acceptodds
Under review as a conference paper at ICLR 2027

SafeDivertor: Faithful Divertor Heat Flux Reconstruction from Macroscopic Plasma State Signals via Time-Frequency Prior Exploitation

Abstract

Divertor heat-flux analysis is essential for understanding plasma-wall interactions and protecting plasma-facing components in magnetic-confinement fusion devices, while conventional infrared-based inversion is usually performed after discharge and requires heat-conduction modeling with device-specific material properties, divertor geometry, and boundary conditions. Rather than accelerating this conventional infrared-based inversion paradigm, we introduce a new online-oriented signal-based reconstruction paradigm that directly reconstructs time-resolved radial heat-flux profiles from multi-source macroscopic plasma-state signals available during discharge. To enable systematic study of this task, we construct DivMPS2HF, a multi-source discharge dataset that provides the data foundation and benchmark for signal-based divertor heat-flux reconstruction. We further propose SafeDivertor, a task-driven framework designed to address the key challenges of signal-based heat-flux reconstruction. Specifically, we design physical prior-aware initialization to provide radial-distribution guidance for unavailable target channels, input perturbation to improve robustness against heterogeneous signal reliability, spectral-aware reconstruction optimization to preserve transient heat-flux dynamics, and progressive training to progressively optimize these complementary objectives. Experiments on DivMPS2HF demonstrate that SafeDivertor achieves the best overall performance among the evaluated time-series baselines across all five metrics, establishing a new performance benchmark for signal-based divertor heat-flux reconstruction.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Time Series, Divertor Heat-Flux Reconstruction, Plasma Diagnostics

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sntYOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SafeDivertor, a framework that estimates divertor heat flux for ITER. Taking T1 model as the backbone, SafeDivertor focuses on online operation, compared to offline infrared-based estimations. Four techniques are used to augment the backbone model: physical prior-aware initialization, input perturbation, spectral-aware reconstruction optimization, and progressive training.

Critical strengths
  1. This work demonstrates both improved accuracy (MSE, MAE) and spectral alignment (LSD, LSD-HF) compared to baselines.
  2. The paper is well-written, and the illustrations are intuitive.
Critical weaknesses
  1. Motivation needs better clarification. In the 2nd paragraph of Introduction, the authors mention that previous approaches including PINN and CNN are limited in terms of online reconstruction. However, Figure 1(c) shows neural network workflow is online-oriented. There appears to be a conflict.
  2. The definition of online-oriented is not clear.
  3. More details of baselines are needed to show fair comparisons.
Questions for authors
  1. Why are earlier NN-based reconstruction work not online-oriented?
  2. Why are methods dependent on thermal measurements and heat-conduction modeling not applicable to online-oriented reconstruction?
  3. Could you justify the comparison with baseline is fair? Are the baselines well-trained, under similar condition with SafeDivertor?
  4. What level of latency does "online-oriented" correspond to?
  5. Can existing work, e.g., the CNN work in Related Work section, be easily adapted to support online-oriented reconstruction?
2 comments

Thank you for your valuable feedback, which helps us improve the quality of our work. We highlight the changes in blue in the revised manuscript and address your questions below.

W1, W2, Q1, and Q2. Online-oriented reconstruction and limits of prior workflows

Response to W1, W2, Q1, and Q2

Thank you for pointing out the ambiguity. Our original wording made the limitations of the two cited workflows sound like limitations of the neural network architectures themselves. In the revised Introduction, we describe the requirements of each method explicitly (Section 1, Lines 46–51).

The two workflows face different constraints on reconstruction during discharge. Aymerich et al. [1] train their PINN with fixed initial and boundary conditions in the loss. When those conditions change, the formulation they demonstrate requires retraining with the new conditions, which limits direct reuse of the trained model across discharges. Looby et al. [2] train and evaluate their CNN on ANSYS-simulated thermocouple signals. Applying their workflow on NSTX-U requires dedicated thermocouples embedded beneath the divertor tile surface. The CNN takes a temperature sequence that includes the thermal response after heating ends, so inference must wait until that part of the sequence has been recorded.

