acceptodds
Under review as a conference paper at ICLR 2027

Calibrated Thermal Twins for Localizing and Quantifying Thermal-Bridge Heat Loss

Abstract

We introduce a calibrated thermal twin that localizes thermal bridges and quantifies heat loss from as-built geometry, material information, and exterior infrared measurements. The central challenge is that distinct interior conductivity fields can produce the same surface temperature, making full-field reconstruction an ambiguous target for learned inversion. Our approach exploits the distinction between this hidden field and the surface transmittance that remains identifiable under known steady-state boundary conditions. The twin combines a geometry-conditioned neural surrogate for surface-temperature prediction with an amortized network for bridge localization. Its diagnostic stage recovers surface transmittance (U-value) through a physical surface-flux identity and localizes the lateral footprint of thermal bridges without reconstructing the interior. Split-conformal calibration provides U-value intervals with distribution-free, finite-sample marginal coverage under exchangeability. On the synthetic junction benchmark generated with a solver validated against ISO 10211 reference cases, direct surface readout achieves a U-value mean absolute error of , compared with when the same quantity is read from an optimized conductivity field. The surface-anomaly localizer achieves a bridge-footprint IoU of , compared with for optimization and at most for the learned field inverses. Across increasing measurement noise, U-value error remains substantially more stable than field-based optimization, while empirical conformal coverage remains at or above the nominal level. Experiments in heat conduction, Darcy flow, and electrical-impedance tomography further show that useful physical quantities can remain recoverable despite inaccurate field reconstruction. Experiments on real building geometry and measured infrared data further assess localization plausibility, while differentiable retrofit optimization demonstrates the downstream use of the thermal twin. These results support designing inverse methods around quantities that boundary observations actually identify.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
thermal digital twins, inverse problems, neural operators, building energy, thermal bridges, infrared thermography, identifiability, uncertainty quantification, conformal prediction, physics-informed machine learning, heat transfer, digital twins

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_ijxZOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a calibrated thermal twin framework for identifying building heat loss from surface temperature observations. The authors argue that interior conductivity fields cannot be uniquely recovered from limited boundary measurements and instead focus on estimating surface U-values and localizing thermal bridge footprints. The framework combines a geometry-conditioned neural operator, a calibrated surface-flux readout, and split-conformal prediction for uncertainty estimation. Experiments on synthetic building datasets and several inverse problems show that direct U-value estimation can outperform methods based on full-field reconstruction. Additional experiments on real building geometries and measured infrared data provide preliminary evidence of practical applicability.

Critical strengths
  1. Clear motivation. The paper distinguishes between recovering an interior conductivity field and estimating quantities supported by boundary observations. The non-uniqueness analysis provides a reasonable justification for focusing on U-values and thermal bridge footprints.
  2. Extensive experimental analysis. The experiments compare direct readout with optimization-based and learned inverse methods across different datasets and noise levels. The paper also reports cases where learned inverses perform better and evaluates uncertainty using conformal prediction.
Critical weaknesses
  1. Limited real-world validation. The main U-value accuracy results rely on synthetic data, while the measured infrared experiments lack ground-truth U-values. It remains unclear whether the calibration learned from synthetic envelopes transfers reliably to real buildings.
  2. Idealized noise and calibration assumptions. The experiments mainly consider independent measurement noise, whereas real thermal images may contain spatially correlated errors, emissivity bias, and uncertain boundary conditions. The conformal coverage guarantee also depends on exchangeability, which may not hold across different buildings or measurement settings.
  3. Inconsistent reporting of experimental results. The paper reports both original (40/32/32) and enlarged (120/96/96) validation sets, with different U-value errors appearing across sections. The relationship between these results and the corresponding experimental settings should be clarified to improve reproducibility.
Questions for authors
  1. Which validation sets support the main results reported in the abstract and tables? How do the original and enlarged evaluation sets relate to each other?
  2. How robust is the U-value estimation to spatially correlated noise, shared measurement bias, or uncertain boundary conditions?
  3. How well does the calibration learned from synthetic data transfer to real buildings, given the absence of measured U-value ground truth?
  4. How should the reported conformal coverage be interpreted under distribution shifts or repeated calibration-test splits with relatively small validation sets?
4 comments

Thank you for the informed and constructive review. We address each point below, in three comments. Every change in the revised PDF is typeset in blue.

