Overview of the physics basis for the arc fusion power plant. Journal of Plasma Physics 92 (3), p. E69. External Links: Document Cited by: §1. [18] C. T. Holcomb, M. A. Makowski, S. L. Allen, W. H. Meyer, and M. A. Van Zeeland (2008) Overview of equilibrium reconstruction on diii-d using new measurements from an expanded motional stark effect diagnostica). Review of Scientific Instruments 79 (10), p. 10F518. External Links: ISSN 0034-6748, Document, https://pubs.aip.org/aip/rsi/article-pdf/doi/10.1063/1.2955711/15669216/10f518_1_online.pdf Cited by: §2.1. [19] C. T. Holcomb, M. A. Makowski, R. J. Jayakumar, S. A. Allen, R. M. Ellis, R. Geer, D. Behne, K. L. Morris, L. G. Seppala, and J. M. Moller (2006) Motional stark effect diagnostic expansion on diii-d for enhanced current and er profile measurements. Review of Scientific Instruments 77 (10), p. 10E506. External Links: ISSN 0034-6748, Document, https://pubs.aip.org/aip/rsi/article-pdf/doi/10.1063/1.2235812/11170995/10e506_1_online.pdf Cited by: §2.1. [20] A. Jalalvand, S. Kim, J. Seo, Q. Hu, M. Curie, P. Steiner, A. O. Nelson, Y. Na, and E. Kolemen (2025) Multimodal super-resolution: discovering hidden physics and its application to fusion plasmas. Nature Communications 16 (1), p. 8506. External Links: ISSN 2041-1723, Document Cited by: §1. [21] J. Kong, J. Kim, and J. Bae (2020) HiFi-gan: generative adversarial networks for efficient and high fidelity speech synthesis. In Advances in Neural Information Processing Systems, H. Larochelle, M. Ranzato, R. Hadsell, M.F. Balcan, and H. Lin (Eds.), Vol. 33, p. 17022–17033. External Links: Link Cited by: §1, §1. [22] R. Langman, A. Jukić, K. Dhawan, N. K. Rao, and J. Li (2025) Spectral codecs: improving non-autoregressive speech synthesis with spectrogram-based audio codecs. External Links: 2406.05298, Link Cited by: §1, §1. [23] J.D. Lawson (1957) Some criteria for a power producing thermonuclear reactor. Proceedings of the Physical Society. Section B 70 (1), p. 6. External Links: Document Cited by: §1. [24] I. Loshchilov and F. Hutter (2019) Decoupled weight decay regularization. In International Conference on Learning Representations, External Links: Link Cited by: §2.3. [25] L. Maes, Q. L. Lidec, D. Scieur, Y. LeCun, and R. Balestriero (2026) LeWorldModel: stable end-to-end joint-embedding predictive architecture from pixels. External Links: 2603.19312, Link Cited by: §2.1. [26] G. R. McKee, R. Ashley, R. Durst, R. Fonck, M. Jakubowski, K. Tritz, K. Burrell, C. Greenfield, and J. Robinson (1999) The beam emission spectroscopy diagnostic on the diii-d tokamak. Review of Scientific Instruments 70 (1), p. 913–916. External Links: ISSN 0034-6748, Document, Link, https://pubs.aip.org/aip/rsi/article-pdf/70/1/913/19142687/913_1_online.pdf Cited by: §2.1. [27] F. Mentzer, D. Minnen, E. Agustsson, and M. Tschannen (2024) Finite scalar quantization: VQ-VAE made simple. In The Twelfth International Conference on Learning Representations, External Links: Link Cited by: §1, §2.1. [28] H. Noh, C. Heo, X. Gao, and Y. Na (2026) PanoMHD: multimodal modelling of plasma dynamics towards tokamak control. External Links: 2603.02672, Link Cited by: §1. [29] NVIDIA, :, N. Agarwal, A. Ali, M. Bala, Y. Balaji, E. Barker, T. Cai, P. Chattopadhyay, Y. Chen, Y. Cui, Y. Ding, D. Dworakowski, J. Fan, M. Fenzi, F. Ferroni, S. Fidler, D. Fox, S. Ge, Y. Ge, J. Gu, S. Gururani, E. He, J. Huang, J. Huffman, P. Jannaty, J. Jin, S. W. Kim, G. Klár, G. Lam, S. Lan, L. Leal-Taixe, A. Li, Z. Li, C. Lin, T. Lin, H. Ling, M. Liu, X. Liu, A. Luo, Q. Ma, H. Mao, K. Mo, A. Mousavian, S. Nah, S. Niverty, D. Page, D. Paschalidou, Z. Patel, L. Pavao, M. Ramezanali, F. Reda, X. Ren, V. R. N. Sabavat, E. Schmerling, S. Shi, B. Stefaniak, S. Tang, L. Tchapmi, P. Tredak, W. Tseng, J. Varghese, H. Wang, H. Wang, H. Wang, T. Wang, F. Wei, X. Wei, J. Z. Wu, J. Xu, W. Yang, L. Yen-Chen, X. Zeng, Y. Zeng, J. Zhang, Q. Zhang, Y. Zhang, Q. Zhao, and A. Zolkowski (2025) Cosmos world foundation model platform for physical ai. External Links: 2501.03575, Link Cited by: §2.1. [30] J. Seo, S. Kim, A. Jalalvand, R. Conlin, A. Rothstein, J. Abbate, K. Erickson, J. Wai, R. Shousha, and E. Kolemen (2024) Avoiding fusion plasma tearing instability with deep reinforcement learning. Nature 626 (8000), p. 746–751. External Links: ISSN 1476-4687, Document Cited by: §1. [31] J. Shen, R. Pang, R. J. Weiss, M. Schuster, N. Jaitly, Z. Yang, Z. Chen, Y. Zhang, Y. Wang, R. Skerrv-Ryan, R. A. Saurous, Y. Agiomvrgiannakis, and Y. Wu (2018) Natural tts synthesis by conditioning wavenet on mel spectrogram predictions. In 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Vol. , p. 4779–4783. External Links: Document, ISSN 2379-190X Cited by: §1. [32] R. Shousha, J. Seo, K. Erickson, Z. Xing, S. Kim, J. Abbate, and