The conventional infrared workflow in Figure 1(b) is typically run after the discharge. The EAST workflow [3] first measures the divertor target temperature with an infrared camera and then reconstructs the heat flux by solving a transient inverse heat-conduction problem. The cost of this numerical inversion limits how quickly the workflow can update heat-flux estimates during a discharge (Appendix A.1).

In SafeDivertor, online-oriented reconstruction means updating the heat-flux profile during the discharge as each diagnostic window becomes available. Once the window's signals are aligned, SafeDivertor reconstructs a time-resolved radial heat-flux profile directly from routinely acquired macroscopic plasma-state signals in a single forward pass (Section 1, Figure 1(c), Lines 82–84). So each rolling update can use the existing plasma diagnostics alone (Section 4.1).

W3 and Q3. Baseline training and fair comparison

Response to W3 and Q3

We provide the detailed training protocol in Appendix B.4 and list the key model hyperparameters in Table 7. SafeDivertor and all baselines use the same data split, preprocessing, input window, and batch size. We train every model for 500,000500{,}000 steps and evaluate its final checkpoint on the same test discharges. Figure 4 shows that the baseline training losses generally decrease and stabilize within this shared budget (Appendix B.4).

Q4. What level of latency does "online-oriented" correspond to?

Response to Q4

Once a 500500-step input window is ready, SafeDivertor takes 11.31311.313 ms for one forward pass at batch size 11 (Section 5, Table 6). The current experimental conditions at EAST prevent us from measuring end-to-end latency on the device.

However, this limitation does not diminish the core contribution of SafeDivertor. Our experiments show that macroscopic plasma-state signals suffice to reconstruct time-resolved radial heat-flux profiles directly. All of these input signals are routinely acquired during plasma discharges. SafeDivertor can update the profile during the discharge with a single forward pass as soon as a diagnostic window is aligned (Section 1, Figure 1(c)). The conventional infrared workflow shown in Figure 1(b) reconstructs the heat-flux profile by post-shot heat-conduction inversion, so the profile is available only after the discharge ends.

Q5. Can existing work, e.g., the CNN work in Related Work section, be easily adapted to support online-oriented reconstruction?

Response to Q5

Our online-oriented reconstruction paradigm is independent of the network architecture. A CNN can implement the same mapping from aligned windows of macroscopic plasma-state signals to time-resolved radial heat-flux profiles.

However, adapting the CNN-based method of Looby et al. [2] to our online-oriented reconstruction task requires changes to its inputs and outputs. Their workflow relies on signals from thermocouples embedded beneath the divertor tile surface, which they simulate in ANSYS for training. Their CNN estimates heat-flux profile parameters from a temperature sequence that includes the thermal response after heating ends. Therefore, the model must wait for that part of the sequence before it can produce an estimate for the heating interval.

References

[1] Physics informed neural networks towards the real-time calculation of heat fluxes at W7-X

[2] Convolutional Neural Networks for Heat Flux Model Validation on NSTX-U

[3] The development of a three-dimensional finite element method code for the heat flux analysis of tungsten monoblock divertor on EAST

Dear Reviewer sntY,

The authors have submitted their rebuttal and we are closing in on the end of the discussion period. Please take some time to carefully read over the rebuttal and respond whether your concerns have been addressed.

Best, AC

Reviewer_dARJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SafeDivertor for reconstructing time-resolved radial divertor heat-flux profiles from macroscopic plasma-state signals. It introduces a dataset, DivMPS2HF, a task-driven framework that integrates physical prior-aware initialization, input perturbation, spectral-aware reconstruction optimization. The reported experiments show improvements over the evaluated time-series baselines on five metrics. The task is relevant, and the combination of reconstruction, spectral supervision, and robustness-oriented training is potentially useful.

Critical strengths

The paper formulates a clearly defined reconstruction task in which the target heat-flux channels are unavailable within the input window. The proposed components are motivated by identifiable challenges, including radial-profile guidance, heterogeneous input reliability, and spectral fidelity.