[W1 & Q3: Real-world validation and calibration transfer]

Absolute U-value accuracy is established on the ISO 10211 validated synthetic corpora and not on measured buildings, because no public dataset pairs infrared imagery of a wall with a heat-flux-meter (ISO 9869) U-value and a layer table. The measured-data appendix claims plausibility of localization and nothing about U, and on the real TUM2TWIN shells, where the twin's prior-equipped inverse locates the inserted bridges at 0.74 and 0.66, U is degenerate at the clear-wall value. The abstract now states this scope (C1). A measured validation needs exterior infrared, interior and exterior air temperatures, a heat-flux plate on the same section and the construction record, and it is the next step.

The two calibration parameters have a physical meaning, which answers the transfer question. Regressing the raw slab readout on true U gives slope 1.11 on junction, 2.03 on hard and 2.37 on box. The readout averages a slab 2 % inside the outer layer, and how far θ has climbed from its face value r_se·U at that depth is set by that layer's conductivity. The affine map absorbs this, so it is corpus-specific and does not transfer (junction's map on hard gives 0.098 against hard's own 0.024). It is a sampling artefact of reading a point cloud inside the wall, not a sim-to-real device. A camera reads the face itself, where θ_face = r_se·U holds exactly and the correction vanishes. U-value MAE (W/m²K), enlarged sets, with two parameter-free replacements.

Corpus (geometries)Raw slabCalibrated (a, b), oursPrior-ratio, no fitClear wall
Junction (120)0.0930.0360.2200.283
Hard (96)0.1600.0240.0330.079
Box (96)0.1820.0330.0330.059

Scaling the raw slab mean by the 1-D prior's own slab-to-face ratio, computed from the layer table with no fit, replaces the calibration within 0.01 W/m²K on hard and box and fails on junction, where the bridges rather than the layer table set the near-face profile. Extrapolating the slab profile to the face fails wherever the slab straddles a layer interface (0.28 and 0.29 on hard and box). Narrowing the slab toward the face shrinks the raw bias as the physics predicts (raw slope 2.03 → 1.71 on hard and 1.11 → 1.06 on junction between 2 % and 0.5 % depth). The two parameters encode where the sampled slab sits inside the outer layer, and the revision reports the non-transfer and the prior-ratio readout in Appendix A.13 (C8).

[W2 & Q2: Measurement error beyond independent noise]

We repeated the junction noise sweep under three error models, exterior slab, five seeds, 40 geometries, the same calibration. (i) Spatially correlated noise, a Gaussian random field with correlation length 5, 15 or 30 % of the face and unit marginal variance, spatial mean kept. (ii) A shared gain θ·(1 + sζ), one ζ per image, the emissivity or drive-temperature case. (iii) A shared offset θ + sζ·mean(θ_face), the reflected-temperature case. ΔT is 20 K, so the largest level s = 0.04 is a 0.8 K error. U-value MAE (W/m²K) at s = 0.04 and the readout's 90 % conformal interval. At s = 0 every model gives 0.039 for the readout, 0.046 to 0.049 for optimization and 0.15 to 0.18 for Transolver. The clear-wall guess is 0.28 to 0.30.

Error modelReadout (ours)OptimizationTransolverCoverage / width of the readout interval
Independent (paper)0.0810.1100.1750.90 / 0.36
Shared gain0.0460.0510.1650.90 / 0.23
Shared offset0.0500.0560.1600.90 / 0.24
Correlated, 5 % of the face0.1230.1560.1810.89 / 0.49
Correlated, 15 %0.2310.2500.2040.89 / 0.94
Correlated, 30 %0.3400.3150.2000.90 / 1.51

A shared gain or offset is absorbed by the calibration, so Assumption 1 was too pessimistic there. Spatially correlated error is the limitation. A smooth perturbation at the slab scale is not averaged down, and at the largest level and length every route reading the surface falls past the clear-wall guess (0.34 and 0.32 against 0.28), while Transolver sits at 0.20 only because it is at 0.15 to 0.21 at every level including zero noise. Coverage stays at 0.89 to 0.91 in every cell while the width grows from 0.22 to 1.5, so the interval reports the failure. The readout beats both field routes at every level at 5 % correlation length (p = 4×10⁻⁴ and 0.047 at s = 0.04), up to s = 0.02 (0.4 K) at 15 % and s = 0.01 (0.2 K) at 30 %. The physical reading is the size of the signal. The exterior face sits r_se·U·ΔT, about 0.4 K, above outdoor air. Independent camera noise of 0.03 to 0.05 K averages over the roughly 80 readout points to a few percent of that signal. A smooth 0.4 K error at the slab scale is the signal itself, and no estimator reading that surface can remove it. The revision states this envelope with the interval flagging anything beyond (Appendix A.12, C7).

