acceptodds
Under review as a conference paper at ICLR 2027

Image Fidelity is Not Field Fidelity: Joint Thermodynamic Reconstruction and Error Localization in Neural Tomography

Abstract

Neural fields for scientific tomography are optimized from 2D images, but the actual quantity of interest is often a latent 3D physical field. Because the forward map is many-to-one, low 2D image error need not certify a correct 3D field. Moreover, the latent field is not directly supervised during training, and its error cannot be evaluated against truth at deployment. We develop CoroNeRF to jointly optimize 3D electron density and temperature fields directly from multiview, multiline intensities through a differentiable atomic-emission renderer. Using solar coronal tomography as a controlled testbed, we evaluate physical-field recovery and test whether cross-seed instability provides a ground-truth-free-at-inference indicator of local physical-field error. We underscore the following two observations. (i) Image fidelity is not field fidelity: spectral ablations show that limited-channel reconstructions can fit their available observations well while recovering substantially worse fields, whereas evaluation on a common richer probe exposes the discrepancy. (ii) Cross-seed instability ranks local physical-field error across tested matched-model conditions, supported by sparsification and physical signal-strength controls. Seed-deviation projections provide complementary directional validation, but shared forward-model mismatch can still produce incorrect cross-seed consensus. These results characterize joint thermodynamic recovery and the usefulness and limits of seed-based error localization in a controlled, single-scene solar tomography testbed.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Implicit Neural Representations, Scientific Machine Learning, Neural Tomography, Uncertainty Quantification, Solar Corona, Thermodynamic Reconstruction, Differentiable Rendering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_LoFSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces CoroNeRF, which jointly reconstructs 3D coronal electron density and temperature from multiview, multiline observations through a differentiable atomic-emission model. It demonstrates that low image error can coexist with inaccurate physical fields and evaluates cross-seed variability as an error-localization signal, including its limitations under shared forward-model mismatch.

Critical strengths
  1. The paper addresses an important scientific reconstruction problem by directly evaluating physical-field accuracy and demonstrating the limitations of image-space validation.

  2. The error-localization analysis includes meaningful controls for signal strength, radial structure, and seed stability. The explicit model-mismatch experiments also help establish when ensemble disagreement becomes unreliable.

Critical weaknesses
  1. All experiments use a single synthetic scene that also guided architecture and loss selection. The shared numerical renderer for data generation and inversion further limits evidence that the findings extend beyond this tuned simulation setting.

  2. The main representation comparison supplies ground-truth temperature and reconstructs only density. It therefore does not establish the effectiveness of the proposed joint density–temperature reconstruction relative to reasonable joint-inversion baselines.

Questions for authors
  1. With the architecture and loss fixed, do reconstruction accuracy and error-localization performance persist on an independent coronal scene? This would substantially affect my assessment of the robustness of the findings.

  2. How does CoroNeRF compare with a joint grid-based or MLP baseline under the same four-channel observations, when both density and temperature are unknown?

2 comments
AuthorsOpenReview ↗

Response to Reviewer LoFS

Hi Reviewer LoFS,

Thanks for the two precise requests! We ran both experiments you asked for. Below, we first summarized the 4 main changes in our revision, and then detailed how they relate to each of your two points. The Sections here refer to the latest revised PDF.

Summary of the revision

  1. Held-out scene (Sec. 5.3, App. D.9): the noiseless and η=9\eta = 9 ensembles are repeated on a second ground truth MAS cube (CR2290) with the rest of the pipeline frozen. The inner-band MAE rises by 16-29%, and the correlations and nAUSE change by at most 0.05.

  2. Quadrature mismatch (Sec. 4, App. D.8): We rerendered the observed data at 4x finer step than the inversion (1/1024 vs 1/256). The field MAE rises by around 2%, and every localization metric stays within the matched 95% interval.