Critical weaknesses

1.In the experimental The dataset is split by discharge shot, which is appropriate, but the stride-one sliding-window evaluation produces highly overlapping samples. The reported number of test windows therefore does not represent an equivalent number of independent observations. Per-shot metrics and uncertainty estimates are needed to substantiate cross-discharge generalization. 2.The reported latency measures only the neural-network forward pass for a 500-step window. It does not establish end-to-end latency, update frequency, or real-time operational feasibility. The authors should clarify the causal availability of the window inputs and distinguish online-oriented inference from validated real-time deployment. 3.Adding progressive training improves MSE and MAE but worsens LSD and LSD-HF relative to the PPI+IP+SRO configuration without progressive training. In addition, Table 3 reports a lower SSIM for multi-scale STFT than for single-scale STFT, despite the accompanying claim that the single-scale loss improves all metrics. These discrepancies should be addressed and the trade-offs discussed explicitly. 4.The heat-flux labels are obtained from post-shot heat-flux analysis. The reported metrics establish agreement with these labels, but do not independently establish physical accuracy beyond them. The dataset is currently available upon reasonable request, and the code is promised after acceptance. More detail on label construction, data access, and reproducible preprocessing would strengthen the work. 5.Important implementation details remain unclear, including the exact multi-scale STFT configurations and weights, the high-frequency band used by LSD-HF, and the training and model-selection protocols for the baselines. These details are necessary to reproduce the results and assess the comparability of the reported experiments.

Questions for authors

Table 3 reports a lower SSIM for multi-scale STFT than for single-scale STFT, despite the accompanying claim that the single-scale loss improves all metrics. These discrepancies should be addressed and the trade-offs discussed explicitly.

2 comments

Thank you for your valuable feedback. We highlight the changes in blue in the revised manuscript and address your questions below.

W1. Per-shot evidence for generalization

Response to W1

Per-shot analysis and uncertainty. To assess generalization across discharges, we compute metrics for each held-out discharge and estimate 95%95\% bootstrap confidence intervals by resampling discharges (Appendix C.2, Table 8 and Figure 6). SafeDivertor achieves the best mean performance across discharges on all five metrics.

Paired comparisons. We compare SafeDivertor with each baseline on the same discharges. The 95%95\% confidence interval for the mean gain lies entirely above zero in 4141 of 4545 comparisons (Appendix C.2, Figure 7, Lines 972–979).

Given the limited dataset size, we evaluate two additional shot-level splits to test whether performance depends on the original split. SafeDivertor performs comparably across all three splits.

SplitMSE ↓MAE ↓SSIM ↑LSD ↓LSD-HF ↓
Original0.2000.2910.8702.4752.729
Split 10.1490.2820.8452.4752.713
Split 20.1530.2620.8792.2762.450

W2. Input availability and end-to-end latency

Response to W2

Input availability. SafeDivertor reconstructs each 500500-step window (0.50.5 s at 11 kHz) once the diagnostic signals are aligned. When the window ends between two samples from a lower-rate diagnostic, linear interpolation requires the later sample (Appendix B.1).

Latency. A forward pass takes 11.31311.313 ms at batch size 11 with the input ready (Table 6). Current EAST conditions preclude measuring end-to-end latency and update frequency.

However, this limitation does not diminish the core contribution of SafeDivertor. Our experiments show that SafeDivertor reconstructs time-resolved radial heat-flux profiles in one forward pass per aligned window using macroscopic plasma-state signals routinely acquired during discharge. The infrared workflow in Figure 1(b) obtains profiles through heat-conduction inversion after the discharge ends.

W3 and Q1. Metric trade-offs in progressive training and STFT

Response to W3 and Q1

Progressive training (PT). PT improves MSE and MAE with slightly higher LSD and LSD-HF (Table 2). Without PT, spectral supervision runs throughout training. PT starts with MSE, adds input perturbations, then uses spectral supervision on clean inputs for the final 100,000100{,}000 steps of the same budget. Fewer spectral updates help explain the trade-off.

Single- and multi-scale STFT. The gains across all five metrics refer to single-scale STFT compared with MSE-only training. Multi-scale STFT further lowers LSD and LSD-HF, with slightly lower SSIM than single-scale STFT (Table 4). The STFT loss targets spectral agreement, while SSIM captures local structure and contrast. Table 2 shows that PPI gives the largest SSIM gain and SRO gives the largest LSD and LSD-HF reductions among the single-component ablations. These gains can diverge.