Uncertain boundary conditions. An error in the outdoor reference temperature enters as the shared offset and an error in the indoor-to-outdoor contrast as the shared gain, both measured above. An uncertain exterior film resistance r_se is a multiplicative error on U for every route, since all read U = θ_face / r_se, so a 25 % error, the spread between calm and windy conditions, costs 25 % of U for every method and for the ISO 6946 formula itself, which is why exterior surveys are specified at low wind. The revision states dU/U = −dr_se/r_se (C7).

[W3 & Q1: The two validation-set sizes]

The 40/32/32 geometries are the original validation sets. The 120/96/96 sets were generated afterwards from the same validation seed and contain the originals as their first draws, so they are supersets evaluated with the same five training seeds and trained models. The revision uses the enlarged sets everywhere in the main text and moves the original-set tables to Appendix A.14 (C3). The hard and box noise sweeps, previously run only on the original sets, are now run on the enlarged sets. U-value MAE (W/m²K), each cell noise s = 0 → s = 0.04, p of the readout against optimization at s = 0.04.

Corpus, geometriesReadout (ours)OptimizationTransolverp
Junction, 400.039 → 0.0810.047 → 0.1100.164 → 0.1756×10⁻⁵
Junction, 1200.036 → 0.0860.042 → 0.1110.183 → 0.1898×10⁻⁸
Hard, 320.029 → 0.0490.042 → 0.0850.047 → 0.0866×10⁻⁵
Hard, 960.024 → 0.0460.037 → 0.0880.037 → 0.0844×10⁻¹⁵
Box, 320.042 → 0.0760.054 → 0.1490.038 → 0.05210⁻⁶
Box, 960.033 → 0.0770.037 → 0.1350.029 → 0.0506×10⁻¹⁴

The enlarged sets sharpen every comparison and change no conclusion, including that on box the Transolver route reads U slightly better than the surface. With the prior-equipped baselines of the 7Zij thread, the readout's margin on hard is significant only on the 96-geometry set (p = 0.010 and 0.007).

[Q4: Conformal coverage under shift and with small sets]

Coverage is marginal under exchangeability for any estimator, and the paper reports repeated 50/50 splits of the validation geometries. The standard deviation of coverage across splits is 0.09 at 20 calibration geometries and 0.055 at 60, reported in Appendix A.13. To test the guarantee off its assumption we calibrated on one corpus and tested on another, which breaks exchangeability deliberately and needs no new training. Empirical coverage of the nominal 90 % interval over the six off-diagonal pairs.

EstimatorOwn corpusWorst pairBest pair
Readout (ours)0.90 to 0.920.800.98
Supervised head0.90 to 0.920.750.99
Optimization route0.90 to 0.920.521.00
Transolver route0.90 to 0.920.381.00

The worst case for the readout is calibrating on hard, the corpus with the narrowest spread of U, and testing on junction. Exchangeability failure costs every estimator its guarantee, the readout loses the least, and the field routes' intervals, whose width is set by their prior, collapse under shift (C8).

[Changes to the paper]

C1 abstract states the validation scope. C3 enlarged sets primary, original sets in Appendix A.14. C7 measurement-error appendix with the operating envelope and the r_se sensitivity, and the envelope pointer in the Discussion. C8 appendix on the calibration parameters, their non-transfer, the prior-ratio readout and cross-corpus coverage. The remaining changes are in the general response.

We hope the new evidence addresses your concerns. We would appreciate it if you could kindly reconsider your score.

I thank the authors for the detailed response. The clarification that the 120/96/96 sets are supersets generated from the same validation seed resolves my concern about inconsistent reporting (W3). However, my two other concerns are not resolved. First, the authors confirm that no dataset pairing wall infrared imagery with ISO 9869 heat-flux U-values is publicly available, so absolute U accuracy remains established only on the ISO 10211 synthetic corpus; the physical-interpretability argument for calibration transfer is a justification, not evidence, and the abstract's claim of "initial evidence for practical application" on real buildings should be accordingly downgraded. Second, the newly added noise study undermines rather than supports the central claim: under 15% and 30% spatially correlated noise the proposed estimator (0.231/0.340) is worse than Transolver (0.204/0.200), and cross-corpus conformal coverage drops to 0.52, which the authors themselves describe as loss of the guarantee. I therefore keep my score at 2.

