{"id":"f1ca1d07-fd86-4fbe-a7c7-d506e8d1524a","arxiv_id":"2502.03424","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A GNN surrogate of FEA fire simulations is used to train a predictor of the worst-case fire location (MFSP), with evaluation mostly against the surrogate's own labels.","lead":"This paper proposes the concept of the Most Fire-Sensitive Point (MFSP), the single location where a fire would cause the worst structural drift, and a two-stage machine learning framework that predicts it using a graph neural network surrogate of finite element fire simulations. The surrogate is trained on a limited set of fire simulations, then used to generate pseudo labels and gradients for training the MFSP predictor.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Headline MFSP accuracies (74.1% top-5) are measured against pseudo-labels produced by the same MIDR surrogate, not against FEA; only a single building (ex2) is checked against brute-force FEA, so the central claim lacks external ground-truth validation.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the MFSP predictor is trained and scored against pseudo-labels produced by the MIDR surrogate, with only one building validated against brute-force FEA. I agree with this assessment. This concern is load-bearing because every headline MFSP number (top-5 accuracy up to 74.1%, average room distance error 1.15, average room rank 6.73) is computed relative to the surrogate's own ranking. The ex2 case study is a useful existence proof, and the open-source code and dataset are positive features, but they do not replace a distribution-level FEA ground-truth evaluation. I do not see an internal inconsistency in the framework, and the use of pseudo-labels is clearly disclosed; the issue is that the evaluation target does not match the claimed target. The proposed concrete test would settle whether the surrogate-to-FEA gap is small enough for the central claim to hold. Because the reader already assigned a CONDITIONAL verdict reflecting exactly this concern, my stress-test does not change the verdict.","tokens_in":33502,"tokens_out":3245,"duration_ms":33157,"concrete_test":"Run χara brute-force FEA on a random sample of 100 held-out test buildings: for each building, simulate every room-center fire point (or a 2 m grid where affordable), compute the true FEA MIDR for each fire point, derive FEA-based MFSP labels, and evaluate the trained MFSP predictor's distance error, room rank, top-5 accuracy, and MIDR regret (1 - MIDR_at_predicted_MFSP / MIDR_at_true_MFSP) against these FEA labels. If top-5 accuracy falls materially below 74.1% or average regret is non-negligible, the surrogate-validated numbers overstate MFSP identification performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's stated objective is to identify the fire location that maximizes the building's MIDR as computed by FEA. However, the MFSP predictor is trained and evaluated against pseudo ground truth generated by the MIDR predictor itself (Eq. 11 for the training loss, Eq. 23 and the room-rank definition for evaluation). The reported top-5 accuracy, room rank, and distance-error metrics therefore measure agreement between the MFSP network and the surrogate, not agreement with FEA. The only brute-force FEA comparison is the single building ex2 with 1,232 grid points; its near-exact hybrid prediction is encouraging, but it cannot establish population-level accuracy. Moreover, an average Spearman correlation of 0.74 (0.91 for severe cases) over 30 fire scenarios does not by itself guarantee argmax fidelity: the worst-case room can shift under modest ranking errors, especially when several rooms have similar MIDR values. Consequently, the central claim that the framework efficiently identifies the true MFSP is not yet supported by external ground truth, and the headline accuracies may substantially overstate performance against FEA.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Most Fire-Sensitive Point (MFSP) concept—the fire location that maximizes the Maximum Interstory Drift Ratio (MIDR) of a building—and proposes a two-stage GNN framework to identify it. In stage 1, a GNN-based MIDR predictor is trained on FEA (χara) simulations; in stage 2, this predictor acts as a differentiable surrogate that pseudo-labels unlabeled structures and guides the training of an MFSP predictor through MSE and hybrid losses. The authors also propose a Transfer Learning scheme and an Edge Update mechanism. Experiments on a generated dataset of 1,573 labeled and 16,050 unlabeled structures report MIDR Spearman correlations up to 0.742 (over 0.91 for severe cases) and