E. Kolemen (2023) Machine learning-based real-time kinetic profile reconstruction in diii-d. Nuclear Fusion 64 (2), p. 026006. External Links: Document Cited by: §1. [33] E. J. Strait (2006) Magnetic diagnostic system of the diii-d tokamak. Review of Scientific Instruments 77 (2), p. 023502. External Links: ISSN 0034-6748, Document, Link, https://pubs.aip.org/aip/rsi/article-pdf/doi/10.1063/1.2166493/16053320/023502_1_online.pdf Cited by: §2.1. [34] A. van den Oord, S. Dieleman, H. Zen, K. Simonyan, O. Vinyals, A. Graves, N. Kalchbrenner, A. Senior, and K. Kavukcuoglu (2016) WaveNet: A Generative Model for Raw Audio. In 9th ISCA Workshop on Speech Synthesis Workshop (SSW 9), p. 125. Cited by: §1. [35] M. A. Van Zeeland and T. N. Carlstrom (2004) Phase error correction method for a vibration compensated interferometer. Review of Scientific Instruments 75 (10), p. 3423–3425. External Links: ISSN 0034-6748, Document, Link, https://pubs.aip.org/aip/rsi/article-pdf/75/10/3423/19251218/3423_1_online.pdf Cited by: §2.1. [36] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin (2017) Attention is all you need. In Advances in Neural Information Processing Systems, I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett (Eds.), Vol. 30, p. . External Links: Link Cited by: §1, §2. [37] R. Villegas, M. Babaeizadeh, P. Kindermans, H. Moraldo, H. Zhang, M. T. Saffar, S. Castro, J. Kunze, and D. Erhan (2023) Phenaki: variable length video generation from open domain textual descriptions. In International Conference on Learning Representations, External Links: Link Cited by: §1. [38] F. Wagner, G. Becker, K. Behringer, D. Campbell, A. Eberhagen, W. Engelhardt, G. Fussmann, O. Gehre, J. Gernhardt, G. v. Gierke, G. Haas, M. Huang, F. Karger, M. Keilhacker, O. Klüber, M. Kornherr, K. Lackner, G. Lisitano, G. G. Lister, H. M. Mayer, D. Meisel, E. R. Müller, H. Murmann, H. Niedermeyer, W. Poschenrieder, H. Rapp, H. Röhr, F. Schneider, G. Siller, E. Speth, A. Stäbler, K. H. Steuer, G. Venus, O. Vollmer, and Z. Yü (1982) Regime of improved confinement and high beta in neutral-beam-heated divertor discharges of the asdex tokamak. Phys. Rev. Lett. 49, p. 1408–1412. External Links: Document, Link Cited by: §1. [39] A. M. Wang, A. Pau, C. Rea, O. So, C. Dawson, O. Sauter, M. D. Boyer, A. Vu, C. Galperti, C. Fan, A. Merle, Y. Poels, C. Venturini, F. Felici, S. Marchioni, and T. C. V. Team (2025) Learning plasma dynamics and robust rampdown trajectories with predict-first experiments at tcv. Nature Communications 16 (1), p. 8877. External Links: ISSN 2041-1723, Document Cited by: §1. [40] Y. Wang, R.J. Skerry-Ryan, D. Stanton, Y. Wu, R. J. Weiss, N. Jaitly, Z. Yang, Y. Xiao, Z. Chen, S. Bengio, Q. Le, Y. Agiomyrgiannakis, R. Clark, and R. A. Saurous (2017) Tacotron: Towards End-to-End Speech Synthesis. In Interspeech 2017, p. 4006–4010. External Links: Document, ISSN 2958-1796 Cited by: §1. [41] M. Xu, W. Dai, C. Liu, X. Gao, W. Lin, G. Qi, and H. Xiong (2021) Spatial-temporal transformer networks for traffic flow forecasting. External Links: 2001.02908, Link Cited by: §1, §2.2, §2. [42] Z. Yang, Z. Yang, W. Tian, J. Li, X. Sun, G. Zheng, S. Liu, N. Wu, R. Li, Z. Xu, B. Li, Z. Shi, Z. Gao, W. Chen, X. Ji, M. Xu, and W. Zhong (2026) FusionMAE, a self-supervised pretrained model to optimize and simplify diagnostic and control of fusion plasma. Communications Physics. External Links: ISSN 2399-3650, Document Cited by: §1. NeurIPS Paper Checklist The checklist is designed to encourage best practices for responsible machine learning research, addressing issues of reproducibility, transparency, research ethics, and societal impact. Do not remove the checklist: The papers not including the checklist will be desk rejected. The checklist should follow the references and follow the (optional) supplemental material. The checklist does NOT count towards the page limit. Please read the checklist guidelines carefully for information on how to answer these questions. For each question in the checklist: • You should answer [Yes] , [No] , or [N/A] . • [N/A] means either that the question is Not Applicable for that particular paper or the relevant information is Not Available. • Please provide a short (1–2 sentence) justification right after your answer (even for [N/A] ). The checklist answers are an integral part of your paper submission. They are visible to the reviewers, area chairs, senior area chairs, and ethics reviewers. You will also be asked to include it (after eventual revisions) with the final version of your paper, and its final version will be published with the paper. The reviewers of your paper will be asked to use the checklist as one of the factors in their evaluation. While [Yes] is generally preferable to [No] , it is perfectly acceptable to answer [No] provided a proper justification is given (e.g., error bars are not reported because it would be too computationally expensive” or “we were unable to find the license for the dataset we used”). In general, answering [No] or [N/A] is not grounds for rejection. While the questions are phrased in a binary way, we acknowledge that the true answer is often more nuanced, so please just use your best judgment and write a justification to elaborate. All supporting evidence can appear either in the main paper or the supplemental material, provided in appendix. If you answer [Yes] to a question, in the justification please point to the section(s) where related material for the question can be found. IMPORTANT, please: • Delete this instruction block, but keep the section heading “NeurIPS Paper Checklist", • Keep the checklist subsection headings, questions/answers and guidelines below. • Do not modify the questions and only use the provided macros for your answers. 1. Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: We have trained a world model for predicting DIII-D discharges. In the abstract and introduction, we claim that we can predict the temporal evolution for a given sequence, based on an initial state and an actuator trajectory. We experimental validate this with objective nRMSE measures, by showing the learning curves of the model, and discussing limitations such as generalization and overfitting. Guidelines: • The answer [N/A] means that the abstract and introduction do not include the claims made in the paper. • The abstract and/or introduction should clearly state the claims made, including the contributions made in the paper and important assumptions and limitations. A [No] or [N/A] answer to this question will not be perceived well by the reviewers. • The claims made should match theoretical and experimental results, and reflect how much the results can be expected to generalize to other settings. • It is fine to include aspirational goals as motivation as long as it is clear that these goals are not attained by the paper. 2. Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We discuss that this world model is focused on the DIII-D nuclear fusion device and we do not claim that this can be transferred to a different machine. We also do not claim that we can generate physically validated diagnostic evolutions when changing the actuator trajectories. We explicitly state that this is part of the future work. Guidelines: • The answer [N/A] means that the paper has no limitation while the answer [No] means that the paper has limitations, but those are not discussed in the paper. • The authors are encouraged to create a separate “Limitations” section in their paper. • The paper should point out any strong assumptions and how robust the results are to violations of these assumptions (e.g., independence assumptions, noiseless settings, model well-specification, asymptotic approximations only holding locally). The authors should reflect on how these assumptions might be violated in practice and what the implications would be. • The authors should reflect on the scope of the claims made, e.g., if the approach was only tested on a few datasets or with a few runs. In general, empirical results often depend on implicit assumptions, which should be articulated. • The authors should reflect on the factors that influence the performance of the approach. For example, a facial recognition algorithm may perform poorly when image resolution is low or images are taken in low lighting. Or a speech-to-text system might not be used reliably to provide closed captions for online lectures because it fails to handle technical jargon. • The authors should discuss the computational efficiency of the proposed algorithms and how they scale with dataset size. • If applicable, the authors should discuss possible limitations of their approach to address problems of privacy and fairness. • While the authors might fear that complete honesty about limitations might be used by reviewers as grounds for rejection, a worse outcome might be that reviewers discover limitations that aren’t acknowledged in the paper. The authors should use their best judgment and recognize that individual actions in favor of transparency play an important role in developing norms that preserve the integrity of the community. Reviewers will be specifically instructed to not penalize honesty concerning limitations. 3. Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [N/A] Justification: This is an applied AIpaper. The results are practical and all originate from experiments to optimize the IGNITE model. Guidelines: • The answer [N/A] means that the paper does not include theoretical results. • All the theorems, formulas, and proofs in the paper should be numbered and cross-referenced. • All assumptions should be clearly stated or referenced in the statement of any theorems. • The proofs can either appear in the main paper or the supplemental material, but if they appear in the supplemental material, the authors are encouraged to provide a short proof sketch to provide intuition. • Inversely, any informal proof provided in the core of the paper should be complemented by formal proofs provided in appendix or supplemental material. • Theorems and Lemmas that the proof relies upon should be properly referenced. 4. Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main experimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and data are provided or not)? Answer: [Yes] Justification: All hyper-parameters and implementation details are provided in order to reproduce all results. The code is publicly available, data is available upon request according to DIII-D guidelines. Guidelines: • The answer [N/A] means that the paper does not include experiments. • If the paper includes experiments, a [No] answer to this question will not be perceived well by the reviewers: Making the paper reproducible is important, regardless of whether the code and data are provided or not. • If the contribution is a dataset and/or model, the authors should describe the steps taken to make their results reproducible or verifiable. • Depending on the contribution, reproducibility can be accomplished in various ways. For example, if the contribution is a novel architecture, describing the architecture fully might suffice, or if the contribution is a specific model and empirical evaluation, it may be necessary to either make it possible for others to replicate the model with the same dataset, or provide access to the model. In general. releasing code and data is often one good way to accomplish this, but reproducibility can also be provided via detailed instructions for how to replicate the results, access to a hosted model (e.g., in the case of a large language model), releasing of a model checkpoint, or other means that are appropriate to the research performed. • While NeurIPS does not require releasing code, the conference does require all submissions to provide some reasonable avenue for reproducibility, which may depend on the nature of the contribution. For example (a) If the contribution is primarily a new algorithm, the paper should make it clear how to reproduce that algorithm. (b) If the contribution is primarily a new model architecture, the paper should describe the architecture clearly and fully. (c) If the contribution is a new model (e.g., a large language model), then there should either be a way to access this model for reproducing the results or a way to reproduce the model (e.g., with an open-source dataset or instructions for how to construct the dataset). (d) We recognize that reproducibility may be tricky in some cases, in which case authors are welcome to describe the particular way they provide for reproducibility. In the case of closed-source models, it may be that access to the model is limited in some way (e.g., to registered users), but it should be possible for other researchers to have some path to reproducing or verifying the results. 5. Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The code is publicly available in our Github repository https://github.com/PlasmaControl/FusionAIHub. We cannot share the data publicly. Following our description to get access to DIII-D data, it is possible to preprocess the data with our tools and reproduce all results. Guidelines: • The answer [N/A] means that paper does not include experiments requiring code. • Please see the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • While we encourage the release of code and data, we understand that this might not be possible, so [No] is an acceptable answer. Papers cannot be rejected simply for not including code, unless this is central to the contribution (e.g., for a new open-source benchmark). • The instructions should contain the exact command and environment needed to run to reproduce the results. See the NeurIPS code and data submission guidelines (https://neurips.c/public/guides/CodeSubmissionPolicy) for more details. • The authors should provide instructions on data access and preparation, including how to access the raw data, preprocessed data, intermediate data, and generated data, etc. • The authors should provide scripts to reproduce all experimental results for the new proposed method and baselines. If only a subset of experiments are reproducible, they should state which ones are omitted from the script and why. • At submission time, to preserve anonymity, the authors should release anonymized versions (if applicable). • Providing as much information as possible in supplemental material (appended to the paper) is recommended, but including URLs to data and code is permitted. 6. Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyperparameters, how they were chosen, type of optimizer) necessary to understand the results? Answer: [Yes] Justification: We describe all the experiments that lead to the final hyper-parameters. Everything can be reproduced, and we provide all optimizers and processing scripts. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The experimental setting should be presented in the core of the paper to a level of detail that is necessary to appreciate the results and make sense of them. • The full details can be provided either with the code, in appendix, or as supplemental material. 7. Experiment statistical significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [No] Justification: While we provide objective measurements of the model, we did not run uncertainty quanitification. And we were not able to train several billion parameter models in parallel due to limited computational power. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The authors should answer [Yes] if the results are accompanied by error bars, confidence intervals, or statistical significance tests, at least for the experiments that support the main claims of the paper. • The factors of variability that the error bars are capturing should be clearly stated (for example, train/test split, initialization, random drawing of some parameter, or overall run with given experimental conditions). • The method for calculating the error bars should be explained (closed form formula, call to a library function, bootstrap, etc.) • The assumptions made should be given (e.g., Normally distributed errors). • It should be clear whether the error bar is the standard deviation or the standard error of the mean. • It is OK to report 1-sigma error bars, but one should state it. The authors should preferably report a 2-sigma error bar than state that they have a 96% CI, if the hypothesis of Normality of errors is not verified. • For asymmetric distributions, the authors should be careful not to show in tables or figures symmetric error bars that would yield results that are out of range (e.g., negative error rates). • If error bars are reported in tables or plots, the authors should explain in the text how they were calculated and reference the corresponding figures or tables in the text. 8. Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the computer resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We describe which hardware we used and how many resources. We also describe the training and inference time. Guidelines: • The answer [N/A] means that the paper does not include experiments. • The paper should indicate the type of compute workers CPU or GPU, internal cluster, or cloud provider, including relevant memory and storage. • The paper should provide the amount of compute required for each of the individual experimental runs as well as estimate the total compute. • The paper should disclose whether the full research project required more compute than the experiments reported in the paper (e.g., preliminary or failed experiments that didn’t make it into the paper). 9. Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.c/public/EthicsGuidelines? Answer: [Yes] Justification: We do not use any personal data for the studies in this paper. Guidelines: • The answer [N/A] means that the authors have not reviewed the NeurIPS Code of Ethics. • If the authors answer [No] , they should explain the special circumstances that require a deviation from the Code of Ethics. • The authors should make sure to preserve anonymity (e.g., if there is a special consideration due to laws or regulations in their jurisdiction). 10. Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [N/A] Justification: IGNITE does not use any kind of private data. It can potentially be used to generate unexpected diagnostic trajectories. However, since every DIII-D -related publication undergoes an internal review process, physicists will easily recognize if unreasonable outputs were generated with IGNITE. Guidelines: • The answer [N/A] means that there is no societal impact of the work performed. • If the authors answer [N/A] or [No] , they should explain why their work has no societal impact or why the paper does not address societal impact. • Examples of negative societal impacts include potential malicious or unintended uses (e.g., disinformation, generating fake profiles, surveillance), fairness considerations (e.g., deployment of technologies that could make decisions that unfairly impact specific groups), privacy considerations, and security considerations. • The