{"id":"770e8a53-ae49-4ea0-aa41-2c07b20b74b8","arxiv_id":"2504.14237","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"FSA-Heat, a frequency-spatial dual-domain network, predicts 2.5D IC temperature and gradient maps with over 99% lower RMSE and 4.23x faster inference than a GCN-based baseline on HotSpot data.","lead":"This paper introduces FSA-Heat, a neural network that predicts temperature and heat-flow gradients across the layers of a 2.5D chiplet package using both frequency and spatial information. It reports up to 99% lower error and 4x faster inference than a graph-network baseline, which could speed up thermal-aware chip design.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'over 99% RMSE reduction' claim rests on an unverified self-implemented GCN+PNA baseline whose reported errors are near-random; until that baseline is fairly tuned or released, the comparative claim is not established.","rationale":"The reader identified HotSpot data fidelity as the weakest assumption. I agree that this is a real limitation for real-chip transfer, but the most load-bearing concern for the paper's headline claim is the fairness and verifiability of the GCN+PNA baseline. The reported baseline RMSE of 5.14 and MAPE of 111% indicate a near-random predictor, so the claimed 99% improvement is unsurprising if the baseline was undertrained or improperly configured. This is a verification problem rather than an internal inconsistency: the ablations in Table II are internally consistent, and the architecture is plausible. However, without released code or a controlled rerun, the central comparative claim cannot be independently checked. A concrete test that retrains GCN+PNA with matched budget and hyperparameter search would settle whether the margin is real. If the margin persists, the conditional acceptance conditions are satisfied; if it collapses, the central claim fails. I therefore keep the reader's CONDITIONAL verdict rather than moving it to ACCEPT or REJECT.","tokens_in":9468,"tokens_out":6042,"duration_ms":59996,"concrete_test":"Obtain the original GCN+PNA implementation from the authors of [11] or require the current authors to release their reimplementation with exact hyperparameters. Then run a controlled comparison on the same 6,000-sample train/test split with matched training budget (epochs, optimizer, learning-rate schedule) and perform a hyperparameter search for GCN+PNA over learning rate, hidden dimensions, graph construction, and training length, reporting the best RMSE over at least 5 seeds. If the best-tuned GCN+PNA RMSE remains above 5 on the test set, the comparison stands; if it drops below 1, the 'over 99% RMSE reduction' claim is an artifact of an undertrained or mis-specified baseline.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim in the abstract and Section IV-B is that FSA-Heat outperforms GCN+PNA by over 99% RMSE and 4.23x inference time. That comparison depends entirely on the authors' own reimplementation of GCN+PNA, for which code, weights, hyperparameters, and training budgets are not released. The reported baseline numbers in Table I (RMSE 5.1424, MAPE 111.235%) are so poor that GCN+PNA is effectively producing near-useless predictions; any reasonable learned model would show a massive RMSE reduction against such a baseline. This is not an internal inconsistency, but it makes the central comparative claim unverifiable from the manuscript alone. A fair comparison requires matched training budgets, hyperparameter tuning, and multiple seeds, or the original authors' code. The standalone accuracy of FSA-Heat (RMSE 0.0415 on HotSpot data) may still be valid, and the ablations in Table II are internally consistent, but the headline 'outperforms GCN+PNA by over 99%' is not yet supported. A secondary concern, also present in the reader's verdict, is that HotSpot-only synthetic data limits transfer to real 2.5D packages, but the baseline-fairness issue is the more immediate blocker for the paper's central quantitative claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes FSA-Heat, a neural network for fast steady-state thermal prediction in 2.5D chiplet ICs. The network combines a frequency-spatial thermal encoder (FSTE) with high-to-low frequency learnable weights, a frequency-domain cross-scale interaction module (FCIFormer), and a frequency-spatial hybrid loss (FSL). The model is trained and evaluated on 6,000 HotSpot-generated samples on a 64x64x4 grid. The authors report RMSE 0.0415, MAPE 0.861%, PSNR 68.078 dB, and an inference speed of 0.0549 s, claiming over 99% RMSE reduction and a 4.23x speedup over a re-implemented GCN+PNA baseline. They also claim robust generalization to unseen thermal conductivity values and heat-source counts, supported primarily by visual comparisons.","tokens_in":9718,"tokens_out":3217,"duration_ms":31523,"significance":"If the reported results hold under transparent evaluation, FSA-Heat would be a useful contribution: it addresses a real need for fast and accurate thermal surrogates in 2.5D chiplet design, and the proposed frequency-spatial architecture is a plausible and nontrivial design. The ablation study