acceptodds
Under review as a conference paper at ICLR 2027

S3THERM: A SOURCE-SHEET NEURAL OPERATOR FOR 3D STEADY HEAT CONDUCTION

Abstract

Repeated thermal analysis requires predicting three-dimensional steady tempera- ture fields for changing volumetric heat sources. A neural predictor must capture how a source’s shape and position affect the full field and how the responses to multiple sources combine. For fixed geometry, material properties, and Robin boundary conditions, temperature rise is linear in the heat source. This makes a thin source layer a useful response unit: it retains an in-plane heat pattern and a representative vertical position, while responses to different layers can be com- bined. We introduce S3Therm, which uses one shared neural operator to predict the three-dimensional response to each layer and sums these responses to predict the field of the full source. The operator is trained from the governing equation and boundary conditions without paired temperature fields. Energy-balance rescaling and symmetry averaging refine the combined prediction. On ten held-out source configurations evaluated at matched COMSOL coordinates across seven planes, S3Therm achieves a three-seed mean per-layout RMSE of 0.397 K, compared with 1.529 K for supervised FactFormer.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Source-sheet neural operator, 3D steady heat conduction, physics-informed learning, volumetric heat sources, source superposition, Robin boundary conditions, integral-balance correction, symmetry averaging, thermal field prediction

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_5ZSmOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

S$^3$Therm predicts 3D steady temperature fields for changing volumetric heat sources under fixed geometry, conductivity, and Robin cooling. The source is reduced to finite-thickness sheets, each with a unit-peak footprint, a centroid height, and an amplitude. A shared height-conditioned branch–trunk operator, trained only from PDE and Robin residuals on individual synthetic sheets, predicts each sheet's response; amplitude-weighted responses are summed, rescaled for global energy balance, and symmetry-averaged. On held-out layouts scored at COMSOL plane samples, it outperforms supervised LANO and FactFormer on every layout; ablations examine the corrections and collapsing source heights into a single sheet.

Critical strengths
  1. Single-sheet physics-informed training generalizes to multi-height layouts. With cross-interval addition and positive homogeneity in amplitude explicit before correction (Eqs. (2), (5)), the operator is trained on individual synthetic sheets (Sec. 4.3), yet the complete predictor has the lowest error on every held-out layout (Table 1). This is consequential because a model trained without paired temperature fields outperforms the evaluated supervised baselines in this fixed-physics setting.

  2. Physically motivated inference-time corrections with modest latency overhead. Rescaling enforces discrete integral balance and averaging enforces in-plane equivariance (Eqs. (7)–(8)); together they give the lowest mean and worst-layout RMSE (Table 2), whereas applying both to LANO does not improve its accuracy (Sec. 5.3). This is consequential because the rescaling addresses a global source–boundary mismatch that local residual training can leave (Sec. 4.5).

Critical weaknesses
  1. The benefit of the sheet decomposition over complete-source inputs is not isolated. The representation control supports resolving source heights over a single collapsed sheet (Sec. 5.3). However, no physics-informed operator using the complete-source input contrasted in Sec. 1 is compared, although DeepOHeat is named as particularly relevant (Sec. 2); this gap is distinct from the per-stage error separation acknowledged in Sec. 4.5. Moreover, without rescaling, the composed predictor is on average worse than both baselines (Tables 1–2). This is consequential because the conclusion presents the decomposition as a useful inductive bias (Sec. 6), yet its independent contribution to the advantage over supervised baselines, beyond the rescaling gain credited above, remains unclear.

  2. The practical cost benefit is not established, and the representation-control protocol is unclear. Setting aside the acknowledged budgeted, native-resolution protocol (Sec. 5.1), repeated-solve cost motivates the work (Sec. 1), yet S$^3$Therm's training time is unreported and its inference time is compared with neither COMSOL nor the baselines (Table 2). The control's multi-sheet reference (Sec. 5.3) also falls below the Table 1 mean over the same layouts, with its checkpoints and settings unstated. This is consequential because these omissions limit assessment of the practical speedup and of how the control's effect relates to the main benchmark.

