Pith. sign in

REVIEW 4 major objections 6 minor 36 references

BlastOFormer: Attention and Neural Operator Deep Learning Methods for Explosive Blast Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read BlastOFormer is a transformer-based surrogate that maps charge and obstacle layouts directly to full-field maximum blast pressure, outperforming CNN and FNO baselines while running in 6.4 milliseconds per scenario.

desk verdict A legitimate operator-transformer application to blast pressure with a promising SDF encoding, but the headline unscaled metrics are confounded by a model-specific postprocessor and the speedup claim is inflated. read the letter →

arxiv 2505.20454 v1 pith:IXMCXWEK submitted 2025-05-26 cs.LG

Reid Graves , Anthony Zhou , Amir Barati Farimani This is my paper
classification cs.LG
keywords blastpressurepredictiontransformerneuraloperatorsigneddistancefunctionsurrogatemodelcomputationalfluiddynamicsattentionmechanismlearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper introduces BlastOFormer, a transformer-based surrogate model that maps a charge and obstacle configuration directly to the full field of maximum blast pressure on a 99-by-99 grid. The authors train it on 1,500 blastFoam CFD simulations and report that it achieves the best accuracy among a CNN, a Fourier Neural Operator, and itself, with an unscaled R² of 0.9516 and a mean absolute error of 484 kPa. At 6.4 milliseconds per prediction, the model is several orders of magnitude faster than the CFD solver, making it a candidate for real-time blast assessment. The claim that matters is not just speed: the attention-based architecture is asserted to capture spatial coherence, including pressure amplification near obstacles, better than the convolutional and Fourier baselines.

What carries the argument

The load-bearing mechanism is the signed-distance-function (SDF) input encoding combined with a grid-patchified transformer: each obstacle becomes its own SDF channel, the charge is encoded as its mass divided by distance, and a ViT-style linear patch projection turns the 99-by-99 multi-channel grid into tokens; a stack of self-attention blocks with rotary positional embeddings (RoPE) encodes the environment, and a pointwise decoder with random-Fourier-feature and cross-attention queries produces the pressure field, followed by a separately trained 'UnscalerCNN' that maps log-domain predictions to unscaled pressure values.

What would settle it

Run the same training pipeline with the UnscalerCNN applied to the CNN and FNO outputs (or remove it from BlastOFormer) and compare unscaled R² and MAE on the same 200-sample test set; if BlastOFormer is no longer the best, its claimed dominance in the unscaled domain is not established.

Watch

Extended reading notes

Core claim

The central discovery is that a grid-to-grid transformer using signed-distance encodings of obstacles and charge, with cross-attention decoding at query points, can reproduce CFD-computed maximum pressure fields from arbitrary charge masses, charge locations, and obstacle layouts more accurately than a CNN or an FNO, and fast enough for real-time use. The model's unscaled predictions, produced by a small post-hoc convolutional residual network that converts log-domain outputs back to physical pressure values, reach R² = 0.9516 and a 21.1% mean absolute percentage error, while the same model reaches R² = 0.9169 in the log domain. The paper interprets this as evidence that attention over spatial and feature tokens captures long-range shock interactions and obstacle effects that local convolutions and fixed Fourier modes miss.

Load-bearing premise

The comparison between BlastOFormer and the baselines is fair: specifically, that the CNN and FNO are not disadvantaged by lacking an equivalent of the separately trained UnscalerCNN that BlastOFormer uses to produce its unscaled predictions.