  3. Joint baselines (Sec. 5.1, App. D.2): a grid (w/ and w/o Tikhonov regularization) and an MLP now invert density and temperature jointly (we removed the density-only baselines and briefly mentioned it at the end of App. D.2). The inner-band MAEs for the baselines are as follows (nen_{\rm e} / TeT_{\rm e}): hash grid is 0.039 / 0.014 dex, MLP is 0.063 / 0.022, and regularized grid is 0.145 / 0.046. At approximately matched training time, the hash grid is still 1.3x and 2.5x more accurate for density.

  4. Deployable control (Sec. 5.2, App. C.1): we implemented b^E\hat b_E, the inverse emissivity of the reconstructed ensemble mean (as opposed to bEb_E, which is the GT inverse emissivity). The nAUSE is 0.11-0.30 for σens\sigma_{\rm ens} vs 0.48-1.11 for b^E\hat b_E across the matched-model conditions.

Weakness 1 / Question 1, independent scene (item 1 in summary; Tables 21-22). We also summarized the values below in a compact table for you, for density / temperature. CR2283 is the original development scene, and CR2290 is the new scene.

MAE [dex]ρ(σens,ϵmean)\rho(\sigma_{\rm ens}, \epsilon_{\rm mean})nAUSE of σens\sigma_{\rm ens}
noiseless, CR22830.037 / 0.0130.761 / 0.7260.115 / 0.140
noiseless, CR22900.048 / 0.0160.725 / 0.7760.117 / 0.113
η=9\eta = 9, CR22830.056 / 0.0190.572 / 0.5900.274 / 0.251
η=9\eta = 9, CR22900.065 / 0.0220.589 / 0.6150.268 / 0.201

Note: both scenes share the same solar-cycle phase by design, but are separated by about 7 rotations with largely unrelated density structures.

Weakness 1, shared numerical renderer (item 2 in summary; Tables 19-20).

  • MAE rises 2.0 ±\pm 1.2% for density and 1.9 ±\pm 0.4% for temperature over the seed pairs.
  • The correlation goes from 0.761 to 0.770 for density and 0.726 to 0.717 for temperature (nAUSE from 0.115 to 0.120 and from 0.140 to 0.145), all still within matched intervals.
  • The seed-deviation error capture moves the most (7.5% to 4.8% and 5.0% to 4.0%), but it's still 23x and 19x above the isotropic reference.
  • Shared emissivity LUT, so atomic-data error is not tested.

Weakness 2 / Question 2, joint baselines (item 3 in summary; Table 13).

  • For tuning, λsmooth∈{0,10,100}\lambda_{\rm smooth} \in \lbrace 0, 10, 100 \rbrace and L∈{6,10}L \in \lbrace 6, 10 \rbrace. We also bound the MLP's outputs smoothly into the LUT range using an algebraic sigmoid (Equation 30).
  • For runtime, the training time for 6060k steps on one RTX 4090 is 3.8 h / 1.8 h / 0.9 h / 0.6 h for hash grid / MLP / Grid-Reg / Grid, respectively.
  • We also test performance at matched-time: the baselines were still improving at 6060k steps, so we trained the MLP and Grid-Reg (one seed, 120k and 240k steps respectively) for about the hash grid's training time. They reach 0.049 / 0.017 and 0.097 / 0.034 dex, so the hash grid is still 1.3x and 2.5x more accurate for density.

You noted that Question 1 would substantially affect your assessment of the robustness of the findings. We hope that these revisions address it, and we would be very grateful if you could take them into account in your updated assessment. Thanks again!

AuthorsOpenReview ↗

Followup to LoFS

Hi Reviewer LoFS,

Since the discussion period closes tomorrow, we wanted to follow up. You wrote that independent-scene results would substantially affect your assessment on the robustness of the results in your first Question: we reported them in our reply above, in Section 5.3 of the updated manuscript, and also Tables 21-22 in the Appendix. And we also reported joint baseline results (your Question 2) where both density and temperature are unknown, in Section 5.1 and Appendix D.2 of the updated manuscript. We would be grateful to know whether or not these address your concerns, and are happy to clarify any points before the discussion period ends. Thanks again!

Reviewer_dNrbOpenReview ↗
Rating
3
Confidence
1(−1)
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The authors propose CoroNeRF which jointly optimizes 3D electron density and temperature fields from multview intensities. Two main findings are presented: image fidelity is not field fidelity, and cross-seed instability ranks local physical-field error.