room-level top-5 MFSP accuracy up to 74.06%. The central claim is that the framework efficiently identifies the worst-case fire location, enabling fire safety pre-checks to simulate only that scenario.","tokens_in":33775,"tokens_out":4279,"duration_ms":42027,"significance":"If the reported accuracy were measured against true FEA ground truth, the paper would be a valuable contribution to fire safety assessment in preliminary structural design. The strengths of the manuscript include a clearly defined problem, an open-sourced dataset and code, a nontrivial GNN architecture with an edge-update mechanism, and a three-orders-of-magnitude speedup relative to FEA. The single external FEA case study (ex2) showing a 0.8 m error for the hybrid loss is encouraging. However, the headline MFSP accuracies are computed against pseudo-labels generated by the same MIDR surrogate that also supervises the MFSP predictor, so the central claim lacks population-level external validation. The need for such validation is in fact acknowledged by the authors in Section VIII-B, which states that rigorous experimental validation and cross-comparison studies are essential.","major_comments":[{"comment":"The MFSP predictor is trained and evaluated entirely against pseudo ground truths produced by the MIDR predictor. Equation (11) trains the MFSP network on pseudo-labeled MFSP coordinates, and Eqs. (23)-(24) and Table VII compute distance error, room distance error, and room rank against those same pseudo-labels. Consequently, the reported top-5 accuracy of 74.06% and the average room rank of about 6.7 measure consistency between the MFSP network and the MIDR surrogate, not agreement with FEA. This is a teacher-student loop that does not establish the central claim. Please add an external evaluation on a random sample of test structures in which FEA is used to compute the true MFSP (e.g., by simulating all 30 room-center fire scenarios, or a grid of points), and report distance error, room rank, and top-5 accuracy against those FEA ground truths. Also report the agreement between the MIDR predictor's argmax and the FEA argmax, especially for structures where several rooms have similar MIDR values.","section":"Section VII-B, Eqs. (11), (23), (24), Table VII"},{"comment":"An average Spearman correlation of 0.74, or even 0.91 for severe cases, is not by itself sufficient to certify argmax fidelity. Spearman correlation is computed over 30 fire scenarios per structure and does not quantify how often the surrogate's highest-MIDR point coincides with the FEA highest-MIDR point; a correlation of 0.91 can still misorder the top several positions, and the MFSP is defined by the argmax. Please report the exact room-level MFSP match rate of the MIDR predictor against FEA, and characterize the distribution of MIDR gaps between the top-ranked and second-ranked fire points. Without such information, the statement in Section VII-A3 that 'the high value of average ρs ≥ 0.90 guarantees the reliability and validity' is not supported.","section":"Section VII-A3, Table V, Fig. 9"},{"comment":"The only brute-force FEA validation of the MFSP predictor is the single building ex2, with 1,232 grid points and one reported hybrid-loss error of 0.8 m. A single example cannot establish population-level accuracy, and the text does not report replication or sensitivity to the 2 m grid spacing. Please extend the FEA-based validation to a sample of structures, preferably including severe cases across the test set, and report both MSE-only and hybrid losses for all of them. This would also allow a direct comparison of the two loss functions against true FEA ground truth rather than against pseudo-labels.","section":"Section VII-B2, case study ex2"},{"comment":"The rule-based thermal load generation model, with hand-chosen parameters r_up, r_down, α_i, and β_i, is an ad hoc approximation of fire propagation. Because the FEA labels themselves are generated from this model, the entire pipeline is validated only against this simplified physics. The manuscript acknowledges in Section VIII-B that 'rigorous experimental validation and cross-comparison studies are essential to verify simulation fidelity,' which is an important limitation. The abstract and conclusions should therefore scope the central claim to structures and thermal fields of the type generated by the authors' rule-based model, and the paper should state explicitly that real-world fire validation is absent.","section":"Section VI-D and Section VIII-B"}],"minor_comments":[{"comment":"The word 'identifing' should be 'identifying' in the description of the MFSP predictor.","section":"Section