Editorial extensions

If this is right

  • If the reported accuracy holds across a broader range of geometries, blast load maps for structural safety assessments can be generated in milliseconds rather than tens of minutes per scenario.
  • The SDF-plus-attention recipe could transfer to other PDE surrogates where the input is a configuration of sources and obstacles and the output is a full field.
  • The gap between log-domain and unscaled performance shows that evaluating surrogates in physical units is necessary; models that look competitive in a transformed space can fail in the physical one.
  • The architecture's injection of explicit obstacle geometry through SDF channels is a cheap way to make a surrogate geometry-aware without meshing.
  • Real-time use in mission planning or parametric studies becomes plausible, since a single forward pass replaces a full CFD re-run.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the UnscalerCNN were applied symmetrically to the CNN and FNO baselines, the unscaled-domain gap might shrink; the paper's claimed superiority in the physical domain depends on this auxiliary network, which only BlastOFormer receives.
  • The dataset's random obstacle placements share a common y-range and non-overlapping x-ranges, so the claimed generalization to 'arbitrary' configurations is only demonstrated within a narrow family; testing on obstacles outside that layout distribution would be a stronger test.
  • The 6.4 ms inference time is for a single sample on one GPU; wall-clock deployments with batching or CPU-only inference would likely be slower, though still far faster than CFD.
  • A natural testable extension is predicting time-dependent pressure fields or 3D fields, both named as future work in the paper.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper introduces BlastOFormer, a Transformer-based operator surrogate for predicting full-field maximum blast pressure from SDF-encoded obstacle and charge configurations. The model is trained on 1500 blastFoam CFD simulations and compared against CNN and FNO baselines on a held-out test set. The authors report that BlastOFormer achieves the highest R2 (0.9516 unscaled) and lowest error metrics, with 6.4 ms inference, and claims more than 600,000x speedup over CFD. The paper also provides qualitative pressure-map and error-map analyses.

Significance. If the empirical claims are validated, BlastOFormer would be a practically useful, fast surrogate for blast pressure field prediction, and the systematic comparison against CNN and FNO baselines on a new 1500-sample blastFoam dataset is a useful contribution. The paper is careful to evaluate on a held-out test set and reports both log-domain and unscaled metrics, which is good practice. However, the central claim of universal outperformance is currently undermined by an asymmetric evaluation setup (the UnscalerCNN postprocessor) and by internally inconsistent quantitative claims. With corrections, the work could be a solid empirical benchmark.

major comments (4)
  1. [§2.2.5 and Table 1] The unscaled metrics for BlastOFormer are produced by passing the log-domain prediction through a separately trained UnscalerCNN, while no equivalent postprocessor is described for the CNN or FNO baselines. Consequently, the unscaled comparison in Table 1 is not a comparison of the three architectures alone, and the claimed superiority in the unscaled domain (abstract and §3.1) is not established. The authors should either apply the same post-processing to all models, compare all models end-to-end, or report BlastOFormer's unscaled metrics without the UnscalerCNN and revise the claims accordingly.
  2. [Table 1, §3.1] The log-domain results contradict the abstract's statement that BlastOFormer 'outperforms ... across both log transformed and unscaled domains': Table 1 reports CNN log R2 = 0.9218 versus BlastOFormer log R2 = 0.9169. The paper acknowledges this in §3.1 but the abstract and conclusion still claim universal outperformance. The claims must be revised to be consistent with the presented numbers.
  3. [Abstract and §3.1] The abstract states that BlastOFormer is 'more than 600,000 times faster than CFD simulations,' while §3.1 reports blastFoam takes 10–12 minutes per sample and BlastOFormer 6.4 ms, which corresponds to a speedup of roughly 100,000–120,000x (about five orders of magnitude), not 600,000x. The text itself says 'exceeding 5 orders of magnitude,' so the abstract's factor appears to be an error and should be corrected.
  4. [§2.1] The dataset description states a timestep of 1×10⁻⁷ s and a total duration of 7.5×10⁻⁶ s, which implies 75 timesteps per simulation, not 'approximately 450 total timesteps.' Please correct the number or explain the discrepancy; as written, the simulation setup is internally inconsistent and hinders reproducibility.
minor comments (6)
  1. [§2.2.5] The subsection title is 'Scaling Convolutional Neural Network' but the text and later references call it 'UnscalerCNN'; please unify the terminology.
  2. [Table 1] The model name in the first column is rendered as 'BlastOF ormer' due to a spacing/typing error; it should read 'BlastOFormer'.
  3. [Figure 2] Figure 2 is not cited in the text; please add an explicit citation where it is first discussed.
  4. [§2.2.4] The Random Fourier Feature bandwidth parameter σ is a free hyperparameter, but its value is not reported; please state the value used in the experiments.
  5. [§2.4.2] The FNO 'Conditioning channels: 21' is not explained; clarify how the 21 conditioning channels are constructed from obstacle and charge configurations.
  6. [Abstract] There is a typo in the abstract: 'more 600,000 times faster' should be 'more than 600,000 times faster' (or the corrected speedup factor).