Questions for authors
  1. (Related to W1; suggested experiment.) On the existing held-out layouts, how does a physics-informed operator using complete-source input (e.g., the composite-volume branch input in Sec. 4.2) compare at comparable capacity, budget, and corrections?
  2. (Related to W2; clarification.) What is S$^3$Therm's training time relative to the baselines' budget (Sec. 5.1), and how do all methods' end-to-end inference times compare with COMSOL solve times on these layouts? Which checkpoints and corrections produce the Sec. 5.3 reference, and why does it differ from Table 1?
  3. (Related to S2; clarification.) How thick are components relative to WW (Eq. (4)), and does the lifted source, whose budget Eq. (7) closes, carry the COMSOL source's integrated power?
  4. (Related to S1; clarification.) How do the held-out stress layouts (Sec. 5.1), once reduced to sheets, compare with the synthetic training sheets in footprint shape and height, and which generator settings produce the latter (Sec. 4.3)?
  5. (Related to S1; suggested experiment.) Both baselines are attention-based; how does a standard convolution-based surrogate (e.g., a U-Net, or the FNO cited in Sec. 2), trained on the same paired layouts and budget, perform on the held-out layouts?

If any part of this review reflects a misunderstanding of the paper, I would appreciate the authors pointing it out, and I am happy to revise my assessment accordingly.

3 comments

Reviewer 5ZSm

Response to Reviewer 5ZSm (W2, Question 2): Numerical Consistency in Section 5.3

Thank you for identifying this inconsistency. When we expanded the experiments to three random seeds, we updated Tables 1 and 2 but overlooked the single-seed values in the four-layout source-representation paragraph in Sec. 5.3. The reported 0.296 K and 1.020 K are therefore single-seed results. We apologize for this oversight.

We have now evaluated both source representations using the same three trained models as in the main comparison. Within each training run, the model weights are held fixed between the two representations, and both use the same balance correction, symmetry averaging, and COMSOL evaluation coordinates. The control preserves the integrated in-plane footprint and discrete source budget while replacing the three retained heights by one power-weighted centroid. We first average the layout RMSEs over the four layouts for each seed, then compute the mean and sample standard deviation across the three seeds.

The updated mean RMSE is 0.393 ± 0.085 K with the source heights retained and 1.104 ± 0.188 K after centroid collapse. The RMSE increases for every layout under all three random seeds. Thus, the benefit of retaining the vertical source distribution is consistent across the three runs.

We will replace the single-seed values in Sec. 5.3, state the averaging procedure explicitly, and provide the per-seed results in the supplement. Table 2 will also identify the results as averages over three random seeds.

Response to Reviewer 5ZSm (Question 3): Component Thickness and Integrated Source Power

Thank you for requesting clarification of the component thickness and source power. The nominal lifted-sheet thickness matches the physical component thickness in our experiments. Each component is 0.05 mm thick. For the 1 mm domain, W=0.05W=0.05 therefore corresponds to the same 50 μm thickness. The parameter PP controls the maximum height interval used to group source slices, whereas WW specifies the finite-thickness source whose response is learned. Their default normalized values coincide, but they enter different steps of the construction.

We traced the power from the specified component geometry through voxel encoding, sheet construction, and finite-thickness lifting. Across all ten layouts, the largest relative difference between the encoded-volume power and the lifted-sheet discrete budget is 4.3×10−84.3\times10^{-8}. Thus, sheet construction preserves the encoded source power to numerical precision in these experiments. Equation (7) balances this discrete source budget against the predicted boundary flux.

The distinction from the geometrically specified source power can be quantified. In Dense small blocks, the geometric power is 31.500 mW, voxel encoding gives 32.288913 mW, and the lifted-sheet discrete budget is 32.288914 mW. Integrating the bilinearly interpolated lifted source gives 32.418458 mW, 2.916% above the geometric power. The trace therefore identifies rasterization and integration as the sources of this discrepancy, while the sheet-construction step preserves the encoded budget.