III-B"},{"comment":"'Hybird' appears instead of 'Hybrid' in the table caption and in the discussion of CDF values.","section":"Table VIII and surrounding text"},{"comment":"The acronym 'MSFP predictor' appears in a few places where 'MFSP predictor' is meant; please correct the transposed spelling.","section":"Section IV and Section VII"},{"comment":"The word 'implemente' should be 'implemented' in the sentence describing the two versions of the MIDR predictor.","section":"Section VII-A1"},{"comment":"The caption labels panel (a) as 'CCDF' while the text refers to a CDF; please standardize the terminology.","section":"Figure 9 caption and Section VII-A3"}],"recommendation":"major_revision","confidential_remarks":"The central validation gap is real and load-bearing, but it is addressable within the manuscript's scope: adding FEA-based MFSP ground truths for a sample of the test structures and re-reporting the headline metrics would substantially strengthen the paper. I would not recommend rejection based on the simplified thermal model alone, because the authors disclose it and the framework may still be useful as a proof of concept. However, the current abstract and conclusions overstate the empirical support for the central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here is my honest read. The paper's real contribution is the MFSP concept and the two-stage workflow: a GNN surrogate for FEA-based MIDR, trained on real FEA labels, then a differentiable 'argmaxer' trained to find the worst fire location. That framing is new and practically motivated. The MIDR surrogate itself is reasonably validated against FEA on 1,573 structures with 30 fire scenarios each, with Spearman correlation 0.74 overall and above 0.91 for severe cases. The speed gain is real, and the ablation shows the edge update and transfer learning help. Code and data are open-sourced.\n\nThe soft spot is exactly where the reader put it. The MFSP predictor is trained and evaluated against pseudo-labels generated by the same MIDR surrogate (Eqs. 11 and 23). The reported top-5 accuracy and room rank therefore measure agreement between two neural networks, not agreement with FEA. The single brute-force FEA check on ex2, where the hybrid loss lands within 0.8 m of the true MFSP, is encouraging, but one building cannot support the headline numbers. The rule-based thermal model also depends on hand-chosen parameters (r_up, r_down, the alphas and betas), and there is no sensitivity analysis showing how robust the conclusions are to those choices. The paper's own limitations section acknowledges the need for experimental validation and higher-fidelity thermal modeling, which is honest but does not fix the evaluation gap.\n\nI do not think this is a fatal flaw. The workflow is plausible, the surrogate is well tested, and the concept has real value for preliminary design. But the central claim — that the framework efficiently identifies the true MFSP — is not yet supported by external ground truth. A multi-building FEA validation with error bars, plus a sensitivity analysis on the thermal parameters, would be the natural fix.\n\nThis paper deserves serious peer review. It is new, reproducible, and the authors engage honestly with the literature. I would send it to reviewers, but I would expect the review to demand that extra validation before the headline accuracies are taken at face value.","headline":"Useful applied paper with a genuinely new problem formulation, but the headline MFSP accuracies are measured against the surrogate's own pseudo-labels; the single FEA check is encouraging but not enough.","tokens_in":34308,"tokens_out":1673,"would_cite":false,"duration_ms":17621,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph neural network trained on finite-element fire simulations can pin down the single fire location that would most damage a steel building, and it does so in milliseconds.","keywords":["Most Fire-Sensitive Point","Graph Neural Network","Maximum Interstory Drift Ratio","Fire safety assessment","Finite element analysis surrogate","Transfer learning","Pseudo-labeling","Steel frame structures"],"falsifier":"Run brute-force FEA over all candidate fire points in several held-out buildings (not just the 30 sampled scenarios), take the true MFSP as the FEA argmax, and compare both the MIDR surrogate's top point and the MFSP predictor's output against it; the claim is falsified if room-level Top-5 accuracy against true FEA labels is far below the 74.1% reported against pseudo-labels.","tokens_in":33263,"feed_emoji":"🔥","tokens_out":8307,"duration_ms":72591,"temperature":0.7,"pith_summary":"The