Critical strengths
  • The paper addresses an important research problem in neural fields: image fidelity does not necessarily imply accurate underlying physical field.
  • The experiments include useful ablations over spectral coverage, number of views, noise level, etc.
  • The analysis of cross-seed instability is reasonably comprehensive.
Critical weaknesses
  • The study is limited to a single synthetic coronal scene and a matched discretized forward model. The generalizability of conclusions is questionable.
  • Baseline comparisons appear to be relatively narrow. More physics-informed neural inverse methods, Bayesian methods, or other uncertainty quantification approaches might strengthen the analysis.
Questions for authors

Authors could refer to weaknesses for rebuttal.

3 comments
AuthorsOpenReview ↗

Response to Reviewer dNrb

Hi Reviewer dNrb,

Thanks for the positive assessment! We ran a few new experiments that now address both of your concerns. We first summarized the 4 main changes in our revision, and then detailed how they relate to each of your points. The Sections here refer to the latest revised PDF.

Summary of the revision

  1. Held-out scene (Sec. 5.3, App. D.9): the noiseless and η=9\eta = 9 ensembles are repeated on a second ground truth MAS cube (CR2290) with the rest of the pipeline frozen. The inner-band MAE rises by 16-29%, and the correlations and nAUSE change by at most 0.05.

  2. Quadrature mismatch (Sec. 4, App. D.8): We rerendered the observed data at 4x finer step than the inversion (1/1024 vs 1/256). The field MAE rises by around 2%, and every localization metric stays within the matched 95% interval.

  3. Joint baselines (Sec. 5.1, App. D.2): a grid (w/ and w/o Tikhonov regularization) and an MLP now invert density and temperature jointly (we removed the density-only baselines and briefly mentioned it at the end of App. D.2). The inner-band MAEs for the baselines are as follows (nen_{\rm e} / TeT_{\rm e}): hash grid is 0.039 / 0.014 dex, MLP is 0.063 / 0.022, and regularized grid is 0.145 / 0.046. At approximately matched training time, the hash grid is still 1.3x and 2.5x more accurate for density.

  4. Deployable control (Sec. 5.2, App. C.1): we implemented b^E\hat b_E, the inverse emissivity of the reconstructed ensemble mean (as opposed to bEb_E, which is the GT inverse emissivity). The nAUSE is 0.11-0.30 for σens\sigma_{\rm ens} vs 0.48-1.11 for b^E\hat b_E across the matched-model conditions.

Weakness 1, single scene and matched discretization (items 1 and 2 in the summary). We note a few remaining limits: (1) both scenes share a solar-cycle phase; (2) data generation and inversion share the same atomic emissivity tables; (3) errors common to all ensemble members, such as those from the abundance mismatch study (App. D.7), are still invisible to σens\sigma_{\rm ens}.

Weakness 2, narrow baselines (item 3 in the summary). Note: the classical grid inversions (with and without regularization) and the MLP use the same renderer as our model.

Weakness 2, other UQ approaches (item 4 in the summary). Our claim is narrower than a calibrated posterior: σens\sigma_{\rm ens} is a GT-free error-ranking proxy, so we compare it against ranking baselines.

  • We will note that we also tested a linearized, Laplace/Gauss-Newton-type reference: the diagonal of the Fisher information matrix of the forward operator linearized at the true field (an oracle like bEb_E, and not in the paper).
  • For reference, its nAUSE is 0.90-1.12 in the noiseless and sparse-view conditions and 0.55-0.65 under high noise, and never better than σens\sigma_{\rm ens} (0.11-0.30).
  • Adapting a full Bayesian neural-field method to the joint emission renderer is beyond the scope of this revision.

Remark on Presentation. To make the Appendix easier to navigate, we added an appendix table of contents. Hopefully this makes it easier to parse!

Thanks again for your support of the paper!