We also evaluated the temperature effect directly on a common 500×500×120500\times500\times120 finite-volume grid, solving for the response to the difference between the lifted and original block sources. The RMS temperature difference at the reference plane coordinates averages 0.0596 K over the ten layouts; the largest layout RMS is 0.3499 K for Dense small blocks. We will clarify PP and WW near Eq. (4), identify the discrete budget in Eq. (7), and add the power trace and temperature comparison to the supplement.

Response to Reviewer 5ZSm (W2, Question 2): Training Budgets under Different Supervision Protocols

Thank you for asking how to compare the budgets of supervised and physics-residual training. Data acquisition is difficult in this field, so the present study uses twelve labeled layouts within the available cost budget. S3Therm runs for 100,000 updates and minimizes the PDE and Robin boundary residuals on synthetic source sheets. Each run instantiates four successive pools of 2,000 sheets, with a simultaneous pool size of 2,000 and continuously sampled source height. The supervised baselines use COMSOL temperature labels from the twelve training layouts, with a nominal training budget of about 10 hours; checkpoints are selected according to the recorded training-loss criterion.

We separate training cost into label generation, data preprocessing, network optimization, and checkpoint selection, and report wall-clock time, GPU count, update count, and sampling scale on the same hardware. This presents label-generation, model-optimization, and model-selection costs separately, allowing the cost structures of supervised and physics-residual training to be compared directly.

Response to Reviewer 5ZSm (W1 and Question 4): Training-Source and Evaluation-Layout Distributions

Thank you for asking about the relationship between the evaluation layouts and the training source sheets. Because data acquisition remains difficult, the current data are still generated in COMSOL. For the 1 mm computational domain, the in-plane size of the rectangular training generator is 0.10 × 0.05 mm, while the evaluation rectangles have xx- and yy-direction sizes of 0.04--0.26 mm and 0.035--0.12 mm, respectively. Training also includes smooth random-field sources. Training source heights are sampled continuously in the normalized interval [0,1], and evaluation component centers lie in 0.18--0.82 mm.

Each evaluation source map is normalized before entering the operator, and its amplitude is applied to the predicted response. The normalized source-map distribution and source-amplitude range are therefore reported separately. The ten layouts contain 1--36 components at one or three heights, with the same in-plane pattern repeated at each occupied height. The experiment evaluates the shared operator on new spatial arrangements and cross-height compositions; the generator settings and evaluation ranges will be presented in the same table.

Response to Reviewer 5ZSm (W2, Question 2): Total Cost and Workloads for Response Reuse

Thank you for asking about the cost implications of training without paired labels and the workloads that benefit from response reuse. Omitting paired temperature-field labels reduces total cost. S3Therm takes 3.434 s per layout with an RMSE of 0.297787 K; FactFormer, LANO, and FNO take 2.056 s, 9.619 s, and 1.854 s, with RMSEs of 0.915161 K, 0.778929 K, and 0.831163 K, respectively. CPU finite volume takes 2.266 s with an RMSE of 0.050769 K on the coarser grid and 5.853 s with an RMSE of 0.046687 K on the finer grid. Among the learned models, S3Therm has the lowest measured RMSE.

The suitable workload is high-throughput design exploration: under fixed geometry and boundary conditions, many heat-source footprints, positions, and height combinations are compared while the same trained operator repeatedly reuses source-sheet responses. When paired COMSOL temperature labels are difficult or costly to obtain, the PDE/Robin training objective also removes this label requirement.

Reviewer_5ZSmOpenReview ↗

Response to Authors

I thank the authors for their careful answers to my questions. The multi-seed results resolve the inconsistency in Sec. 5.3 and confirm the benefit of retaining source heights, and the power trace answers my question on component thickness. However, the added timings show a CPU finite-volume solver to be both faster and more accurate on its coarser grid, and the comparison with a physics-informed operator using complete-source input is not addressed, so the practical benefit and the contribution of the decomposition remain unclear. I maintain my assessment of 2 (weak rejection).