W4. Label validity and access

Response to W4

Reference labels. Appendix B.1 details preprocessing (Lines 678–715). The labels are derived from infrared target temperatures using the finite element heat-conduction inversion of Yang et al. [1]. They report a maximum relative discrepancy below 4%4\% against ANSYS under constant-temperature and constant-heat-flux boundary conditions. We evaluate agreement with these profiles.

Data and code access. The processed DivMPS2HF dataset is available for academic use upon request to the corresponding author. We will release the code upon acceptance.

W5. STFT settings and training protocol

Response to W5

Training STFT. Two successive average-pooling operations with kernel and stride 22 produce scales of 500500, 250250, and 125125 steps. Each uses a centered STFT with a 6464-point FFT, a 6464-sample Hann window, and a 1616-sample hop. The log-magnitude loss weights are 0.20.2, 0.10.1, and 0.10.1 (Section 4.2, Lines 300–301).

Evaluation STFT. We use a one-sided 6464-point FFT, a 6464-sample Hann window, and 3232 samples of overlap on the original sequences. LSD uses all log-power bins, and LSD-HF uses the frequency bins beyond the 70%70\% cutoff.

Training Protocol. All models share the shot-level split, preprocessing, and input window. We train each for 500,000500{,}000 steps with Adam, batch size 1616, and cosine annealing from 10−410^{-4} to 10−610^{-6}. We save checkpoints every 50,00050{,}000 steps and evaluate the final one on the same test set (Appendix B.4). Table 7 lists key hyperparameters. Figure 4 shows that baseline losses generally stabilize within this budget.

References

[1] The development of a three-dimensional finite element method code for the heat flux analysis of tungsten monoblock divertor on EAST

Dear Reviewer dARJ,

The authors have submitted their rebuttal and we are closing in on the end of the discussion period. Please take some time to carefully read over the rebuttal and respond whether your concerns have been addressed.

Best, AC

Reviewer_sHeHOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SafeDivertor, a framework for reconstructing time-resolved divertor heat-flux profiles directly from macroscopic plasma-state signals. The authors construct the DivMPS2HF dataset and extend an existing time-series imputation backbone with statistical heat-flux prior initialization, input perturbation, multi-scale spectral supervision, and progressive training. Experiments demonstrate improvements over several time-series baselines in numerical accuracy, structural similarity, and spectral consistency, with relatively low inference overhead.

Critical strengths

Meaningful application and task formulation. The paper addresses an important problem in fusion plasma monitoring by investigating direct heat-flux reconstruction from routinely available plasma-state signals, potentially reducing reliance on computationally expensive offline inversion during inference.

Critical weaknesses

a. Limited generalization evidence. Although the dataset contains many overlapping temporal windows, it consists of only 77 independent discharge shots, with 10 reserved for testing. The evaluation does not sufficiently establish robustness across diverse operating conditions or challenging transient heat-load events. This limits the evidence supporting the proposed framework's reliability in practical plasma monitoring.

b. Limited methodological novelty. SafeDivertor primarily combines established techniques, including mean-based target initialization, Gaussian input perturbation, multi-scale STFT supervision, and progressive training, on an existing T1 backbone. While these components improve performance, the paper offers limited new methodological insights beyond their task-specific integration. Moreover, the proposed physical prior is essentially a statistical average rather than a constraint derived from governing physical laws.

Questions for authors
  1. Can the authors provide additional evaluation across different discharge conditions, particularly high heat-flux peaks and rapid transient events? How stable are the results under alternative shot-level train/test splits?

  2. How does the proposed statistical prior compare with simpler alternatives, such as a learned target-channel embedding or directly predicting heat-flux profiles from the observed signals without target placeholders?

1 comment

Thank you for helping us improve this work. All changes are marked in blue in the revised manuscript.

W1 and Q1. How robust is SafeDivertor to alternative shot-level splits, high heat loads, and rapid transients?

Response to W1 and Q1

We agree that independent discharges are scarce. With no more available, we assess generalization using the existing data.

Generalization across shot-level splits. We evaluate SafeDivertor on two additional splits. Performance remains comparable across all three splits.

