{"id":"801ce1a6-5914-4f87-a694-d5a924ad32e2","arxiv_id":"2501.18189","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A leaky integrate-and-fire spiking network matches or beats recurrent convolutional networks on simulated crack growth and Turing pattern prediction while using far fewer parameters.","lead":"Brain-inspired spiking neural networks predict simulated fatigue crack growth and Turing pattern formation with accuracy matching or beating conventional neural networks while using far fewer parameters. The result points toward low-power, on-device prediction for structural health monitoring and materials processing.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 27.71% STCLIF error reduction rests on 5 effectively independent Turing test trajectories due to overlapping sliding windows; with no error bars or seed repeats, the headline improvement is statistically ungrounded.","rationale":"The central claim of the paper is comparative: STCLIF outperforms state-of-the-art ANN models on microstructure evolution benchmarks at lower cost. The quantitative foundation is the 27.71% error reduction on Turing patterns. The evaluation protocol undermines the statistical basis of that comparison: with only 5 independent test trajectories and sliding windows producing about 48 overlapping samples per trajectory, the effectively independent sample size is 5, not 240. Any measured difference between models on such a small sample can be due to sampling variability, especially because no error bars or multiple-seed results are provided. The paper's supporting evidence, such as weight distributions and connectivity density in Fig. 5, is mechanistically plausible but does not establish that the accuracy advantage is real; it explains a possible mechanism only if the advantage exists. The FCG results are less vulnerable because 108 test trajectories are available, but the Turing result is the one cited in the abstract and Section 3.3. Therefore the appropriate assessment remains conditional: the claim is acceptable only if the statistical robustness check confirms the effect; otherwise the headline generalization should be scaled back. The reader's weakest assumption identifies the same issue, so agreement is 'agree'.","tokens_in":11655,"tokens_out":3952,"duration_ms":40104,"concrete_test":"Re-evaluate the Turing benchmark using only one separated window per test trajectory (e.g., the final 10-frame output from each of the 5 test runs, or windows separated by at least 20 frames). Compute per-trajectory MAE for STCLIF, ConvLSTM, PredRNN++, and SimVP, and run a paired permutation or bootstrap test across the 5 trajectories; also retrain each model with at least 5 random seeds. If STCLIF does not beat the best ANN in at least 4 of 5 trajectories, or the paired test is not significant at p<0.05, the 27.71% reduction is a small-sample artifact and the headline claim should be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's flagship quantitative result is the 27.71% error reduction for STCLIF over ANN models on Turing pattern prediction (Section 3.3, Fig. 4c). The test set behind this number is described in Supplementary Note 1, 'Input-output configurations': 5 independent trajectories are used for testing, and each is expanded by a sliding window of length 20 with step 1. Because adjacent windows share 19 of 20 frames, the reported 240 test predictions are not independent; they are highly autocorrelated samples from only 5 initial conditions. No standard deviations, confidence intervals, per-trajectory errors, or multiple-seed repeats are reported, so the measured advantage could be dominated by one or two trajectories. The same overlap issue affects the FCG benchmark, though less severely there because the test set contains 108 independent trajectories. The architecture analysis (reduced weight variance, sparser connectivity) is mechanistically suggestive but does not by itself establish that the accuracy difference is real. Thus the central claim of superior predictive accuracy is not statistically supported as stated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"Zhao and Xu compare artificial and spiking neural networks for predicting microstructure evolution in synthetic digital libraries of fatigue crack growth (FCG) and Turing patterns. They evaluate base RNN/LSTM/SNN models on FCG and advanced spatiotemporal models (ConvLSTM, PredRNN++, SimVP, STCLIF) on Turing patterns, using MAE as the accuracy metric and parameter count/memory estimates as cost metrics. The paper reports that the spiking STCLIF model yields a 27.71% lower MAE than the ANN alternatives on Turing-pattern prediction while using an order of magnitude fewer parameters, and interprets this advantage via reduced weight variance and connectivity density. The authors argue that SNNs can overcome the usual accuracy-cost tradeoff for these microstructure evolution tasks.","tokens_in":11844,"tokens_out":6180,"duration_ms":58106,"significance":"If the central claim is established, the work is a valuable demonstration that event-driven models can outperform recurrent convolutional architectures on spatiotemporal microstructure benchmarks at lower memory cost, with implications for on-device structural health monitoring and materials processing. The paper contributes a clearly specified benchmark protocol (digital libraries from XFEM and Gray-Scott simulations) and compares a wider set of architectures than is typical in the domain. However, the claim of superior accuracy currently rests on a small effective test set with no uncertainty