Reviewer_iGNZOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper addresses repeated prediction of 3D steady-state temperature fields in a fixed setting (1 mm cube, uniform conductivity, linear Robin boundary) where only the volumetric heat source changes. S3Therm thresholds the source along z, groups the retained slices into thin height intervals ("source sheets"), and represents each sheet by a unit-peak 2D footprint, a centroid height, and an amplitude. A single DeepONet-style operator (transformer branch over the footprint, Fourier-feature trunk over (x, y, z, z0)) is trained only from PDE and Robin residuals on 2,000 synthetic sheets, and sheet responses are summed by linear superposition. At inference, the summed field is rescaled by a scalar so that total boundary flux matches total source power, then averaged over four in-plane reflection symmetries. On ten held-out layouts evaluated against COMSOL samples on seven planes, the full predictor reaches a mean RMSE of 0.397 K vs. 1.529 K (FactFormer) and 1.660 K (LANO), both supervised on twelve labeled layouts. Ablations quantify the two inference-time corrections, and a four-layout control shows that collapsing all sheets into a single centroid raises RMSE from 0.296 K to 1.020 K.

Critical strengths
  1. Label-free training with consistent wins on held-out layouts. The operator is trained from PDE and boundary residuals only, yet the complete predictor outperforms both supervised baselines on all ten held-out layouts with low seed variance (std 0.099 K). Removing the need for paired solver data is practically meaningful when reference solves are expensive.

  2. The sheet decomposition is supported by a targeted control. Collapsing three sheet responses into one power-weighted centroid raises RMSE from 0.296 K to 1.020 K, giving direct evidence that height-resolved composition matters for this problem. The limitations section is also candid about the non-additivity of the corrected predictor and the plane-only evaluation.

Critical weaknesses
  1. Limited methodological novelty for an ML venue. Each component has close precedent:
  • Physics-informed DeepONet training without paired fields for 3D-IC thermal analysis, with 2D/3D power maps as inputs (DeepOHeat, Liu et al., DAC 2023; DeepOHeat-v1, Yu et al., 2025).
  • Exploiting linearity/superposition of linear PDE solution operators via Green's-function learning (e.g., Neural Green's Functions, Yoo et al., NeurIPS 2025, which additionally generalizes across geometries and source/boundary functions).
  • Post-hoc enforcement of an integral conservation law (ProbConserv, Hansen et al., ICML 2023; not cited).
  • Group averaging for equivariance (frame averaging, Puny et al., ICLR 2022; G-FNO, Helwig et al., ICML 2023). What is new is mainly the granularity of the decomposition (z-interval sheets rather than point sources or the full source). This is a reasonable discretization choice for the application, but layer-wise power maps are already a standard representation in 3D-IC thermal modeling. The paper provides neither analysis (e.g., approximation error of thresholding/centroiding vs. interval width) nor evidence of generality (geometry, conductivity, and boundary coefficients are all fixed) that would turn this into a transferable ML contribution. In its current form, the work seems better suited to an EDA or computational-physics venue.
  1. The experiments do not isolate the claimed contribution, and the comparison set is weak.
  • In Table 2, the raw sheet composition (the learned operator itself) has a mean RMSE of 2.030 K (worst 6.643 K), worse than both supervised baselines. Most of the improvement comes from a single global scalar rescaling (2.030 -> 0.678 K). This suggests the residual-trained operator does not satisfy global energy balance, and that the headline result depends mainly on a domain-specific post-correction rather than on the learned sheet operator.
  • The corrections are applied only to LANO; FactFormer + corrections is not reported. The most relevant baselines, DeepOHeat (described in the paper as "particularly relevant") and Neural Green's Functions, are cited but not compared. There is also no ablation with a physics-informed DeepONet that takes the full source without sheet decomposition, under the same training budget and corrections.
  • The supervised baselines are trained on only twelve labeled layouts under a ~10-hour budget, and LANO's seed std (1.557 K) is close to its mean (1.660 K), indicating unstable or under-tuned training.
  • The efficiency motivation is not quantified. COMSOL wall-clock time is not reported, and S3Therm's inference cost (~3.6 s and ~6 GB GPU memory per layout) is not compared with classical fast alternatives for a fixed-geometry linear problem (e.g., multigrid/reduced-order solves or the cited compact thermal models HotSpot/3D-ICE).
Questions for authors
  1. (Could change my assessment) What are the results of (a) a physics-informed DeepONet/DeepOHeat that takes the full 3D source without sheet decomposition, trained with the same residual budget, and (b) the same model with the balance and symmetry corrections? This comparison is needed to attribute the gain to the sheet decomposition rather than to the corrections.
  2. Please report FactFormer with the same corrections, and the distribution of the balance factor alpha_g across layouts and seeds. How far is alpha_g from 1 for the raw composition, and what does this imply about convergence of the residual training?
  3. What is the COMSOL wall-clock time per layout at the reference resolution? How does S3Therm compare with a classical fast solver that exploits fixed geometry and linearity (e.g., a multigrid FD solve on the same 500x500x120 grid, a reduced-order model, or a compact thermal model)?
  4. How were the size (12 layouts) and training budget of the supervised baselines chosen? How do they perform with more labeled layouts, given that labels come from the same solver used for evaluation?
  5. How sensitive is accuracy to the retention threshold (1%), maximum interval width (0.05), and lifted thickness W? Can the representation error introduced by centroiding within an interval be characterized?
  6. (Future work, not required for this submission) Can the approach extend to varying conductivity, boundary coefficients, heterogeneous layers, or geometry, as handled by DeepOHeat-v1 and Neural Green's Functions?
3 comments