AuthorsOpenReview ↗

Followup to dNrb

Hi Reviewer dNrb,

Thanks again for your positive assessment! Since the discussion period ends tomorrow, we wanted to followup. We hope the new experiments address the two weaknesses you raised (items 1 and 2 address the single scene and matched forward model, while items 3 and 4 address the baselines and UQ). Let us know if you would like any further clarification on any points before the discussion period ends. Thanks again!

Thank the authors for their rebuttal. The addition of experiments and baselines mostly addressed my previous concerns. I would keep my original rating.

Reviewer_jFLcOpenReview ↗
Rating
2
Confidence
4
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces CoroNeRF for jointly reconstructing three-dimensional coronal electron density and temperature from multiview, multiline intensity images. A multiresolution hash-grid field predicts (\log_{10} n_e) and (\log_{10} T_e), which are mapped to line emissivities through CHIANTI/pyCELP lookup tables and integrated along each ray. Training uses image supervision and an asinh-L1 loss.

Experiments use one static PSI MAS coronal cube. In the noiseless four-line setting, the reported MAEs over (1.1 \le r/R_\odot \le 2.0) are 0.039 dex for density and 0.014 dex for temperature. Spectral ablations show that low held-out image error on a model's own observed channels can coexist with substantially worse fields. A common four-line probe provides a diagnostic of the missing spectral constraints. The noise/view sweep, using noise-matched training loss scales, shows accurate recovery with 100-300 views within this scene and substantial degradation with 5-20 views.

The paper also evaluates the voxelwise standard deviation of ten seed reconstructions as a local error-ranking proxy that requires no ground truth at inference. Its association with absolute ensemble-mean field error is assessed using correlations, sparsification, signal-strength and radial controls, leave-one-seed-out checks, and error projections onto seed-deviation subspaces. A Si-abundance mismatch exposes shared reconstruction errors that ensemble spread can miss. The claims concern empirical ranking within the tested scene and conditions, not probabilistic calibration or generalization across coronal states.

Critical strengths
  1. A physically meaningful joint reconstruction formulation. All spectral channels constrain a shared density-temperature state through a differentiable atomic-emission renderer. The spectral ablation quantifies the benefit of complementary line responses: density MAE falls from 0.530-0.605 dex with the tested single lines to 0.039 dex with all four. This is a useful application of neural fields to recovering interpretable physical quantities. The forward model and experimental settings are documented in detail.

  2. Careful validation of an error-ranking heuristic and its failure modes. The ensemble analysis includes sparsification, an oracle signal-strength control, within-shell correlations, leave-one-seed-out stability, disjoint-seed projection tests, and a longitude-rotation null. These complementary checks strengthen the evidence for local error ranking under matched-model conditions. The Si-abundance experiment explicitly demonstrates incorrect consensus: at Si scale 1.4, density correlation is -0.017 and nAUSE is 0.932, close to random performance. The authors clearly distinguish inference-time outputs from diagnostics that require ground truth.

Critical weaknesses
  1. Independent validation is limited. The evidence is limited along three distinct axes. First, the architecture and objective were selected using physical-field recovery on the same single MAS cube used for evaluation (Section 4, lines 216-223), leaving validation on an independently chosen scene untested. Second, synthetic observation generation and inversion use the same emissivity LUT and numerical quadrature (Appendices A.1-A.2). The Si-scale experiment probes a channel-amplitude mismatch, not a change in rendering discretization (Appendices A.4 and D.7). Table 1 reports a 7.5% relative-L2 image difference and an asinh image error of (1.7 \times 10^{-3}) between production quadrature and the finest reference, which uses one-quarter the step size. The stability of recovered fields and error rankings when finer-generated observations are inverted with production quadrature remains untested. Third, the localization ensembles vary training seeds while keeping one fixed noise realization per noisy condition (Appendix C.1, lines 968-970), so their results remain conditional on that realization. Longitude-slice bootstrap intervals quantify within-scene resampling variability, not variation across scenes or independent noise realizations. The separate noise/view ablation uses different realizations (Appendix D.6), but does not assess localization robustness across realizations. Validation with a fixed pipeline on an independently chosen scene, under a changed rendering discretization, or across independent noise realizations would materially strengthen the contribution.