Circularity Check

1 steps flagged · score 2.0 of 10

The headline unscaled metrics are produced by a separately trained UnscalerCNN postprocessor rather than by the BlastOFormer operator itself; this makes that portion of the comparison partially fitted, while the log-domain comparison and held-out evaluation keep the central claim substantially independent.

  1. fitted input called prediction [Section 2.2.5; Table 1; Abstract]
    "To improve fidelity in the unscaled pressure domain, we introduce an auxiliary convolutional neural network that learns a residual mapping from log transformed model predictions to their corresponding unscaled values. The network is trained post hoc using log domain predictions from the trained BlastOFormer model. Specifically, predicted log pressure maps are passed through the UnscalerCNN, which is trained to minimize the mean squared error with respect to the unscaled blastFoam ground truth data."

    The abstract and Table 1 attribute the best unscaled R2 (0.9516) and lowest unscaled MAE to 'BlastOFormer'. By the paper's own description, those numbers are not produced by the proposed attention-based operator alone; they are outputs of a six-layer UnscalerCNN trained post hoc to regress unscaled blastFoam ground truth from BlastOFormer's log-domain predictions. The CNN and FNO baselines receive no equivalent postprocessor. Thus the unscaled superiority claim reduces partly to a fitted auxiliary regressor rather than to the operator architecture itself. The log-domain metrics are independent, and the test split is held out, so the circularity is partial rather than total; this is primarily a benchmark-asymmetry problem.

full rationale

No derivation in the paper is self-definitional: the SDF/charge encoding is an input representation, not a target quantity; the loss is L1 on maximum pressure fields; and the test set is separate from training. The only load-bearing step approaching circularity is the UnscalerCNN. It is a learned postprocessor fit to ground truth via BlastOFormer's log outputs, so the unscaled scores reported as BlastOFormer are partly scores of that auxiliary network; because baselines lack it, the unscaled comparison is not symmetric. This does not invalidate the log-domain analysis, which is independent, though Table 1 shows CNN has higher log R2 (0.9218 vs 0.9169) while BlastOFormer has lower MAE and MAPE. Self-citations to OFormer, ViT, and prior PDE-surrogate work are architectural lineage and are not used as load-bearing evidence, so they do not raise the score. Overall: minor partial circularity in the unscaled headline metric; the central method still has independent held-out content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on trusting blastFoam as ground truth, assuming the encoded parameters fully determine the pressure field, and assuming the narrow training distribution represents the claimed 'arbitrary' configurations. The model itself contributes fitted weights and hand-chosen hyperparameters, and the UnscalerCNN contributes an additional fitted component used for the headline unscaled metrics.