Reviewer iGNZ

Response to Reviewer iGNZ (W1, Critical Weakness 1): Methodological Contribution and Relation to Prior Work

Thank you for asking us to clarify the methodological contribution and its relation to prior work. S3Therm uses a finite-thickness source sheet as the common unit of physics-informed training and response reuse. Under fixed geometry, conductivity, and linear Robin boundary conditions, a shared height-conditioned operator maps a normalized two-dimensional heat-source distribution to its three-dimensional temperature response. The operator is trained from the heat equation and boundary residuals without paired temperature-field training data. The predicted responses are multiplied by their respective source amplitudes and summed across heights, followed by balance correction and symmetry averaging.

The sheet representation is motivated by the geometry of heat generation. For a source concentrated in a few planar layers, the number of physical source layers can remain fixed as the computational grid is refined. For a source grid of size Nx×Ny×NzN_x\times N_y\times N_z, representing these regions by MM retained sheets requires MNxNyMN_xN_y footprint values instead of NxNyNzN_xN_yN_z volumetric values. When M<NzM<N_z, this reduces source storage while retaining the in-plane variation within each aggregated footprint. An explicit row-wise or point-wise partition introduces more response terms as the source resolution increases. The sheet representation exploits localization in the thickness direction without assuming that the in-plane source distribution is separable. All sheet responses share the same learned parameters.

Relative to DeepOHeat, the distinction lies in this source decomposition and reuse of sheet responses. In our attempted high-resolution benchmark setting, the original DeepOHeat exceeded the available GPU memory and could not be evaluated. DeepOHeat-v1 also reports that the original model cannot be trained on its 101×101×56101\times101\times56 collocation grid because of memory limitations. Its separable training improves coordinate-wise basis evaluation and differentiation; S3Therm changes the source representation and the unit of response prediction.

The distinction from Green-function methods concerns both computation and training. GreensONet evaluates source and boundary integrals at a cost proportional to the product of query points and integration elements. S3Therm directly learns the temperature response to an entire finite source sheet, incorporating the effect of in-plane integration into that response. Prediction then consists of footprint encoding and summation of the predicted sheet responses. Neural Green’s Functions uses a low-rank representation to reduce integral-evaluation costs and is trained using FEM solutions and mass matrices. Our response operator uses PDE and Robin boundary residuals for self-supervised training.

We will clarify these distinctions in the Introduction and Related Work, with the computational rationale described in Sec. 4.2. The related formulations also motivate future work on how the choice of source decomposition affects spatial representation and computational cost. We will cite the relevant conservation and symmetry literature when describing the inference-time corrections.

Response to Reviewer iGNZ (W1, Source-Decomposition Evidence): Numerical Update for the Four-Layout Control

We would also like to correct the single-seed values cited for the control experiment in Sec. 5.3, which we inadvertently left unchanged when updating Tables 1 and 2. Averaging the four layout RMSEs within each run and then summarizing across the three random seeds gives 0.393 ± 0.085 K with the source heights retained and 1.104 ± 0.188 K after centroid collapse (mean ± sample standard deviation). The RMSE increases for every layout under all three seeds.