in Table II is internally coherent and suggests that each proposed component contributes to accuracy. However, the central comparative claim depends on a self-implemented, unreleased GCN+PNA baseline whose reported errors are near-random; the paper does not provide code, data, error bars, or seed-level variability. The generalization claims are also not quantified. These issues currently prevent verification of the headline results.","major_comments":[{"comment":"The headline claim of 'over 99% RMSE reduction' is computed against a GCN+PNA implementation that is described only as 'implemented according to the detailed descriptions in [11]'. The baseline RMSE of 5.1424 and MAPE of 111.235% are so poor that they indicate the baseline is performing near chance; any reasonable learned model would show a massive RMSE reduction against it. The manuscript provides no hyperparameters, training budget, model selection procedure, or code for this baseline, and no independent verification that it matches the original GCN+PNA method. This makes the central comparative claim unverifiable. The authors should release code and trained weights, use the original authors' implementation if available, or perform a fair, tuned comparison with matched training budgets and multiple seeds, reporting the resulting baseline performance.","section":"Section IV-B, Table I"},{"comment":"The generalization experiments for unseen thermal conductivities (50% and -50%) and unseen heat-source counts (10, 60, 80) are reported only through qualitative image comparisons. No numeric metrics such as RMSE, MAE, MAPE, or PSNR are given for these test sets, so the claim of 'robust generalization capabilities' is not quantitatively established. The authors should provide tables or plots with the same metrics used in Table I for these extrapolation sets, along with corresponding baseline results and error bars.","section":"Section IV-B, Figures 3 and 4"},{"comment":"All reported metrics are single-run point estimates without standard deviations or confidence intervals. The ablation differences in Table II (e.g., RMSE 0.0415 vs. 0.0499 vs. 0.0519) are small and may be within run-to-run variance for deep networks. Without multiple seeds, the internal ablation conclusions and the claimed improvements over the spatial-only baseline are not statistically supported. The authors should report mean and standard deviation over at least three runs, or otherwise justify that the differences exceed noise.","section":"Section IV-A and Table II"}],"minor_comments":[{"comment":"The statement that 'the 3D-DCT and 3D-IDCT utilize even symmetry continuation to satisfy the implicit Neumann adiabatic boundary condition at the edges' needs a derivation or reference. DCT-II's implicit even extension enforces a zero derivative at the extension center, which is not generally identical to the Neumann condition at the physical grid boundary unless the grid point locations are explicitly matched. Since FSTE is a feature extractor, this does not invalidate the empirical results, but the physical justification should be corrected or qualified.","section":"Section III-B1, Eq. (2)-(5)"},{"comment":"The text defines RMSE as 'root mean absolute error'; the correct expansion is 'root mean squared error' or 'root mean square error'.","section":"Section IV-B"},{"comment":"The figure contains placeholder symbols such as '?1', '?2', '?3', '?4', and '?', which obscure the description of the network. These should be replaced with concrete labels or a caption explains their meaning.","section":"Figure 1"},{"comment":"The paper does not report how the 6,000 samples are split into training, validation, and test sets, nor whether the test cases are generated from the same distribution as the training cases. Please specify the split and ensure that no test samples overlap with training samples.","section":"Section IV-A"},{"comment":"The caption '?? increased by -50%' is confusing because a negative increase is a decrease. Rephrase as 'decreased by 50%'.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The main blocking issue is reproducibility and fairness of the baseline comparison. The paper's standalone accuracy on HotSpot data and the internal ablation consistency are promising, but without release of code/data or a properly tuned and documented baseline, the abstract's central quantitative claim cannot be assessed by readers. The generalization section also needs quantitative support. I would be willing to see a revised version that addresses these points. The topic fits the journal, but the current evidence level is insufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: worth a read, but the headline comparison is not yet established. What is new is the specific combination of 3D-DCT frequency filtering with learned band weights, cross-scale cross-attention in the style of CrossViT, and a frequency-plus-spatial loss, all applied to 2.5D IC thermal prediction on a 64x64x4 grid. That package is new, and the paper applies it to the right kind of problem. The ablations in Table II are internally consistent: removing the frequency branch, the spatial branch, or the frequency loss each degrades accuracy in a plausible direction. The generalization tests on unseen thermal conductivities and heat source counts are the right check for a surrogate model.