SplitMSE ↓MAE ↓SSIM ↑LSD ↓LSD-HF ↓
Original0.2000.2910.8702.4752.729
Split 10.1490.2820.8452.4752.713
Split 20.1530.2620.8792.2762.450

Beyond the split comparison, we examine the results for individual discharges in the original test set (Appendix C.2, Table 8). SafeDivertor achieves the best mean performance on all five metrics (Appendix C.2, Figure 6, Lines 969–971), and the paired bootstrap analysis yields 95% confidence intervals entirely above zero for the mean improvement in 41 of 45 baseline–metric comparisons (Appendix C.2, Figure 7, Lines 972–979).

High heat-flux peaks and rapid transients. Appendix C.3 compares two discharges (Figures 8–9, Lines 1012–1022).

Under high heat load in discharge 154956, SafeDivertor closely reproduces the shape and intensity of the main heat-flux band and captures some of its local variations over time. T1 captures the approximate position of the band, with weaker localized peaks. The profiles reconstructed by Sonnet, TimeFilter, and Timer-XL are smoother (Appendix C.3, Figure 8, Lines 1012–1017).

During rapid strike-point motion in discharge 152916, SafeDivertor and T1 track the band's outward shift and preserve its contour. Sonnet's band is nearly stationary. TimeFilter captures some motion, and Timer-XL produces a more diffuse band (Appendix C.3, Figure 9, Lines 1017–1022).

W2. Methodological contribution and physical interpretation of the prior

Response to W2

Methodological novelty. To the best of our knowledge, we are the first to formulate an online-oriented, signal-based reconstruction paradigm that shifts divertor heat-flux analysis from offline infrared-based inversion to direct reconstruction from macroscopic plasma-state signals routinely available during discharge. Our method can also serve as a data-driven virtual sensor for the divertor heat-load state. Its time-resolved radial heat-flux estimates provide a basis for representing divertor heat loads in future fusion digital twins and for monitoring those loads to inform feedback control.

Component selection. Each component is chosen for its role in solving a specific reconstruction problem. PPI supplies radial information for the unobserved heat-flux channels, and IP reduces reliance on a few diagnostics whose reliability varies across discharges. SRO addresses the loss of rapid heat-flux variations under pointwise supervision, and PT stages the training objectives to ease early optimization. The ablations show how these choices complement each other (Table 2).

Physical interpretation of the prior. The Eich function describes the radial heat-flux profile at the outer divertor target by combining exponential decay in the scrape-off layer with Gaussian spreading along the divertor leg [1]. We fit this function to the mean profile used by PPI over the interval spanning the main peak and its radial decay. The fit gives R2=0.982R^2 = 0.982, showing that the prior captures the peak and decay described by the Eich function (Appendix C.1, Eq. (26), Figure 5, Lines 859–873 and 902–907). The full mean profile also retains a secondary peak near 92.5 mm that may be associated with a gap between divertor monoblocks, consistent with observations on EAST [2].

Q2. Comparison with simpler alternatives

Response to Q2

Comparison with simpler alternatives. We compare PPI with zero-filled placeholders, a learned heat-flux channel embedding, and direct prediction without placeholders (Section 5, Lines 409–417 and Table 3).

Initialization MethodMSE ↓MAE ↓SSIM ↑LSD ↓LSD-HF ↓
Zero-filled placeholders0.2510.3220.8522.5112.753
Learned embedding0.2400.3190.8592.5342.798
Direct prediction0.2350.3280.8502.6042.796
PPI0.2000.2910.8702.4752.729

The gains over zero-filled placeholders and direct prediction show the value of adding a heat-flux prior. PPI also outperforms the learned embedding on all five metrics, supporting the effectiveness of the empirical radial prior.

References

[1] Scaling of the tokamak near the scrape-off layer H-mode power width and implications for ITER

[2] Characteristic of the power decay length of small ELMs measured by infrared camera on EAST

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Manuscript Revisions

Dear Area Chair,

Thank you and the reviewers for your feedback. We have added experiments and analyses and clarified the manuscript, with revisions marked in blue. R1, R2, and R3 refer to Reviewers sHeH, dARJ, and sntY, respectively. We appreciate R2 and R3’s feedback that our rebuttal addressed their concerns. R2 retained a rating of 3, and R3 raised their rating from 2 to 3. We would also welcome any further feedback from R1.