quantification, so the significance is conditional.","major_comments":[{"comment":"The headline result that STCLIF achieves a 27.71% error reduction over ANN models for Turing-pattern prediction is not statistically supported. The Turing test set is generated by expanding 5 independent trajectories with a sliding window of length 20 and step 1, producing 48 windows per trajectory (240 total) whose successive windows share 19 of 20 input frames. The effective number of independent test cases is therefore about 5, far smaller than the nominal 240. No per-trajectory errors, confidence intervals, or multiple-seed runs are reported, so the reported reduction could be dominated by one or two trajectories. Please report per-trajectory errors and a significance test that respects trajectory boundaries (e.g., a paired bootstrap over the 5 trajectories, or a non-overlapping window evaluation), and repeat training with multiple random seeds. The same overlapping-window issue affects the FCG test set (Supplementary Note 1) and should be addressed there as well.","section":"Supplementary Note 1; Sec. 3.3; Fig. 4c"},{"comment":"The abstract and conclusions claim that STCLIF's accuracy benefit arises from reduced weight range and sparser connections, but Section 3.4 provides only a correlational comparison of weight statistics for a single trained pair (ConvLSTM vs. STCLIF). These models differ in many architectural respects (spiking dynamics, autaptic connections, layer counts, etc.), so the observed differences in weight variance and connectivity density do not establish causation. A controlled experiment, e.g., imposing the same weight regularization or connectivity constraints on the ANN baseline, or ablating components of STCLIF, is needed to support the claimed mechanism. As written, the explanation is a plausible post-hoc interpretation rather than a demonstrated cause.","section":"Sec. 3.4; Fig. 5; Abstract"}],"minor_comments":[{"comment":"There are several typos and unclear notations: 'to to enable' in Section 2.2.2, 'machine larning' in Supplementary Note 1, and 'p(r, t) or p(r, t)' in Section 3.1 where a symbol appears twice. Please proofread the manuscript carefully.","section":"Sec. 2.2.2; Supplementary Note 1; Sec. 3.1"},{"comment":"For the Turing-pattern data, a time series of length 68 with a sliding window of length 20 and step 1 yields 49 possible windows per trajectory, not 48 as stated. Please clarify the extraction rule or correct the counts.","section":"Supplementary Note 1"},{"comment":"The axes of Figure 4 are described only as 'observation duration' and 'time step' in the text; please define them in the captions and include absolute MAE values and error bars where available.","section":"Fig. 4"},{"comment":"The estimate that 'spiking neurons save about 3/4 of memory compared to artificial neurons in LSTM' is based only on a parameter-count comparison of the temporal cells; actual memory usage depends on implementation details such as precision, batching, and whether inference uses hardware-specific spike optimizations. Please add a caveat.","section":"Sec. 3.1"},{"comment":"The crack-thickness numbers ('8 for base ANN/LSTM and 1 for base SNN') should be defined: how is thickness measured in pixels, and why is a smaller thickness considered better (presumably because it indicates a sharper interface)?","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper addresses an interesting and timely question, and the benchmark setup is useful. The central quantitative claim, however, is not yet statistically grounded because of the small effective test set and the absence of uncertainty quantification. I would not reject the manuscript; the issues are fixable with additional experiments and re-analysis. I would also encourage the authors to make code and data available, as this would strengthen reproducibility, but this is not mandatory for the journal."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear Colleague,\n\nThe paper applies spiking neural networks (SNNs) to predict fatigue crack growth and Turing pattern evolution from simulated digital libraries, comparing them to standard ANN baselines. That specific application is new, and the authors deserve credit for a clean comparison across ConvLSTM, PredRNN++, SimVP, and STCLIF. The parameter-count advantages of the spiking models are real and clearly documented.\n\nThe problem is the headline quantitative claim. The 27.71% STCLIF error reduction on Turing patterns comes from a test set of only five independent trajectories, expanded by overlapping sliding windows into 240 predictions. Adjacent windows share 19 of 20 frames, so the effective number of independent test cases is about five. No error bars, no multiple seeds, no per-trajectory errors are reported. On that evidence the improvement is not statistically grounded. The FCG benchmark is less affected (108 independent trajectories), but the Turing result is the one in the abstract.\n\nThe architecture analysis (reduced weight variance, sparser connectivity in STCLIF) is suggestive but post-hoc; it does not by itself validate the accuracy difference. I also note the memory claims are based on parameter counts, not measured energy or latency on hardware, which is a minor overstatement.\n\nThe paper is otherwise clearly written and honestly frames the limitations of base models. The \"digital library\" framing is not conceptually new—it's training on simulation data—but the application domain and the SNN comparison are legitimate and useful. The circularity concern raised by the reader does not hold; no results are derived from fitted parameters.