paper introduces the Most Fire-Sensitive Point (MFSP) of a building, defined as the fire origin that would cause the largest structural damage as measured by the Maximum Interstory Drift Ratio (MIDR). It claims that a two-network machine-learning pipeline can find this point without brute-force simulation of every possible fire: a graph neural network first learns to imitate finite element analysis (FEA) of fire scenarios, and a second network is trained as an 'argmaxer' over that surrogate to output the worst-case fire point directly. On a large synthetic dataset of steel frames the surrogate's MIDR ranking correlates with FEA at 0.74 on average and above 0.91 for severe fire cases, while the MFSP predictor reaches 74.1% room-level Top-5 accuracy with millisecond inference. If the claim holds, a fire-safety pre-check could simulate only the single worst-case fire scenario instead of enumerating all candidate locations.","feed_headline":"GNN finds a building's worst fire spot in milliseconds","feed_subtitle":"A learned surrogate for finite-element fire analysis hits 74.1% top-5 room accuracy, promising single-scenario pre-checks.","key_machinery":"The machinery is a message-passing graph neural network in which beams and columns are edges and their joints are nodes, together with an edge-update module that refreshes each edge vector from its two endpoint node embeddings at every layer so that the graph itself degrades as the fire develops. Node features encode the extended coordinates of both the structural node and the fire source, their differences and Euclidean distance. For the MFSP task, a randomized virtual fire point occupies the fire-source input slot during training, forcing the network to rely on global structure; the frozen MIDR predictor generates pseudo ground-truth labels at room centers, and the hybrid loss combines the negative surrogate output with an MSE term on those pseudo-coordinates.","core_discovery":"The central claim is that the worst-case fire location for a framed steel building can be predicted as an optimization over a learned surrogate rather than by physical brute force. The MIDR predictor, a message-passing GNN whose edge attributes are updated each layer to reflect fire-induced degradation, is trained on FEA-generated labels and then frozen. The MFSP predictor reuses that GNN as a feature extractor and is trained, using pseudo-labels obtained by evaluating the surrogate at every room center plus a differentiable loss that maximises predicted MIDR, to output the fire point that maximises the surrogate's MIDR. The paper reports room-level Top-5 accuracy up to 74.1%, and argues that the transfer-learning initialization and the edge-update mechanism are each responsible for a measurable share of that performance.","pith_inferences":["The reported Top-5 accuracy compares the MFSP predictor with the MIDR surrogate's pseudo-labels, not with FEA; the evaluation would be strengthened by measuring both networks against brute-force FEA on several buildings, as done once for ex2.","The differentiable-agent-plus-argmaxer recipe transfers to any problem of finding an input that maximizes an expensive simulator output, such as the most contaminant-sensitive point in a ventilation layout.","A direct test of the pseudo-labeling assumption is to compute the Spearman correlation between the surrogate's MIDR rankings and the FEA rankings on the existing labeled set; if it drops sharply for non-severe buildings, the practical usefulness of the pipeline is concentrated in severe cases.","Because the inference-time virtual fire point is the building's geometric center, one could probe robustness by testing buildings whose true MFSP lies at that center; if the predictor systematically errs there, the VFP trick is biasing the learned argmax."],"forward_implications":["Fire-safety pre-checks during preliminary design could run the worst-case fire scenario only, replacing exhaustive enumeration of fire origins with a single simulation.","For buildings where any fire is severe (some fire case exceeds 2% MIDR), the surrogate's ranking correlation exceeds 0.91, so the identified worst-case point is most reliable exactly in the cases that matter.","Because the surrogate computes MIDR for all scenarios in milliseconds per batch versus tens of seconds per scenario for FEA, iterative design exploration becomes interactive.","Since the framework is agnostic to the performance metric, replacing MIDR by another damage indicator extends the same pipeline to other failure modes.","The reported gains of edge-update and transfer learning imply that physics-relevant graph updates and