Reviewer_7ZijOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies heat-loss diagnostics from building scans. The argument is that from a single exterior thermal observation the interior conductivity field is not identifiable, so scoring learned inverses by field recovery mostly scores their prior. Instead the paper targets what the boundary does determine: the section U-value and the lateral footprint of thermal bridges. U is read from a surface flux identity with a two parameter affine calibration fit on training geometries and wrapped with a split conformal interval, and the footprint comes from thresholding the deviation from a 1D analytic clear wall field. A geometry conditioned neural operator is trained as a forward surrogate but the headline U result does not use it.

Critical strengths
  • Proposed calibaration matches the regression head baseline
  • Negative results are reported and not hidden (Transolver winning on box, TBBR below chance), and the solver is ISO 10211 validated
  • The finding that regressing the functional value direclty outpeforms the indirect apporaches is vis consistent
Critical weaknesses
  • The pipeline seems to never be evaluated end-to-end Table 4 includes field predictions, while Tables 1 & 2 reported based on observation (or noisy obsevration)
  • The manuscript is difficult to read and not self contianed, too many discussions referrign to appendix. While Table 1 & 2 are underexplained
  • Even with the reprodcubilty statement its not clear how baselines were trained
  • The train/calibration slipts are not speicfied clearly
  • The comparison is asymmetric. Only the proposed routes are fit against U, the baselines are trained on fields and then scored on U, and it is not stated whether they get the analytic prior that the proposed pipeline is built on.
Questions for authors
  • What is the mean and std of U values in your datasets?
  • Do baselines get the analytical prior as input?
  • Is the recovery of field appraoach estiablited in literature or a baseline that you propose
  • Can you provide more details on the training of baselines? excat inputs output how the recovery is processed what targets are being used?
4 comments

Thank you for the careful review. We address each point below, in three comments. Every change in the revised PDF is typeset in blue.

[W1: Roles of the two passes]

The synthesis pass, the neural operator, is the differentiable forward model that the optimization route of Tables 1, 2 and 4 inverts and that the retrofit demonstration uses. The diagnostic pass reads U from the observation itself, so it does not inherit the surrogate's error, whereas any route through a recovered field inherits both the surrogate's and the inverse's (Table 4). The revision states under the roster which pass produces each table (C2). We also ran the end-to-end variant you ask for, replacing the anomaly localizer's 1-D analytic reference by the surrogate's predicted surface for the nominal construction. Footprint IoU (higher is better), five seeds, paper (1-D reference) → chained (surrogate reference), p from a paired Wilcoxon test over geometries.

Corpus (geometries)LabelsPaper → chainedp
Junction (120)threshold0.31 → 0.1410⁻¹⁹
Junction (120)true prism0.36 → 0.265×10⁻⁴
Hard (32, original set)threshold0.16 → 0.160.96
TUM2TWIN LoD2 (20)true prism0.27 → 0.082×10⁻⁶
TUM2TWIN LoD3 (20)true prism0.38 → 0.152×10⁻⁶

Chaining the passes makes localization worse because the anomaly then inherits the surrogate's error on the nominal surface. That is the measured reason the diagnostic pass reads from the observation, and the control enters Appendix A.11 (C6).

[W2: Material in the appendix]

Reading guides under Tables 1 and 2 state what each cell is. The roster moves to the main text (page 7), every number is tied to the component that produced it, and "route" is defined once and used consistently (C2, C3, C10).

[W3 & Q4: How every route is trained]

Every route sees the same exterior-slab observation, a band of 8 % of the wall depth holding about 326 of the 4096 cloud points (2048 on box). The roster is on page 7 and the recipe in the Reproducibility statement (C2, C4).