Response to Reviewer iGNZ (W2, Question 2): Matched Corrections and the Meaning of the Balance Factor

Thank you for suggesting that FactFormer should receive the same correction settings. Under the same ten-layout protocol, we evaluated balance correction, symmetry averaging, and both corrections for S3Therm, FactFormer, and LANO. With both corrections, the mean RMSEs are 0.408 ± 0.097 K for S3Therm, 0.953 ± 0.078 K for FactFormer, and 0.832 ± 0.394 K for LANO; these values are the mean and sample standard deviation across three runs. With identical post-processing, S3Therm therefore retains a 57.2% lower mean RMSE than FactFormer.

The height-wise sheet construction intentionally introduces an approximation in order to obtain an approximately linear response representation. Applying a global correction in the subsequent stage provides a direct way to correct this discrepancy.

The factor has the following interpretation:

  • αg>1\alpha_g>1: the boundary budget of the assembled prediction is smaller than the reconstructed source budget, so the model underestimates the overall boundary response and the predicted field is scaled up.
  • αg<1\alpha_g<1: the boundary budget of the assembled prediction is larger than the reconstructed source budget, so the model overestimates the overall boundary response and the predicted field is scaled down.
  • αg≈1\alpha_g\approx1: the raw prediction is already close to discrete source--boundary integral closure.

Response to Reviewer iGNZ (W2, Question 4): Label Count, Training Budget, and Checkpoint Rules

Thank you for asking how the label count, training budget, and checkpoint rules affect the comparison. Data acquisition is difficult in this field, so the present study uses twelve labeled layouts within the available cost budget. With the same training and test split, the supervised baseline has a more direct target and therefore achieves better performance, whereas S3Therm is trained from the heat-equation and Robin boundary residuals without paired temperature-field targets.

Each S3Therm training pool contains 2,000 synthetic source sheets and is refreshed during optimization. LANO and FactFormer are trained with COMSOL temperature fields from the twelve layouts. Data-generation cost and model-training time are reported separately.

Response to Reviewer iGNZ (W1 and Question 5): Source-Representation Error and Validity Conditions

Thank you for asking how source-representation parameters affect accuracy. Reducing the interval width PP from 0.05 to 0.025 changes the mean RMSE across the three seeds from 0.408 to 0.403 K; 14 of 30 paired cases improve and 16 worsen, indicating a small aggregate change under this setting.

We also fixed the finite-volume operator and measured the temperature difference produced by replacing the original block source with the lifted source. The mean layout RMS at the reference coordinates is 0.0596 K, and the largest value is 0.3499 K for Dense small blocks. For that layout, the interpolated lifted source has 2.916% more power than the geometric source.

For the ideal centroid approximation, the separable source construction gives pointwise cancellation of the first vertical moment. A second-order approximation bound follows in regions where the second derivative of the response kernel with respect to source height is bounded. The revision will distinguish this analytical statement from the measured effect of finite-thickness lifting and report the parameter comparisons together.

Response to Reviewer iGNZ (W2, Question 3): Total Cost and Workloads for Response Reuse

Thank you for asking whether avoiding paired temperature-field labels reduces total cost and when response reuse is useful. The completed aligned-output timing covers ten layouts, three repeats per layout, and the full inference path. S3Therm takes 3.434 s per layout with a reference-plane RMSE of 0.297787 K; FactFormer takes 2.056 s with 0.915161 K; LANO takes 9.619 s with 0.778929 K; and FNO takes 1.854 s with 0.831163 K. The CPU finite-volume solver takes 2.266 s at 250×250×60 with an RMSE of 0.050769 K and 5.853 s at 500×500×120 with an RMSE of 0.046687 K. Under this timing protocol, S3Therm has the lowest RMSE among the learned models.