free parameters (4)
  • UnscalerCNN weights = not reported
    A separately trained CNN maps log-domain BlastOFormer predictions to unscaled pressures. The headline unscaled R2 and MAE are computed from its output, and the baselines are not described as having an equivalent network.
  • Transformer hyperparameters = patch size 1, encoder dim 96/256, 6 layers, 4 heads, learning rate 1e-4, batch size 4
    Chosen by hand; the paper reports examining patch sizes 1, 3, 9, 11 but uses only patch size 1 in the reported comparisons.
  • Random Fourier Feature bandwidth sigma = not specified
    The decoder uses RFF with B sampled from N(0, sigma^2), but the value of sigma is never stated, leaving an uncontrolled tuning parameter.
  • SDF outside-obstacle constant = -1
    The signed distance map returns -1 for all points inside obstacles or at the charge location, rather than a true signed distance; this hand-chosen value shapes the input representation.
assumptions (4)
  • domain assumption blastFoam CFD outputs are accurate ground truth for blast pressure fields.
    All 1,500 training samples and all error metrics treat blastFoam simulations as the reference; no experimental or analytical validation is included.
  • domain assumption The pressure field is fully determined by charge mass, charge location, and the three box obstacles as encoded.
    The dataset only varies these parameters and the model is trained to map the SDF inputs directly to pressure; other physical effects (ground, material, turbulence) are not represented.
  • domain assumption The 99 by 99 probe grid at z = 1 m captures the maximum pressure field of interest.
    Predictions and ground truth are only evaluated on this grid; no resolution or height sensitivity study is performed.
  • domain assumption The narrow sampling distribution is representative of 'arbitrary obstacle and charge configurations'.
    Charge y is restricted to [-4.9, 2.0], obstacles have narrow x ranges and a shared y range; the paper claims arbitrary configurations but provides no out-of-distribution evaluation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of BlastOFormer: Attention and Neural Operator Deep Learning Methods for Explosive Blast Prediction." pith.science (2026). https://pith.science/paper/IXMCXWEK

@misc{pith2026250520454,
  author       = {Pith},
  title        = {Pith review of: BlastOFormer: Attention and Neural Operator Deep Learning Methods for Explosive Blast Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IXMCXWEK}},
  note         = {Machine review of arXiv:2505.20454}
}
read the original abstract

Accurate prediction of blast pressure fields is essential for applications in structural safety, defense planning, and hazard mitigation. Traditional methods such as empirical models and computational fluid dynamics (CFD) simulations offer limited trade offs between speed and accuracy; empirical models fail to capture complex interactions in cluttered environments, while CFD simulations are computationally expensive and time consuming. In this work, we introduce BlastOFormer, a novel Transformer based surrogate model for full field maximum pressure prediction from arbitrary obstacle and charge configurations. BlastOFormer leverages a signed distance function (SDF) encoding and a grid to grid attention based architecture inspired by OFormer and Vision Transformer (ViT) frameworks. Trained on a dataset generated using the open source blastFoam CFD solver, our model outperforms convolutional neural networks (CNNs) and Fourier Neural Operators (FNOs) across both log transformed and unscaled domains. Quantitatively, BlastOFormer achieves the highest R2 score (0.9516) and lowest error metrics, while requiring only 6.4 milliseconds for inference, more than 600,000 times faster than CFD simulations. Qualitative visualizations and error analyses further confirm BlastOFormer's superior spatial coherence and generalization capabilities. These results highlight its potential as a real time alternative to conventional CFD approaches for blast pressure estimation in complex environments.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 29 canonical work pages

  1. [1]

    & Zinno, R

    Filice, A., Mynarz, M. & Zinno, R. Experimental and empirical study for predic- tion of blast loads. Applied Sciences 12 (2022). URL https://www.mdpi.com/ 2076-3417/12/5/2691

    work page 2022

  2. [2]

    Kang, M. A. & Park, C. H. Prediction of peak pressure by blast wave propagation between buildings using a conditional 3d convolutional neural network. IEEE Access 11, 26114–26124 (2023). 18

    work page 2023

  3. [3]

    A., Pannell, J

    Dennis, A. A., Pannell, J. J., Smyl, D. J. & Rigby, S. E. Prediction of blast loading in an internal environment using artificial neural networks. International Journal of Protective Structures 12, 287–314 (2021). URL https://doi.org/10. 1177/2041419620970570

    work page 2021

  4. [4]

    & Hao, H

    Li, Q., Wang, Y., Chen, W., Li, L. & Hao, H. Machine learning prediction of bleve loading with graph neural networks. Reliability Engineering & System Safety 241, 109639 (2024). URL https://www.sciencedirect.com/science/article/ pii/S0951832023005537

    work page 2024

  5. [5]

    Remennikov, A. M. & Rose, T. A. Modelling blast loads on buildings in complex city geometries. Computers & Structures 83, 2197–2205 (2005). URL https: //www.sciencedirect.com/science/article/pii/S0045794905002002

    work page 2005

  6. [6]