RouteInputOutputTraining targetLossPrior
Calibrated readout (ours)θ on the observed slabUa, b on training Uaffine, two parametersno
Supervised head (ablation)masked θ, coordinates, signed distanceUtrue Uheteroscedastic Gaussianno
Optimization routeobserved θlog k per pointnone, per instanceresidual + 5·10⁻³ sparsity + 10⁻³ proximity + 5·10⁻³ total variationthrough the surrogate
Amortized inverses (PointNet++, DeepONet, Transolver)[θ·mask, mask], coordinates, signed distancelog k per pointtrue log kmean squared errorno → yes in the new runs
Anomaly ruleobserved θ, θ_priorfootprintnonenoneyes
Clear walllayer tableUnonenoneyes

The amortized inverses use AdamW at 10⁻³ with a cosine schedule, 200 epochs, batch size 1, θ and mask set to 0 off the slab, and DeepONet and Transolver are re-instantiated as θ to log k regressors. The optimization route initialises log k at the clear-wall value and runs Adam at 0.05 for 400 steps (300 in the noise sweeps) through the frozen PointNet++ surrogate, and the ensemble averages five runs. U is read from every recovered field by re-simulating it through the surrogate and applying the same slab readout and calibration as the direct readout, so every U in Tables 1 and 2 is on one footing.

[W4: Splits]

CorpusTraining (seed 1337)Original validation (seed 99)Enlarged validation (seed 99, superset)
Junction15040120
Hard963296
Box963296

The enlarged sets contain the originals as their first draws. Calibration is fit on training geometries only, conformal intervals use repeated 50/50 splits of the validation geometries, no geometry is shared between training, calibration and test, and the five model seeds 1337, 1, 2, 3 and 4 are shared by every route (splits table, page 11, C4).

[Q1: Distribution of U]

U in W/m²K on the enlarged sets with two reference predictors. These columns enter Table 2 (C3).

Corpus (geometries)U, mean ± sdMAE, training meanMAE, clear wallMAE, readout (ours)
Junction (120)0.553 ± 0.2490.2060.2830.036
Hard (96)0.346 ± 0.0780.0660.0790.024
Box (96)0.325 ± 0.0920.0690.0590.033

On the original sets the means are 0.565 ± 0.243, 0.372 ± 0.083 and 0.341 ± 0.120.

[W5 & Q2: The asymmetry]

(a) Only the proposed routes are fit against U. The readout is fit to U only through a two-parameter affine map, fit once on training geometries and applied identically to every route, including every baseline's recovered field. The one estimator trained on U labels is the supervised head, and the readout matches it without labels. A parameter-free reading that scales the raw slab mean by the prior's own slab-to-face ratio reaches 0.033 on hard and box against the calibrated 0.024 and 0.033, so the advantage is not the two fitted constants.

(b) Baselines trained on fields and scored on U. That is the thesis, what the field the literature's inverses recover implies for the quantity a practitioner needs, under one formula for every route. The head, fit directly to U, is in every table and the readout matches it, so fitting the baselines to U would not change the ranking.

(c) The analytic prior. The amortized inverses did not receive it, so every one was re-trained with θ_prior as an extra input channel, on every corpus, five seeds. Field relative L2 error on junction falls from 0.77 to 0.24 (PointNet++), 0.68 to 0.23 (Transolver) and 0.80 to 0.50 (DeepONet), and the recovered volume becomes the prior's layers (volumetric IoU against the layer table 0.17 → 0.80 for PointNet++). Footprint IoU (higher is better), inverse columns without → with the prior, prism labels on junction and the shells, threshold labels on hard and box. The rule marks observed points only, so it is capped at the ceiling.

Corpus (geometries)PointNet++ (ours)TransolverDeepONetRule, fixedRule, calibratedCeiling
Junction (120)0.09 → 0.420.06 → 0.500.06 → 0.160.360.500.72
Hard (96)0.57 → 0.600.60 → 0.830.53 → 0.530.180.200.77
Box (96)0.54 → 0.570.68 → 0.760.63 → 0.660.13n/an/a
TUM2TWIN LoD2 (20)0.05 → 0.740.13 → 0.630.04 → 0.030.270.280.40
TUM2TWIN LoD3 (20)0.12 → 0.660.15 → 0.660.22 → 0.160.380.390.58

The field a learned inverse recovers is its prior (every junction prism spans the insulation layer, so volumetric IoU equals footprint IoU and the inverses report the prior's layer rather than depth). With the prior the inverses locate the footprint, matching the rule on junction (p = 0.06 for PointNet++) and exceeding it and its ceiling on the real shells (p = 2×10⁻⁶). The sentence in contribution 2 that learned field inverses do not locate the footprint overstated the case. It now reads that the rule locates the footprint with zero training where inverses without the prior do not, and that given the prior the twin's inverse matches it on junction and exceeds it on real geometry (C5).