The method is intended for design sweeps: when geometry, material properties, and boundary conditions are fixed, it repeatedly evaluates new heat-source layouts or multi-height combinations and reuses source-sheet responses across candidates. When COMSOL temperature labels are difficult or costly to obtain, PDE/Robin residual training also removes the need for these paired labels.

Response to Reviewer iGNZ (W1, Question 6): Applicability and Future Extensions

Thank you for suggesting these extensions. The shared source-sheet response method provides a starting point for conditioning the operator on material and boundary parameters. For each fixed linear physical configuration, source responses remain additive; an operator conditioned on physical parameters can learn responses across different configurations. We plan to include conductivity and Robin coefficients as operator inputs and train over parameter ranges to study this direction.

Geometric variation requires a corresponding geometric representation. Temperature-dependent conductivity and nonlinear boundaries require nonlinear response treatment because the current additive construction is based on a linear governing problem. We will distinguish these directions in the discussion; the present evaluation remains focused on source-layout variation under fixed physical conditions.

Reviewer_1tfLOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper developed a pipeline to predict 3D temperature field produced by a changing volumetric heat source. The main idea, based on the superposition principle, is to decompose the volume into multiple slices, predict the temperature distribution caused by each slice, then add them together.

Critical strengths

Based on the superposition principle, a divide-and-conquer idea makes sense.

Critical weaknesses
  1. The architecture and training scheme closely follow DeepOHeat (DAC 2023), which feeds the full map into the network and predicts the temperature directly. The key idea here is to decompose, predict separately, and add the results. But within each slice there is no further decomposition. All blocks at the same height are merged into one footprint and passed to the network together, so the network must still learn how these blocks combine. This is exactly the limitation the introduction attributes to full-source input. Since superposition holds for any partition of the source, a more consistent design would also decompose each slice, for example into individual blocks. The authors should compare per-block decomposition with the current per-height decomposition, or explain why decomposition stops at the slice level.

  2. If decomposition and addition is the key idea, any correction should be applied to each slice before the addition, not to the sum afterwards. Energy balance holds for each slice on its own, so a per-slice scaling factor can be computed in the same way. Once each slice is corrected, the slices can be added without further adjustment. Instead, the paper applies one global scaling factor after the addition, and this step plays a critical role in the final accuracy. This suggests that the individual slice predictions are not accurate on their own, which weakens the central claim that slice responses are reusable units that can simply be added. The authors should (a) report the values of the scaling factor across layouts, and (b) compare per-slice correction with the current global correction.

Post-rebuttal: I read the authors’ response and new results. My main concerns on W1 and W2 remain (see my comment below), so I keep my rating.

Questions for authors

See weaknesses.

3 comments

Reviewer 1tfL

Response to Reviewer 1tfL (W1, Critical Weakness 1): Rationale for Decomposition by Source Height

Thank you for asking why the decomposition stops at the sheet level. We choose a sheet to retain the distributed in-plane source pattern while making composition across source heights explicit. For sources confined to a few thin, planar regions, the number of such regions need not increase when the spatial grid is refined. Blocks within each retained height interval are intentionally encoded together as one two-dimensional footprint. The shared operator learns the three-dimensional response to that footprint, conditioned on its height; amplitude-weighted summation then combines responses from different intervals.

This choice separates two tasks: learning the response to a spatially distributed source and composing responses across heights. The first task is handled by the sheet operator, trained using PDE and Robin boundary residuals without paired temperature fields. The second is implemented directly by superposition. We will revise the Introduction to state this distinction clearly, including that the operator jointly models the in-plane distribution within each sheet.

There is also a computational reason for this choice. For a layout with BB blocks, each contained in one of MM retained height intervals, decomposition by height requires MM sheet-response evaluations, whereas decomposition by block requires BB block-response evaluations, before applying the same symmetry transforms. A sheet can therefore account for several blocks at the same height in one response evaluation. Each input remains a two-dimensional footprint and a height coordinate, and increasing the number of intervals does not change the model parameters.