    Kingery, C. N. & Bulmash, G. Air blast parameters from tnt spherical air burst and hemispherical surface burst. Technical Report, U.S. Army Ballistic Research Laboratory, Aberdeen Proving Ground, MD (1984)

    work page 1984

  7. [7]

    A review of methods for predicting bomb blast effects on buildings

    Remennikov, A. A review of methods for predicting bomb blast effects on buildings. Faculty of Engineering - Papers 6 (2003)

    work page 2003

  8. [8]

    Modelling of blast loading on aboveground structures—i

    Beshara, F. Modelling of blast loading on aboveground structures—i. general phenomenology and external blast. Computers & Structures 51, 585–596 (1994). URL https://www.sciencedirect.com/science/article/pii/0045794994900663

    work page arXiv 1994

Show all 36 references
  1. [9]

    A novel and efficient computational peridynamic framework for modeling explosive blast effects on solid plates

    Abdoh, D. A novel and efficient computational peridynamic framework for modeling explosive blast effects on solid plates. Computers & Structures 299, 107381 (2024). URL https://www.sciencedirect.com/science/article/pii/ S004579492400110X

    2024

  2. [10]

    A., Genidi, M

    Talaat, M., Yehia, E., Mazek, S. A., Genidi, M. M. & Sherif, A. G. Finite element analysis of rc buildings subjected to blast loading. Ain Shams Engineering Jour- nal 13, 101689 (2022). URL https://www.sciencedirect.com/science/article/pii/ S2090447921004676

    2022

  3. [11]

    & Rigby, S

    Dib, N., Z´ ehil, G.-P. & Rigby, S. On the blast-wave shielding effect of porous buildings. Journal of Fluids and Structures 115, 103787 (2022). URL https: //www.sciencedirect.com/science/article/pii/S0889974622001797

    2022

  4. [12]

    Dennis, A. A. & Rigby, S. E. The direction-encoded neural network: A machine learning approach to rapidly predict blast loading in obstructed environments. International Journal of Protective Structures 15, 455–483 (2024). URL https: //doi.org/10.1177/20414196231177364

    2024 doi

  5. [13]

    & Hao, H

    Li, Q., Wang, Y., Shao, Y., Li, L. & Hao, H. A comparative study on the most effective machine learning model for blast loading prediction: From gbdt to transformer. Engineering Structures 276, 115310 (2023). URL https://www. 19 sciencedirect.com/science/article/pii/S0141029622013864

    2023

  6. [14]

    & Sun, J

    He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition (2015). URL https://arxiv.org/abs/1512.03385. 1512.03385

    2015 arXiv

  7. [15]

    Vaswani, A. et al. Attention is all you need (2023). URL https://arxiv.org/abs/ 1706.03762. 1706.03762

    2023 arXiv

  8. [16]

    & Barati Farimani, A

    Li, Z., Shu, D. & Barati Farimani, A. Scalable transformer for pde surrogate modeling. Advances in Neural Information Processing Systems 36 (2023)

    2023

  9. [17]

    & Barati Farimani, A

    Shu, D., Li, Z. & Barati Farimani, A. A physics-informed diffusion model for high-fidelity flow field reconstruction. Journal of Computational Physics 478, 111972 (2023). URL https://www.sciencedirect.com/science/article/pii/ S0021999123000670

    2023

  10. [18]

    & Farimani, A

    Li, Z., Meidani, K. & Farimani, A. B. Transformer for partial differential equations’ operator learning. Transactions on Machine Learning Research(2023). URL https://openreview.net/forum?id=EPPqt3uERT

    2023

  11. [19]

    & Barati Farimani, A

    Lorsung, C. & Barati Farimani, A. Mesh deep q network: A deep reinforcement learning framework for improving meshes in computational fluid dynamics. AIP Advances 13, 015026 (2023). URL https://doi.org/10.1063/5.0138039

    2023 doi

  12. [20]