pre-trained structural representations are worth keeping in later versions of the framework."],"supporting_citations":[{"why":"It is the FEA solver that generated the 47,190 fire-case MIDR labels used to train the MIDR predictor.","marker":"[33]"},{"why":"It supplies the structural graph-network framework and the story-matched layer-count strategy adopted to avoid over-smoothing.","marker":"[8]"},{"why":"It provides the temperature-dependent reduction factors for steel material properties used in the FEA simulations.","marker":"[11]"},{"why":"It defines the standard fire-resistance heating curve used in the rapid-heating stage of the rule-based thermal load model.","marker":"[21]"},{"why":"It supplies the growth, flashover, and decay model of fire development that motivates the two-stage temperature evolution.","marker":"[38]"},{"why":"Its Markov-chain fire-spread model is used to align the adopted spread-time parameters with observed flashover behavior.","marker":"[7]"},{"why":"It provides the actor-critic reinforcement learning framing that motivates the hybrid loss using the differentiable surrogate.","marker":"[39]"},{"why":"It supplies empirical upward fire-spread observations supporting the floor-dependent spread-time parameterization.","marker":"[19]"}],"fun_headline_variants":["GNN predicts a building's worst fire spot fast","Worst-fire location found by GNN surrogate","Surrogate GNN finds the most fire-sensitive point","Machine learning speeds up worst-fire scenario finding","GNN pinpoints the fire that threatens a building most"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the MIDR surrogate's ranking of fire locations is faithful to the true finite-element ranking; the reported accuracies measure agreement with the surrogate's own pseudo-labels, and only one building is checked against brute-force FEA.","fun_headline_variants_meta":{"raw":{"variants":["GNN predicts a building's worst fire spot fast","Worst-fire location found by GNN surrogate","Surrogate GNN finds the most fire-sensitive point","Machine learning speeds up worst-fire scenario finding","GNN pinpoints the fire that threatens a building most"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000206,"raw_usage":{"total_tokens":1388,"prompt_tokens":926,"completion_tokens":462,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":542,"completion_tokens_details":{"reasoning_tokens":388}},"tokens_in":542,"tokens_out":462,"duration_ms":4700,"temperature":1.0,"reasoning_tokens":388,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:47:00.046156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run brute-force FEA over all candidate fire points in several held-out buildings (not just the 30 sampled scenarios), take the true MFSP as the FEA argmax, and compare both the MIDR surrogate's top point and the MFSP predictor's output against it; the claim is falsified if room-level Top-5 accuracy against true FEA labels is far below the 74.1% reported against pseudo-labels.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"It is the FEA solver that generated the 47,190 fire-case MIDR labels used to train the MIDR predictor."},{"cited_title":"Jean, Kai-Hung Chang, Yin-Nan Huang, and Chuin-Shan Chen","cited_arxiv_id":null,"evidence_quote":"It supplies the structural graph-network framework and the story-matched layer-count strategy adopted to avoid over-smoothing."},{"cited_title":"EN 1993-1-2:2005 (E), 2005","cited_arxiv_id":null,"evidence_quote":"It provides the temperature-dependent reduction factors for steel material properties used in the FEA simulations."},{"cited_title":"ISO, Geneva, Switzerland, 1999","cited_arxiv_id":null,"evidence_quote":"It defines the standard fire-resistance heating curve used in the rapid-heating stage of the rule-based thermal load model."},{"cited_title":"Centre for Advanced Engineering, University of Canterbury, 2008","cited_arxiv_id":null,"evidence_quote":"It supplies the growth, flashover, and decay model of fire development that motivates the two-stage temperature evolution."},{"cited_title":"Hadjisophocleous","cited_arxiv_id":null,"evidence_quote":"Its Markov-chain fire-spread model is used to align the adopted spread-time parameters with observed flashover behavior."},{"cited_title":"MIT Press, 2018","cited_arxiv_id":null,"evidence_quote":"It provides the actor-critic reinforcement learning framing that motivates the hybrid loss using the differentiable surrogate."},{"cited_title":"Hokugo, Y","cited_arxiv_id":null,"evidence_quote":"It supplies empirical upward fire-spread observations supporting the floor-dependent spread-time parameterization."}],"review_version":1}