[W5 & Q2, continued: The rule under the same calibration, and U once the baselines have the prior]

With threshold and sign set on training geometries and frozen, the rule reaches 0.50 on junction (p = 2×10⁻¹⁰ against the fixed rule). On the real shells calibration changes the rule by less than 0.015, so the gain is specific to prisms that reach into the slab, and we say so.

The U-value comparison does not change once the baselines have the prior. U-value MAE (W/m²K, lower is better), amortized columns without → with the prior, readout, optimization and head as in Table 2.

Corpus (geometries)Readout (ours)PointNet++TransolverDeepONetOptimizationHeadp, readout against PointNet++ / Transolver
Junction (120)0.0360.129 → 0.0840.175 → 0.0880.206 → 0.0820.0420.0446×10⁻¹⁵ / 3×10⁻¹⁶
Hard (96)0.0240.039 → 0.0300.036 → 0.0300.040 → 0.0400.0370.0250.010 / 0.007
Box (96)0.0330.045 → 0.0360.029 → 0.0310.028 → 0.0280.0370.0260.010 / 0.15 (DeepONet 4×10⁻⁴, better)

Reading U from the surface stands with the asymmetry removed. On box the DeepONet route reads U better and the head is best, as the paper reports.

[Q3: Field recovery routes]

Optimization through a differentiable forward model is PDE-constrained inversion (Raissi et al. 2019, Zhong et al., NeurIPS 2026). Amortized inversion is the neural inverse operator of Molinaro et al. (ICML 2023). The DeepONet and Transolver inverses are our re-instantiations of published forward architectures (Lu et al. 2021, Wu et al., ICML 2024) as inverse regressors, and the roster says so.

[Changes to the paper]

C2 roster with the pass per table. C3 reading guides and the complete Table 2 with the Q1 columns. C4 split table and training recipe. C5 contribution 2 and the footprint table rewritten with the rows above. C6 appendix with the prior-equipped runs and the end-to-end control. C10 route defined once. The remaining changes are in the general response.

We hope the new evidence addresses your concerns. We would appreciate it if you could kindly reconsider your score.

The revision does not change my assessment. The re-uploaded PDF remains unchecked: Figure 1 is a placeholder, Appendix A.10 is empty, and seven citations are unresolved, including DeepONet and FNO. More importantly, the chained system fails end-to-end, with junction IoU dropping from 0.31 to 0.14 and TUM2TWIN from 0.27 to 0.08. The headline U result instead relies on a two-parameter affine readout that bypasses the surrogate. With the same analytic prior, baselines match or outperform the proposed method: Transolver achieves 0.50 on junction and 0.74/0.66 on real shells, versus 0.50 and 0.27/0.38 for the calibrated rule. On box, Transolver outperforms the readout at every noise level, with the reported win rate falling to 24%. The footprint claim also relies on prism labels available only for junction, while the calibrated readout underperforms the supervised head on box (0.033 vs. 0.026) and fails to transfer across construction types. These issues remain unresolved, and my score stands.

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

This paper introduces a calibrated thermal twin that localizes thermal bridges and quantifies section-level heat loss from as-built geometry, material information, and exterior infrared measurements. Experiments on synthetic datasets show that the results are decent.

Critical strengths
  1. The experimental tables demonstrate promising results.
Critical weaknesses
  1. What is the actual title of the paper? There is a discrepancy between the title in the PDF and the title listed in the review system. For now, I will base my review on the title provided in the system.
  2. This addresses a very niche application-oriented problem (thermal bridges). It is unclear whether such a topic aligns with the scope of ICLR, and I question whether this represents a genuine, significant problem.
  3. The experiments appear to be conducted on a self-built synthetic dataset created by the authors. The lack of proper baselines severely undermines the persuasiveness of the evaluation.
  4. The manuscript lacks clarity in its presentation. The "Thermal Twins" approach seems to rely on geometry-based inference and related information, but the core methodological novelty remains unclear.
Questions for authors
  1. Please address the concerns and questions raised in the Weaknesses section above.
2 comments

Thank you for the review. We address each point below, in two comments. Every change in the revised PDF is typeset in blue.