  2. Comparative evidence for joint reconstruction and deployable error ranking is incomplete. The representation benchmark reconstructs density with ground-truth temperature supplied (Appendix D.2, Table 12), rather than jointly inferring both fields. It therefore does not establish whether the hash-grid representation outperforms the grid or MLP baselines in the paper's main joint reconstruction task. For localization, the comparisons include random ranking and inverse emissivity computed from truth. No informative alternative ranking derived from observations or the reconstructed field and available at inference is tested (Appendices C.1-C.2). Comparing such controls against the same ensemble-mean field-error target would clarify the incremental ranking value of ensemble spread. Ten reconstructions require approximately 30 aggregate A100 GPU-hours of training per condition, whereas computing spread afterward is inexpensive (Appendix D.4, lines 1553-1558). A control computed from the ensemble mean would test ranking value, not training-cost savings; assessing a cheaper single-reconstruction pipeline would additionally require reporting its field accuracy and ranking quality.

Questions for authors
  1. Independent validation: With the architecture, loss, and hyperparameters fixed, do the main reconstruction and localization results persist on an independently selected MAS cube? If that experiment is unavailable, repeating a representative localization condition with independent noise realizations would provide a narrower robustness check. Please report field MAE, Spearman correlation, and nAUSE separately for each scene or realization. These results would directly affect my confidence in the empirical findings.

  2. Discretization sensitivity: When observations are generated with the finer quadrature in Table 1 and inversion retains the production quadrature, how much do field MAE and localization metrics change? The reported image convergence test does not establish this. A representative test would show whether the conclusions depend materially on sharing the same discrete renderer.

  3. Comparative evidence: Can the regularized grid or MLP baseline be extended to predict both density and temperature under the same four-line observations, with appropriate tuning and accuracy/runtime reporting? For localization, how does ensemble spread compare with a simple available-at-inference control, such as inverse emissivity computed from the reconstructed fields? This would assess the performance of the full joint solver and the additional ensemble cost. These are requests for evidence relevant to the present assessment.

2 comments
AuthorsOpenReview ↗

Response to Reviewer jFLc

Hi Reviewer jFLc,

Thanks for the detailed review and the precise requests! We ran the experiments you asked for except for noise realization-varying ensembles. We first summarized the 4 main changes in our revision, and then detailed how they relate to each of your points. The Sections here refer to the latest revised PDF.

Summary of the revision

  1. Held-out scene (Sec. 5.3, App. D.9): the noiseless and η=9\eta = 9 ensembles are repeated on a second ground truth MAS cube (CR2290) with the rest of the pipeline frozen. The inner-band MAE rises by 16-29%, and the correlations and nAUSE change by at most 0.05.

  2. Quadrature mismatch (Sec. 4, App. D.8): We rerendered the observed data at 4x finer step than the inversion (1/1024 vs 1/256). The field MAE rises by around 2%, and every localization metric stays within the matched 95% interval.

  3. Joint baselines (Sec. 5.1, App. D.2): a grid (w/ and w/o Tikhonov regularization) and an MLP now invert density and temperature jointly (we removed the density-only baselines and briefly mentioned it at the end of App. D.2). The inner-band MAEs for the baselines are as follows (nen_{\rm e} / TeT_{\rm e}): hash grid is 0.039 / 0.014 dex, MLP is 0.063 / 0.022, and regularized grid is 0.145 / 0.046. At approximately matched training time, the hash grid is still 1.3x and 2.5x more accurate for density.

  4. Deployable control (Sec. 5.2, App. C.1): we implemented b^E\hat b_E, the inverse emissivity of the reconstructed ensemble mean (as opposed to bEb_E, which is the GT inverse emissivity). The nAUSE is 0.11-0.30 for σens\sigma_{\rm ens} vs 0.48-1.11 for b^E\hat b_E across the matched-model conditions.

Weakness 1 / Question 1, independent scene (item 1 in summary; Tables 21-22). We also summarized the values below in a compact table for you, for density / temperature. CR2283 is the original development scene, and CR2290 is the new scene.