\n\nNow the soft spots. The biggest is the comparison to GCN+PNA. The reported baseline RMSE of 5.14 and MAPE of 111% are close to random-guess numbers. Any reasonable learned model would show a massive relative improvement against that. The authors say they implemented the baseline according to the paper and matched accuracy on the original dataset, but they do not release code, weights, hyperparameters, or training budgets, so the fairness of that comparison cannot be checked. That makes the central quantitative claim unverifiable. A smaller issue: the paper claims 3D-DCT even symmetry satisfies the implicit Neumann adiabatic boundary condition, but the top heatsink boundary is convective, not adiabatic. That is a mismatch in the justification; it may not affect the learned predictions, but it should be addressed. Also, all numbers are single-run point estimates with no error bars or seed variance.\n\nWho should read this: people building fast thermal surrogates for 2.5D/3D IC design-space exploration. If the authors fix the baseline comparison, release artifacts, and clarify the boundary comment, the architecture is worth publishing. As it stands, treat the standalone FSA-Heat accuracy as plausible but the comparative advantage as unproven. I would send this to peer review—it has enough substance to warrant referee time—but I would expect major revision.","headline":"Architecture is a sensible new combination and the ablations hold together, but the 99% RMSE claim rests on a near-random self-implemented baseline and missing artifacts.","tokens_in":10325,"tokens_out":2384,"would_cite":false,"duration_ms":20361,"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":"FSA-Heat claims that adding high-to-low frequency encoding to a 3D convolutional network predicts steady-state temperature fields in 2.5D chiplets with over 99% lower RMSE and 4.23x faster inference than the GCN+PNA baseline.","keywords":["Thermal analysis","2.5D ICs","frequency-spatial domain","multi-scale interaction","thermal prediction","chiplet","deep learning"],"falsifier":"Generate the same chiplet layouts with a finite-element thermal simulator or measured silicon data and compare FSA-Heat's predictions to the solver's output at the hotspots; if the RMSE advantage over GCN+PNA does not stay near the reported 99% or the errors degrade under that distribution, the central claim is refuted.","tokens_in":9236,"feed_emoji":"🔥","tokens_out":8543,"duration_ms":66978,"temperature":0.7,"pith_summary":"The paper proposes FSA-Heat, a neural network that predicts steady-state temperature distributions across the four layers of 2.5D chiplet packages—heat source, thermal interface material, heat spreader, and heatsink—directly from geometric, heat-exchange, and heat-source parameters on a 64×64×4 grid. It aims to replace repeated runs of thermal solvers during design optimization, where speed matters as much as accuracy. The central claim is that extracting features in both the spatial and frequency domains lets the network capture global heat-dissipation structure that plain CNNs and graph networks miss, yielding over 99% RMSE reduction and a 4.23x inference speedup over the GCN+PNA baseline while generalizing to unseen thermal conductivities and heat-source counts. If the claim holds, thermal-aware floorplanning and hotspot checks for chiplet designs could run in milliseconds without a full simulation.","feed_headline":"Thermal prediction error drops 99% with frequency-spatial network","feed_subtitle":"New deep network predicts 2.5D chip temperatures 4.23x faster than a graph baseline.","key_machinery":"The central object is a frequency-spatial thermal encoder block (FSTE) built around a 3D-DCT/IDCT branch with learnable anisotropic frequency weights $W_f = e^{-(E_u u^2 + E_w w^2 + E_p p^2)}$ plus a multiplicative gating signal, run in parallel with a residual 3D-convolution spatial branch. A second module, FCIFormer, replaces plain skip connections with cross-attention computed in the frequency domain between four encoder scales, and a frequency-spatial hybrid loss adds an L1 frequency-domain term to spatial MSE. Together these components give the network a global receptive field while preserving local boundary smoothness via even-symmetry continuation that mirrors the adiabatic Neumann boundary condition.","core_discovery":"The authors claim that a frequency-spatial dual-domain architecture can serve as a fast surrogate for steady-state thermal analysis in 2.5D ICs. Their FSA-Heat network encodes the input thermal-conductance field through a 3D discrete cosine transform, multiplies the frequency coefficients by learnable high-to-low frequency weights, and transforms back, while a parallel 3D-convolution branch keeps local spatial detail; a cross-scale interaction module then fuses the multi-resolution encoder outputs in the frequency domain before a convolutional decoder emits the temperature field. Trained with a loss that combines