\n\nThis deserves serious peer review. The authors should be asked to run multiple seeds, report error bars, and ideally release code and data so the Turing result can be checked. If the 27.71% survives re-evaluation, the paper would be a solid contribution to ML-for-scientific-surrogates.\n\nBest regards.","headline":"Plausible new application of spiking networks to microstructure surrogates, but the headline Turing accuracy gain rests on five independent test trajectories and needs robustness checks before it can be taken at face value.","tokens_in":12317,"tokens_out":2614,"would_cite":false,"duration_ms":24399,"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":"Spiking neural networks beat artificial networks on microstructure prediction at lower cost.","keywords":["spiking neural networks","microstructure evolution","fatigue crack growth","Turing patterns","digital libraries","leaky integrate-and-fire","spatiotemporal predictive learning","event-driven computation"],"falsifier":"Rerun the Turing-pattern comparison on a test set of many independently initialized simulations with non-overlapping windows, and require that STCLIF's 27.71% mean-absolute-error reduction over the ANN models persists with confidence intervals; the claim collapses if the gap shrinks to within noise. A supporting check is whether the same ranking holds under multiple random seeds and across an order-of-magnitude range of training-set sizes.","tokens_in":71,"feed_emoji":"⚡","tokens_out":6093,"duration_ms":97509,"temperature":0.7,"pith_summary":"The paper tries to establish that spiking neural networks, which encode information as threshold-triggered events rather than continuous activations, can predict the evolution of material microstructures—fatigue crack growth and Turing patterns—as accurately as recurrent convolutional artificial networks while using far fewer parameters and less memory. On the Turing-pattern benchmark, the spatiotemporal circuit leaky integrate-and-fire (STCLIF) model reports a 27.71% error reduction relative to the ANN models tested. The claimed mechanism is that interface-dominated spatiotemporal patterns are naturally event-driven, matching the spike-based representation, and the trained spiking networks exhibit smaller weight magnitudes and sparser connections. If correct, the result breaks the usual accuracy-cost trade-off of computer vision models and makes low-power, real-time structural health monitoring and materials processing control feasible.","feed_headline":"Spiking nets beat ANNs on material microstructure forecasts","feed_subtitle":"STCLIF reports 27.7% lower error with an order of magnitude fewer parameters, enabling low-power monitoring.","key_machinery":"The load-bearing object is the spatiotemporal circuit leaky integrate-and-fire (STCLIF) neuron, a spiking unit whose membrane potential decays over time and that adds an autaptic synaptic circuit—a self-connection—so that spike history shapes later firing. It processes space and time in a single integrated scheme, unlike ANN baselines that separate a CNN spatial encoder from an RNN/LSTM temporal processor. The paper also relies on the 'digital library' concept: time-series image data generated by a fatigue-crack XFEM solver and Gray-Scott reaction-diffusion simulations replace PDE-constrained modeling, so the network learns evolution purely from data. The argument is carried by comparing STCLIF and its base-SNN counterpart against ConvLSTM, PredRNN++, and SimVP on two benchmarks with contrasting interface types (sharp crack vs diffuse Turing patterns).","core_discovery":"The central claim is that on microstructure-evolution benchmarks, spiking neural networks do not merely approximate the accuracy of artificial neural networks at lower cost—they exceed ANN accuracy while reducing parameters and memory. Concretely, for Turing-pattern prediction the STCLIF model achieves the lowest mean absolute error of all models compared, a 27.71% reduction over the ANN baselines (ConvLSTM, PredRNN++, SimVP), with an order of magnitude fewer parameters; for fatigue crack growth the base SNN matches ANN/LSTM accuracy with roughly four to five orders of magnitude fewer parameters. The authors attribute this to the event-driven nature of spiking neurons and to architectural properties emerging after training: reduced weight variance, smaller weight magnitudes, and a connectivity density roughly 22% lower than ConvLSTM's, which act like built-in regularization. The message is that the dominant practice of scaling dense, pixel-regressing ANNs is not the only route to predictive accuracy on physical interface problems.","pith_inferences":["A caveat drawn from the paper's own supplementary setup: the Turing test set is built from 5 independent trajectories expanded into 48 overlapping windows each, so successive test samples share 19 of 20 frames; the 27.71% error reduction may partly reflect very few independent test cases, and an independent-trajectory validation would test this.","The pixel-versus-vector analogy suggests the spiking advantage should grow with resolution: as interfaces occupy fewer pixels relative to the bulk, event-driven sparse representations should waste less computation on homogeneous regions, a prediction that could be checked by varying image resolution and measuring the accuracy-cost gap.","The scaling discussion implies 3D extensions are the next test; if STCLIF retains its advantage with 3D convolutions on voxel data, the memory savings (about 3/4 of LSTM-based alternatives) would