conference expects that many papers will be foundational research and not tied to particular applications, let alone deployments. However, if there is a direct path to any negative applications, the authors should point it out. For example, it is legitimate to point out that an improvement in the quality of generative models could be used to generate Deepfakes for disinformation. On the other hand, it is not needed to point out that a generic algorithm for optimizing neural networks could enable people to train models that generate Deepfakes faster. • The authors should consider possible harms that could arise when the technology is being used as intended and functioning correctly, harms that could arise when the technology is being used as intended but gives incorrect results, and harms following from (intentional or unintentional) misuse of the technology. • If there are negative societal impacts, the authors could also discuss possible mitigation strategies (e.g., gated release of models, providing defenses in addition to attacks, mechanisms for monitoring misuse, mechanisms to monitor how a system learns from feedback over time, improving the efficiency and accessibility of ML). 11. Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pre-trained language models, image generators, or scraped datasets)? Answer: [N/A] Justification: We are not allowed to release the data used to train IGNITE without users signing user agreements with DIII-D . So, there is no safety risk as we will not release the dataset. Model checkpoints will become available upon reasonable request. Guidelines: • The answer [N/A] means that the paper poses no such risks. • Released models that have a high risk for misuse or dual-use should be released with necessary safeguards to allow for controlled use of the model, for example by requiring that users adhere to usage guidelines or restrictions to access the model or implementing safety filters. • Datasets that have been scraped from the Internet could pose safety risks. The authors should describe how they avoided releasing unsafe images. • We recognize that providing effective safeguards is challenging, and many papers do not require this, but we encourage authors to take this into account and make a best faith effort. 12. Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: We are the code owners, and the code is in our Github repository. We used Claude Code for software development. We cite all utilized foundations for model development. We acknowledge the cluster that we used for model training. We acknowledge the DIII-D team for allowing us to collect and use the dataset. Guidelines: • The answer [N/A] means that the paper does not use existing assets. • The authors should cite the original paper that produced the code package or dataset. • The authors should state which version of the asset is used and, if possible, include a URL. • The name of the license (e.g., C-BY 4.0) should be included for each asset. • For scraped data from a particular source (e.g., website), the copyright and terms of service of that source should be provided. • If assets are released, the license, copyright information, and terms of use in the package should be provided. For popular datasets, paperswithcode.com/datasets has curated licenses for some datasets. Their licensing guide can help determine the license of a dataset. • For existing datasets that are re-packaged, both the original license and the license of the derived asset (if it has changed) should be provided. • If this information is not available online, the authors are encouraged to reach out to the asset’s creators. 13. New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [Yes] Justification: We describe the training and validation process of the IGNITE model in details, and the limitations of the model. License is in the FAITH repository. Guidelines: • The answer [N/A] means that the paper does not release new assets. • Researchers should communicate the details of the dataset/code/model as part of their submissions via structured templates. This includes details about training, license, limitations, etc. • The paper should discuss whether and how consent was obtained from people whose asset is used. • At submission time, remember to anonymize your assets (if applicable). You can either create an anonymized URL or include an anonymized zip file. 14. Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? Answer: [N/A] Justification: Our research did not involve crowdsourcing or research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Including this information in the supplemental material is fine, but if the main contribution of the paper involves human subjects, then as much detail as possible should be included in the main paper. • According to the NeurIPS Code of Ethics, workers involved in data collection, curation, or other labor should be paid at least the minimum wage in the country of the data collector. 15. Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals (or an equivalent approval/review based on the requirements of your country or institution) were obtained? Answer: [N/A] Justification: Our research did not involve crowdsourcing or research with human subjects. Guidelines: • The answer [N/A] means that the paper does not involve crowdsourcing nor research with human subjects. • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research. If you obtained IRB approval, you should clearly state this in the paper. • We recognize that the procedures for this may vary significantly between institutions and locations, and we expect authors to adhere to the NeurIPS Code of Ethics and the guidelines for their institution. • For initial submissions, do not include any information that would break anonymity (if applicable), such as the institution conducting the review. 16. Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the core methodology, scientific rigor, or originality of the research, declaration is not required. Answer: [No] Justification: We designed and developed the model based on the Genie paper [5]. Every development step was decided by us. We used Claude Code to assist with the implementation. However, we carefully reviewed the code and double-checked the results. We used unit-tests to guide the software development properly. For the manuscript, we used AI tools to make the main figure more concise, after preparing the initial version. Guidelines: • The answer [N/A] means that the core method development in this research does not involve LLMs as any important, original, or non-standard components. • Please refer to our LLM policy in the NeurIPS handbook for what should or should not be described. Appendix A Utilized diagnostics and actuators Table 1 and Table 2 summarize the diagnostics and actuators that were used for training the IGNITE world model. We grouped the diagnostics as specified in Section 2.1 and resampled the diagnostics in each group. Table 1: Overview of the diagnostics used for this study. Diagnostic Measured properties Channels Sampling freq. Resampled freq. Modality type CER Ion temperature (TiT_i), plasma rotation (vtorv_tor) 2×48$2$×$48$ 100 Hz100\,Hz – Slow time series MSE Magnetic pitch angle 6969 100 Hz100\,Hz – Slow time series TS Electron temperature (TeT_e) and density (nen_e) 2×44$2$×$44$ & 2×10$2$×$10$ 100 Hz100\,Hz – Slow time series Filterscopes Line emissions 88 10 kHz10\,kHz – Fast time series BES Fluctuations 1616 1 MHz1\,MHz 500 kHz500\,kHz Spectrogram ECE Electron temperature 4848 500 kHz500\,kHz – Spectrogram CO2 Line-averaged electron density 44 1.3 MHz1.3\,MHz 500 kHz500\,kHz Spectrogram Magnetics Magnetic field 66 1 MHz1\,MHz 500 kHz500\,kHz Spectrogram Table 2: Overview of the actuators used for this study. Actuator Actuated property Channels Sampling frequency Electron Cyclotron Heating (ECH) Power, toroidal angle, poloidal angle, polarization 88 10 kHz10\,kHz Gas Voltage, gas flow 88 10 kHz10\,kHz Neutral Beam Injection (NBI) Power, torque, voltage 88 10 kHz10\,kHz Resonant Magnetic Perturbation (RMP) Coil current 88 10 kHz10\,kHz I-Coil Coil current 1212 10 kHz10\,kHz Appendix B Details on the modality-specific tokenizers Figure 4 illustrates the processing pipeline for slow time-series. Figure 4: Tokenizer for slow time series. The channel-time array is grouped into four channel groups and patched at 5 samples × channels; a shared linear layer projects each patch by Linear →5×128→ 5×$128$; 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 illustrates the processing pipeline for fast time-series. Figure 5: Tokenizer for fast time series. A 50 ms50\,ms window of 8 channels sampled at 10 kHz10\,kHz gives 500500 samples per channel; a Conv1D front end reduces each channel to 5 frames, patched at 5×85× 8 over frames and channels. A shared Linear →8×256→ 8×$256$ projects each patch, a 256256-wide time encoding is added, and the token sequence is processed by the transformer to produce the model output. Figure 6 illustrates the processing pipeline for spectrograms. Figure 6: Tokenizer for spectrograms. SEach of the C channels contributes a time-frequency map of 98 frames, cut into 16×1616× 16 patches, giving 192192 patches per channel. Patches at the same grid position are concatenated across channels and an MLP reduces them to 192×256$192$×$256$. A frequency encoding (32×256$32$×$256$) and a time encoding (6×2566×$256$) form a 32×632× 6 table whose entries are 256-dimensional vectors, added to the token at that grid position, before the transformer produces the model output.