spatial MSE with L1 penalties on the magnitude and phase of the 2D DFT, the model reports RMSE 0.0415 versus 5.1424 for GCN+PNA on the same 64×64×4 test grid, a PSNR of 68.078 versus 28.659, and 4.23x faster inference than the graph baseline.","pith_inferences":["The frequency branch's global receptive field suggests the same architecture could transfer to full 3D chiplet stacks or larger grids, but the paper only demonstrates 64×64×4 and excludes the DeepOHeat operator baseline, so cross-method comparison on 3D cases remains open.","The boundary-condition argument points to a testable variant: replacing the 3D-DCT with a sine transform for fixed-temperature Dirichlet boundaries would show whether the even-symmetry assumption is what carries the accuracy.","The claimed 99% RMSE reduction is measured against a single graph baseline on HotSpot-generated synthetic data; transfer to measured silicon temperatures or to other compact thermal solvers is not established by the paper."],"forward_implications":["If the reported accuracy holds outside the training distribution, thermal-aware placement and floorplanning loops can call the network thousands of times per optimization run at millisecond latency instead of invoking a thermal solver each iteration.","The even-symmetry continuation of the 3D-DCT means the network respects the adiabatic edge condition by construction, so temperature predictions near package edges should not need special boundary handling.","Because the network also outputs the temperature dissipation gradient, it can directly support adaptive grid partitioning for higher-resolution thermal simulation.","A 64×64×4 grid prediction in 0.0549 seconds makes the method suited to interactive design-space exploration for chiplet packages."],"supporting_citations":[{"why":"Supplies the HotSpot 7.0 compact thermal model and the 6,000-sample, four-layer 64×64×4 dataset used for training and ground truth.","marker":"[7]"},{"why":"Defines the GCN+PNA graph-convolution baseline whose RMSE, PSNR, and inference speed are the comparison targets of the paper.","marker":"[11]"},{"why":"Provides the cross-attention multi-scale transformer and position-embedding design that the FCIFormer module adapts for frequency-domain cross-scale fusion.","marker":"[16]"},{"why":"Motivates the claim that CNNs under-represent high-frequency components, which is the paper's stated reason for adding a frequency branch.","marker":"[10]"},{"why":"Names the DeepOHeat operator-learning baseline that the paper excludes from comparison, defining the scope of the claimed superiority.","marker":"[12]"}],"fun_headline_variants":["FSA-Heat cuts 2.5D IC thermal error by 99%","Dual-domain network predicts 2.5D chip heat 4.23x faster","Frequency-spatial net reduces 2.5D IC thermal error 99%","Fast thermal prediction for 2.5D ICs with frequency-spatial net","New network: 99% less thermal error, 4.23x faster in 2.5D ICs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim rests on HotSpot-generated synthetic data—Gaussian power densities, a 64×64×4 grid, and thermal conductivities varying by ±30%—faithfully representing real 2.5D chiplet packages.","fun_headline_variants_meta":{"raw":{"variants":["FSA-Heat cuts 2.5D IC thermal error by 99%","Dual-domain network predicts 2.5D chip heat 4.23x faster","Frequency-spatial net reduces 2.5D IC thermal error 99%","Fast thermal prediction for 2.5D ICs with frequency-spatial net","New network: 99% less thermal error, 4.23x faster in 2.5D ICs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000821,"raw_usage":{"total_tokens":3612,"prompt_tokens":980,"completion_tokens":2632,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":2515}},"tokens_in":596,"tokens_out":2632,"duration_ms":16296,"temperature":1.0,"reasoning_tokens":2515,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:53:46.753573+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Generate the same chiplet layouts with a finite-element thermal simulator or measured silicon data and compare FSA-Heat's predictions to the solver's output at the hotspots; if the RMSE advantage over GCN+PNA does not stay near the reported 99% or the errors degrade under that distribution, the central claim is refuted.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the HotSpot 7.0 compact thermal model and the 6,000-sample, four-layer 64×64×4 dataset used for training and ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the GCN+PNA graph-convolution baseline whose RMSE, PSNR, and inference speed are the comparison targets of the paper."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the cross-attention multi-scale transformer and position-embedding design that the FCIFormer module adapts for frequency-domain cross-scale fusion."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the claim that CNNs under-represent high-frequency components, which is the paper's stated reason for adding a frequency branch."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Names the DeepOHeat operator-learning baseline that the paper excludes from comparison, defining the scope of the claimed superiority."}],"review_version":1}