become more valuable for real-time monitoring.","One could test the regularization interpretation directly by penalizing weight magnitude and connectivity in the ANN baselines to see whether the accuracy gap narrows; if it does, sparsity rather than spike timing is the operative advantage."],"forward_implications":["If the result holds, event-driven spiking models become a practical choice for embedded structural-health monitoring, where energy and memory budgets are tight but interface predictions must be accurate.","The finding implies that model architecture should be chosen to match spatiotemporal complexity: localized interface problems need only about 1,000 parameters (base SNN), while collective diffuse patterns need millions (STCLIF), so resource allocation can be scaled to the physics.","The digital-library approach can be extended to problems such as grain growth and phase transformations, using either high-fidelity simulation data or experimental data, freeing prediction from the 'mathematical overconstraints' of governing PDEs.","Weight sparsity and small weight ranges—not raw parameter count—are presented as the source of the spiking model's generalization, pointing to architecture-level regularization as a design principle for physical prediction tasks."],"supporting_citations":[{"why":"Supplies the STCLIF model with autaptic synaptic circuits that the paper's central comparison is built around.","marker":"[9]"},{"why":"Provides the ConvLSTM baseline and the spatiotemporal predictive-learning setup the ANN comparison uses.","marker":"[6]"},{"why":"Provides the PredRNN++ baseline with ST-LSTM cells against which STCLIF's accuracy and cost are measured.","marker":"[7]"},{"why":"Provides the SimVP baseline, the SOTA CNN video-prediction model that STCLIF outperforms on Turing patterns.","marker":"[8]"},{"why":"Defines the leaky integrate-and-fire neuron dynamics used for the base SNN and STCLIF spiking neurons.","marker":"[12]"},{"why":"Supplies the spiking-neuron implementation equations the paper uses to define the SNN models.","marker":"[25]"},{"why":"Constructs the fatigue-crack-growth digital library via extended finite element path-slicing, the FCG benchmark's data source.","marker":"[13]"},{"why":"Defines the Gray-Scott reaction-diffusion equations used to generate the Turing-pattern digital library.","marker":"[20]"}],"fun_headline_variants":["Spiking nets beat ANNs on microstructure with 27.7% lower error","Event-driven spiking nets outperform ANNs on material growth predictions","STCLIF spiking model wins on microstructure accuracy and efficiency","Spiking neurons learn material patterns with fewer parameters than ANNs"],"cache_read_input_tokens":14592,"weakest_assumption_plain":"The reported accuracy advantage rests on test sets built from a handful of independent simulated trajectories, because sliding windows overlap heavily (Turing test samples share 19 of 20 frames), so the model comparison may have effectively only about five independent test cases.","fun_headline_variants_meta":{"raw":{"variants":["Spiking nets beat ANNs on microstructure with 27.7% lower error","Event-driven spiking nets outperform ANNs on material growth predictions","STCLIF spiking model wins on microstructure accuracy and efficiency","Spiking neurons learn material patterns with fewer parameters than ANNs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000737,"raw_usage":{"total_tokens":3275,"prompt_tokens":908,"completion_tokens":2367,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":524,"completion_tokens_details":{"reasoning_tokens":2292}},"tokens_in":524,"tokens_out":2367,"duration_ms":21749,"temperature":1.0,"reasoning_tokens":2292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T00:20:38.021517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Rerun the Turing-pattern comparison on a test set of many independently initialized simulations with non-overlapping windows, and require that STCLIF's 27.71% mean-absolute-error reduction over the ANN models persists with confidence intervals; the claim collapses if the gap shrinks to within noise. A supporting check is whether the same ranking holds under multiple random seeds and across an order-of-magnitude range of training-set sizes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the STCLIF model with autaptic synaptic circuits that the paper's central comparison is built around."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the ConvLSTM baseline and the spatiotemporal predictive-learning setup the ANN comparison uses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the PredRNN++ baseline with ST-LSTM cells against which STCLIF's accuracy and cost are measured."},{"cited_title":"IEEE Comput","cited_arxiv_id":null,"evidence_quote":"Provides the SimVP baseline, the SOTA CNN video-prediction model that STCLIF outperforms on Turing patterns."},{"cited_title":"Neurophysiol","cited_arxiv_id":null,"evidence_quote":"Defines the leaky integrate-and-fire neuron dynamics used for the base SNN and STCLIF spiking neurons."},{"cited_title":"132 108–120 11","cited_arxiv_id":null,"evidence_quote":"Supplies the spiking-neuron implementation equations the paper uses to define the SNN models."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Constructs the fatigue-crack-growth digital library via extended finite element path-slicing, the FCG benchmark's data source."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the Gray-Scott reaction-diffusion equations used to generate the Turing-pattern digital library."}],"review_version":1}