    & Barati Farimani, A

    Zhou, A. & Barati Farimani, A. Predicting change, not states: An alternate frame- work for neural pde surrogates. Computer Methods in Applied Mechanics and Engineering 441, 117990 (2025). URL https://www.sciencedirect.com/science/ article/pii/S0045782525002622

    2025

  13. [21]

    Dosovitskiy, A. et al. An image is worth 16x16 words: Transformers for image recognition at scale. ICLR (2021)

    2021

  14. [22]

    Kovachki, N. et al. Neural operator: learning maps between function spaces with applications to pdes. J. Mach. Learn. Res. 24 (2023)

    2023

  15. [23]

    URL https://github.com/synthetik-technologies/ blastfoam

    blastFoam: A solver for compressible multi-fluid flow with application to high-explosive detonation. URL https://github.com/synthetik-technologies/ blastfoam

  16. [24]

    & Brewer, T

    Heylmun, J., Vonk, P. & Brewer, T. blastFoam version 6.0 User Guide (2022). URL https://github.com/synthetik-technologies/blastfoam

    2022

  17. [25]

    OpenFOAM v12 User Guide (The OpenFOAM Foundation, London, UK, 2024)

    Greenshields, C. OpenFOAM v12 User Guide (The OpenFOAM Foundation, London, UK, 2024). URL https://doc.cfd.direct/openfoam/user-guide-v12

    2024

  18. [26]

    Su, J. et al. Roformer: Enhanced transformer with rotary position embed- ding. Neurocomputing 568, 127063 (2024). URL https://www.sciencedirect.com/ science/article/pii/S0925231223011864. 20

    2024

  19. [27]

    & Fedkiw, R

    Osher, S. & Fedkiw, R. Signed Distance Functions , 17–22 (Springer New York, New York, NY, 2003). URL https://doi.org/10.1007/0-387-22746-6 2

    2003 doi

  20. [28]

    & Zhu, W

    Chan, T. & Zhu, W. Level set based shape prior segmentation , Vol. 2, 1164–1170 vol. 2 (2005)

    2005

  21. [29]

    L., Kiros, J

    Ba, J. L., Kiros, J. R. & Hinton, G. E. Layer normalization (2016). URL https: //arxiv.org/abs/1607.06450. 1607.06450

    2016 arXiv

  22. [30]

    & May, J

    Gheini, M., Ren, X. & May, J. Cross-attention is all you need: Adapting pre- trained transformers for machine translation (2021). URL https://arxiv.org/abs/ 2104.08771. 2104.08771

    2021 arXiv

  23. [31]

    & Chen, X

    Hou, R., Chang, H., Ma, B., Shan, S. & Chen, X. Cross attention network for few-shot classification (Curran Associates Inc., Red Hook, NY, USA, 2019)

    2019

  24. [32]

    R., Fan, Q

    Chen, C.-F. R., Fan, Q. & Panda, R. CrossViT: Cross-Attention Multi-Scale Vision Transformer for Image Classification , 347–356 (2021)

    2021

  25. [33]

    & Recht, B

    Rahimi, A. & Recht, B. Random Features for Large-Scale Kernel Machines , Vol. 20 (Curran Associates, Inc., 2007). URL https://proceedings.neurips.cc/ paper files/paper/2007/file/013a006f03dbc5392effeb8f18fda755-Paper.pdf

    2007

  26. [34]

    & Ong, C

    Tran, A., Mathews, A., Xie, L. & Ong, C. S. Factorized fourier neural operators (2023). URL https://arxiv.org/abs/2111.13802. 2111.13802

    2023 arXiv

  27. [35]

    Tancik, M. et al. Fourier features let networks learn high frequency functions in low dimensional domains (2020). URL https://arxiv.org/abs/2006.10739. 2006. 10739

    2020 arXiv

  28. [36]

    Lee, J. et al. Set transformer: A framework for attention-based permutation- invariant neural networks (2019). URL https://arxiv.org/abs/1810.00825. 1810. 00825. 21 5 Appendix 5.1 Additional blastFoam Configuration phase parameter value c4 Reactants equation of state Murnaghan...

    2019 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.