[W1: Title]

We thank the reviewer for highlighting this inconsistency. Although we revised the title before the original submission deadline, the submission form inadvertently retained the earlier title and abstract. We apologise for this oversight and have now corrected both fields to align with the revised manuscript.

[W2: Scope and significance]

We appreciate the reviewer's concern regarding the scope and significance of this work. A growing body of research at ICLR and related venues has addressed closely connected problems in thermal modeling, heat transfer, physical property estimation, and the built environment, reflecting the importance of this research direction. For example, inverse heat conduction from thermography has been studied at NeurIPS 2026 (Neural Field Thermal Tomography, which we cite), while thermal 3D reconstruction has been explored at ICLR 2025 (ThermalGaussian) and ECCV 2024 (Thermal3D-GS, which also models heat conduction). The ThermoScenes facades used in our evaluation were introduced at ECCV 2024 (ThermoNeRF). Research on building energy has also appeared at NeurIPS through BuildingsBench (2023), a dataset covering 900,000 buildings, and the CityLearn competitions (2022 and 2023). Building reconstruction from scans has been studied at ICCV (BuildingNet 2021, Building3D 2023), while the Tackling Climate Change with Machine Learning workshops at NeurIPS, ICLR, and ICML have featured more than fifty papers on buildings, including facade retrofit assessment from scans (ICLR 2023) and infrared sensing of facades (NeurIPS 2024). From a methodological perspective, the neural operators evaluated in our work have also been developed and published at these venues, including FNO (ICLR 2021), GNOT and neural inverse operators (ICML 2023), GINO (NeurIPS 2023), and Transolver (ICML 2024). Collectively, these studies establish a substantial research foundation for learning-based thermal analysis and physical modeling. However, none directly addresses the task of converting building scans and infrared measurements into calibrated heat-loss estimates, which is the specific contribution of our work. The ICLR 2027 call for papers also explicitly recognizes applications to the physical sciences and sustainability.

The problem is large in cost. Buildings use 20 to 40 % of final energy in developed countries (Pérez-Lombard et al., 2008), thermal bridges raise the heating demand of an insulated building by up to 30 % (Theodosiou and Papadopoulos, 2008) and their share grows as walls improve, and two international standards, ISO 10211 and ISO 14683, govern their calculation. A bridge is still found by a thermographer with a camera, which shows where a surface is warm but not how much heat the section loses. The revised introduction gives these figures with their sources (C9).

[W3: Synthetic data]

A fair evaluation of this task needs four things together, the as-built geometry with its construction layer table, an exterior surface-temperature observation under known indoor and outdoor temperatures, ground truth for the interior field, the bridge footprint and the section U-value, and enough instances across construction types to train, calibrate and hold out geometries. No existing dataset provides more than two. TBBR has aerial thermal images with bridge boxes and no geometry or U-value, ThermoScenes has RGB and thermal views of facades with no interior ground truth, and in-situ heat-flux studies give U for single walls without a scan. The interior field cannot be measured without opening the wall and U needs a heat-flux plate on the section, so the learned-inverse literature we compare against also trains and tests on solver-generated fields.

We therefore built corpora that provide all four and checked them. The solver reproduces the two ISO 10211 reference cases to 0.039 K and 9.491 against 9.5 W/mK (Appendix A.2, page 13). The corpora are multilayer constructions from layer tables with ISO 6946 surface films, in three families of increasing difficulty, with 150/96/96 training and 120/96/96 held-out geometries from a separate seed, five model seeds shared by every route and no geometry shared between training, calibration and test (splits table, page 11). Real data are used wherever they exist. The twin runs on the real TUM2TWIN building shells at two levels of detail, and Appendix A.9 (page 16) checks the localizer on ThermoScenes facade infrared, airborne infrared over TUM2TWIN and TBBR, reporting the TBBR result as the negative result it is. What measured data cannot yet give is a U-value ground truth, and the abstract now says so (C1). Code, solver and corpora with their seeds are released on acceptance.