Task and methodological novelty (R1 W2). To our knowledge, we are the first to formulate online-oriented reconstruction of time-resolved radial divertor heat flux directly from plasma-state signals acquired during discharge, and we establish the DivMPS2HF benchmark. This shift from post-shot infrared inversion provides a basis for virtual sensing and feedback monitoring. We design four components for specific reconstruction challenges. PPI supplies radial guidance for unobserved heat-flux channels, IP reduces sensitivity to varying diagnostic reliability, SRO preserves rapid heat-flux variations that pointwise losses tend to smooth, and PT stages the objectives to ease joint optimization. Ablations on the same T1 backbone show their contributions (Table 2).

Generalization across discharges (R1 W1 and Q1, R2 W1). We evaluate two additional shot-level splits and obtain comparable performance across all three.

SplitMSE ↓MAE ↓SSIM ↑LSD ↓LSD-HF ↓
Original0.2000.2910.8702.4752.729
Split 10.1490.2820.8452.4752.713
Split 20.1530.2620.8792.2762.450

On the original test split, SafeDivertor achieves the best mean performance across discharges on all five metrics. In 41 of 45 paired baseline–metric comparisons, the 95% bootstrap confidence interval for the mean gain lies entirely above zero (Appendix C.2, Table 8 and Figures 6–7). Figures 8–9 show that SafeDivertor recovers the main heat-flux band under high heat load and tracks its shift during rapid strike-point motion (Appendix C.3).

PPI and its physical interpretation (R1 W2 and Q2). Physical Prior-aware Initialization (PPI) outperforms all three alternatives on all five metrics (Table 3).

Initialization MethodMSE ↓MAE ↓SSIM ↑LSD ↓LSD-HF ↓
Zero-filled placeholders0.2510.3220.8522.5112.753
Learned embedding0.2400.3190.8592.5342.798
Direct prediction0.2350.3280.8502.6042.796
PPI0.2000.2910.8702.4752.729

The gains over zero-filled placeholders and direct prediction show the value of a heat-flux prior. The comparison with a learned embedding supports the effectiveness of the empirical radial prior. The Eich function fits the prior's main peak and radial decay (R2=0.982R^2 = 0.982), and we discuss a possible link between the secondary peak and a gap between divertor monoblocks on EAST (Appendix C.1, Figure 5).

Prior reconstruction methods (R3 W1, Q1, Q2, and Q5). We clarify the requirements of each workflow in the Introduction. Conventional infrared inversion is typically performed after discharge. Aymerich et al.'s PINN requires retraining when initial or boundary conditions change. Looby et al.'s CNN requires post-heating responses from thermocouples embedded in the target. Our signal-based paradigm is independent of the network architecture. Adapting that CNN requires changes to its inputs and outputs.

Online-oriented inference (R2 W2, R3 W2 and Q4). SafeDivertor uses diagnostics routinely acquired during discharge and updates the heat-flux profile in one forward pass per aligned window. If the window ends between two lower-rate samples, linear interpolation requires the later sample. Once the 500-step window is ready, the forward pass takes 11.313 ms at batch size 1 (Table 6). Current EAST conditions preclude on-device measurements of end-to-end latency and update frequency.

Metric trade-offs (R2 W3 and Q1). PT improves MSE and MAE. It applies spectral supervision only in the final stage, which helps explain the slight increases in LSD and LSD-HF (Table 2). Single-scale STFT improves all five metrics over MSE-only training. The multi-scale variant further lowers both spectral errors, with slightly lower SSIM (Table 4). Gains in structural similarity and spectral fidelity can diverge.

Reproducibility (R2 W4 and W5, R3 W3 and Q3). All models share the data split, preprocessing, and 500,000-step training budget. We report the common protocol and loss curves (Appendix B.4, Table 7 and Figure 4), training and evaluation STFT settings (Appendix B.5), and how reference labels are reconstructed by finite-element inversion of target temperatures measured with infrared cameras (Appendix B.1). The processed dataset is available for academic use from the corresponding author. We will release the code upon acceptance.

Thank you for considering our revision.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.