MAE [dex]ρ(σens,ϵmean)\rho(\sigma_{\rm ens}, \epsilon_{\rm mean})nAUSE of σens\sigma_{\rm ens}
noiseless, CR22830.037 / 0.0130.761 / 0.7260.115 / 0.140
noiseless, CR22900.048 / 0.0160.725 / 0.7760.117 / 0.113
η=9\eta = 9, CR22830.056 / 0.0190.572 / 0.5900.274 / 0.251
η=9\eta = 9, CR22900.065 / 0.0220.589 / 0.6150.268 / 0.201

Note: both scenes share the same solar-cycle phase by design, but are separated by about 7 rotations with largely unrelated density structures.

Weakness 1 / Question 2, discretization (item 2 in summary; Tables 19-20).

  • MAE rises 2.0 ±\pm 1.2% for density and 1.9 ±\pm 0.4% for temperature over the seed pairs.
  • The correlation goes from 0.761 to 0.770 for density and 0.726 to 0.717 for temperature (nAUSE from 0.115 to 0.120 and from 0.140 to 0.145), all still within matched intervals.
  • The seed-deviation error capture moves the most (7.5% to 4.8% and 5.0% to 4.0%), but it's still 23x and 19x above the isotropic reference.
  • Shared emissivity LUT, so atomic-data error is not tested.

Weakness 1, noise realizations. This is not tested, and all ensemble members still share exactly one realization. CR2290's noisy ensemble has a new realization, and localization carries over, but the scene changes too. We state this explicitly in App. C.1 and list it as future work in the conclusion.

Weakness 2 / Question 3, joint baselines (item 3 in summary; Table 13).

  • For tuning, λsmooth∈{0,10,100}\lambda_{\rm smooth} \in \lbrace 0, 10, 100 \rbrace and L∈{6,10}L \in \lbrace 6, 10 \rbrace. We also bound the MLP's outputs smoothly into the LUT range using an algebraic sigmoid (Equation 30).
  • For runtime, the training time for 6060k steps on one RTX 4090 is 3.8 h / 1.8 h / 0.9 h / 0.6 h for hash grid / MLP / Grid-Reg / Grid, respectively.
  • We also test performance at matched-time: the baselines were still improving at 6060k steps, so we trained the MLP and Grid-Reg (one seed, 120k and 240k steps respectively) for about the hash grid's training time. They reach 0.049 / 0.017 and 0.097 / 0.034 dex, so the hash grid is still 1.3x and 2.5x more accurate for density.

Weakness 2 / Question 3, inference-time control (item 4 in summary).

  • With 100-300 views b^E\hat b_E ranks like the oracle bEb_E (density nAUSE 0.80 vs 0.81 for noiseless, 0.52 vs 0.52 for η=9\eta = 9). For reference, σens\sigma_{\rm ens} reaches 0.12 and 0.27.
  • Controlling for it still leaves a partial correlation ρ(σens,ϵmean∣b^E)\rho(\sigma_{\rm ens}, \epsilon_{\rm mean} \mid \hat b_E) (for density, it is 0.76 for noiseless and 0.45 for η=9\eta = 9).
  • Under abundance mismatch 1.4x, b^E\hat b_E complements σens\sigma_{\rm ens}.

Weakness 2, cost. One reconstruction takes around 3 A100 GPU-hours vs 30 for K=10K = 10 (App. D.4). We didn't evaluate a single-member ranking pipeline, and thus we don't make any cost-effectiveness claim.

We hope that these revisions address your concerns, and we would be very grateful if you could take them into account in your updated assessment. Thanks again!

AuthorsOpenReview ↗

Followup to jFLc

Hi Reviewer jFLc,

Since the discussion period closes tomorrow, we want to follow up. In your first question, you mentioned that the independent-scene results would directly affect your confidence in the empirical findings: we reported them in our reply above, in Section 5.3 of the updated manuscript, and also Tables 21-22 in the Appendix. We also added a quadrature test (your Question 2) and joint baselines and a deployable b^E\hat{b}_\mathcal{E} control (your Question 3). We would be grateful to know whether or not these address your concerns, and are happy to clarify any points before the discussion period ends. Thanks again!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.