[W3: Baselines]

Table 2 (page 9) compares the readout with seven other estimators of U, three of them also in a prior-equipped variant, and Appendix A.1 adds three classical inversions. The seven are PDE-constrained optimization through the forward model (sparsity, proximity and total-variation regularisation, five-member ensemble), three amortized neural inverses (PointNet++, DeepONet, Transolver), a supervised regressor trained directly on U, the clear-wall value from the layer table and the training mean. Appendix A.1 adds Tikhonov, total-variation and Bayesian inversion on controlled walls. They cover every way the literature recovers U, through a field by optimization, through a field by a learned inverse, directly by regression and from the construction table without a measurement. Every route sees the same observation, seeds and splits, every U is read with one calibration fit on training geometries, and during the discussion the three learned inverses were re-trained with the analytic prior as an input so that they have every input our pipeline has. The submission showed three routes in Table 2 and left the rest to the appendix. The revision puts every route in Table 2 and lists each route's inputs, output, training target and loss in a roster on page 7 (C2, C3).

[W4: Clarity and novelty]

Tables 1 and 2 gain reading guides, the roster moves into the main text followed by which pass produces each table, and "route" is defined once and used consistently (C2, C3, C10).

The novelty is not an architecture but a system that did not exist. No prior method takes an as-built scan and an exterior infrared measurement and returns where the thermal bridges are and how much heat the section loses, with a distribution-free interval on that quantity. Three things had to come together. An identifiability argument that says what to predict, since a surface measurement does not determine the interior field but does determine the surface transmittance. A two-pass design built on it, a geometry-conditioned neural operator as the differentiable forward model and a diagnostic pass that reads U from a surface-flux identity and locates bridges from the surface anomaly or with the twin's prior-equipped inverse. Evidence that the choice of target matters, on three PDEs and on every corpus. The closest prior work does one part each. Infrared U-value estimation reads one-dimensional walls without geometry or localization (Albatici and Tonelli 2010, Tejedor et al. 2017). Learned bridge detectors locate without quantifying (Mayer et al. 2023). Bayesian wall inference reconstructs one-dimensional sections (De Simon et al. 2018). Neural Field Thermal Tomography inverts for the field and returns no interval (Zhong et al., NeurIPS 2026). Goal-oriented inference and task-driven conformal prediction (Lieberman and Willcox 2013, Wen et al., ECCV 2024) supply the abstract idea of a target functional, which we instantiate as a physically identifiable one. Section 2 states this (C9).

[Changes to the paper]

C1 abstract states the validation scope. C2 and C3 roster, pass per table, reading guides and the complete Table 2. C9 significance figures and the Section 2 paragraph. C10 route defined once. The remaining changes are in the general response.

We hope the new evidence addresses your concerns. We would appreciate it if you could kindly reconsider your score.

Comments on the paper

2 comments

We sincerely thank all reviewers for their careful evaluation and constructive feedback. Their comments have helped us identify important gaps in the original evaluation, clarify the scope of our contributions, and strengthen the experimental evidence. In response, we have conducted additional experiments to address the concerns raised, including ensuring fair comparisons by providing the learned baselines with the same analytical prior used by our method. The revised results continue to support our principal contribution while also revealing findings that warranted corrections to our original claims. We summarize the key results across two comments, with further details provided in the individual reviewer threads. All revisions in the updated PDF are highlighted in blue. Unless otherwise stated, the new experiments follow the paper's original evaluation protocol, using the same five random seeds, metrics, and calibration fitted exclusively on training geometries.

[Summary of revisions to the paper]

We have revised the manuscript to address the reviewers' concerns while retaining the 10-page limit. The changes include:

  • C1: Clarified the scope of U-value validation in the abstract and Section 6.
  • C2: Added a complete roster of methods in Section 5, identifying which methods contribute to each table.
  • C3: Improved the presentation of Tables 1 and 2 with reading guides, reported the enlarged evaluation sets as the primary results, and extended Table 2 to include every route. Results from the original sets are retained in the appendix.
  • C4: Added a table documenting the training, calibration, and test splits.
  • C5: Revised the second contribution and footprint localization results to account for prior-equipped inverses, the calibrated rule, and the observation ceiling.
  • C6: Added an appendix evaluating the analytical prior as an input to learned inverses, including an end-to-end control.
  • C7: Extended the measurement-error analysis to characterize the method's operating limits.
  • C8: Documented the calibration parameters and evaluated cross-corpus coverage.
  • C9: Strengthened the motivation in Section 1 with quantitative evidence of the problem's significance and clarified the novelty in Section 2.
  • C10: Improved terminology and readability by defining "route" consistently and introducing each quantity before its mathematical notation.

Dear Authors,

The revised PDF is not reviewable. Please take action promptly. Please treat this as a serious revision-quality issue.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.