We appreciate the suggestion to investigate finer source decompositions and plan to pursue this direction in future work. Our present two-dimensional input retains the source’s in-plane shape and spatial distribution in each operator evaluation. With a finer decomposition, this spatial information would need to be represented through the individual source descriptors and their composition. We will therefore examine the choice of decomposition granularity together with the source encoding, prediction accuracy, and computational cost. The rationale for the present sheet representation will be clarified in Sec. 4.2.

Response to Reviewer 1tfL (W2 and Related Questions): Per-Sheet Versus Global Balance Correction

Thank you for suggesting a direct comparison of the two correction orders. Your proposal is theoretically sound and was also consistent with our early design intuition. Our subsequent analysis showed that, in an engineering setting, applying energy-balance correction to each sheet separately has limited effect when the individual sheet errors are small; correcting each sheet before summation can then introduce a bias in the assembled result. We therefore compared the current procedure with correction applied to each sheet before summation using three random seeds. Both settings use the same symmetry averaging. Across the ten layouts, the mean RMSE is 0.411 K for per-sheet correction and 0.408 K for global correction. In this experiment, per-sheet balance does not show a consistent performance improvement over correction of the assembled prediction.

The height-wise sheet construction intentionally introduces an approximation in order to obtain an approximately linear response representation. Applying a global correction in the subsequent stage provides a direct way to correct this discrepancy. The layout-wise scaling factors have mean 1.307600, sample standard deviation 0.562847, and range 0.414291--4.552238. We will report the factor distribution, paired results, and correction definition together.

I am not fully convinced.

  1. W1: The blocks at the same height are merged into one input, so the network still has to learn how these blocks combine. But this is exactly the limitation the authors pointed out for methods like DeepOHeat and what this work aimed to tackle.

  2. W2: The authors added experiments to apply the scaling to each sheet, and got similar results to the global scaling setting. However, it is unclear (not reported) how much the per-sheet factors vary. If the factors are not close to 1, then each sheet prediction is not accurate on its own, and the sum before scaling (the original design) is not accurate either. To me, per-sheet scaling first keeps the predictor additive , and it is more consistent with the paper's main idea.

Reviewer 1tfL

Response to Reviewer 1tfL (W1): Distinguishing 2D Surface-Source Input from 3D Volumetric-Source Composition

Thank you for pointing out that blocks at the same height are jointly encoded and that the network therefore still models their in-plane combination. We agree with this description of the present representation. Our clarification concerns the distinction between this within-sheet modeling and the source-composition limitation discussed for methods such as DeepOHeat.

In the reported DeepOHeat experiment, the source is provided as a two-dimensional power map on a three-dimensional computational domain, with heat generation confined to a single heated plane or layer. The three-dimensional temperature field is the output of the three-dimensional physical model, but the source input is not a decomposition of a general three-dimensional volumetric source into reusable two-dimensional responses. Extending the source input by adding a spatial dimension would provide a 3D source input, but it would still treat the complete volumetric configuration as one network input rather than explicitly composing responses from separate source heights.

S3Therm uses a different composition structure. Each retained height interval is represented by a two-dimensional footprint and a height coordinate, and the operator predicts the three-dimensional response of that finite-thickness source sheet. Responses from different heights are then composed explicitly through the linear superposition rule. Consequently, S3Therm removes the need for the network to learn cross-height composition from complete volumetric inputs. It does not remove the need to learn how multiple blocks within the same height interval combine; those blocks are intentionally merged into one footprint and are modeled jointly by the operator.

Response to Reviewer 1tfL (W2): Per-Sheet Balance Correction and Distribution of Scaling Factors

We agree that the final RMSE comparison alone does not establish whether each sheet response is accurate before composition. The previous response reported the comparable performance of per-sheet and global correction, but did not report the distribution of the per-sheet factors. That diagnostic is necessary for assessing the accuracy and balance of individual sheet responses.

We will therefore adopt the per-sheet correction order in the revised predictor: each finite-thickness sheet response will first be assigned its own balance factor, and the corrected sheet responses will then be summed across heights before symmetry averaging. For a fixed sheet decomposition, this preserves additive assembly at the response level. We will also report the per-sheet factors by layout and seed, including their mean, sample standard deviation, range, and deviation from one, together with the corresponding uncorrected and corrected sheet-response diagnostics.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.