Pith. sign in

REVIEW 2 major objections 1 minor 23 references

Thermal transmittance prediction based on the application of artificial neural networks on heat flux method results

T0 review · 2 major / 1 minor · reviewed 2026-05-25 · grok-4.3

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Artificial neural networks can predict heat flux through building walls from interior and exterior air temperatures.

desk verdict This is a straightforward empirical test of four off-the-shelf ANNs on real HFM temperature-flux data from two walls, with results that look usable on one but limited on the other, and no clear link from flux error to final U-value uncertainty. read the letter →

arxiv 2103.14995 v1 pith:K7A3XSE2 submitted 2021-03-27 cs.LG cs.AIeess.SP

Sanjin Gumbarević , Bojan Milovanović , Mergim Gaši , Marina Bagarić This is my paper · ORCID
classification cs.LGcs.AIeess.SP
keywords thermaltransmittanceU-valueheatfluxmethodartificialneuralnetworksbuildingenvelopein-situmeasurementtemperatureto
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tests whether neural networks trained on heat flux sensor data can accurately estimate heat flux rates when given only air temperature readings. If successful, this would let a single sensor collect initial data at one spot, then move elsewhere while the model supplies the missing flux values needed for U-value calculations. Four network types were trained and compared on measurements from one multi-layer wall. The temperature-to-flux predictions performed well enough on that wall to suggest shorter overall test times. A follow-up test on a second wall revealed limits that the authors flag for future work.

What carries the argument

Artificial neural networks (multilayer perceptron with three hidden neurons, LSTM with 100 units, GRU with 100 units, and a 50/50 LSTM-GRU hybrid) that map interior and exterior air temperatures to heat flux rate.

What would settle it

Compute U-values for the same wall segment once with the full direct heat-flux record and once with the network-predicted fluxes; the two U-values must agree within the tolerance allowed by the standard.

Watch

Extended reading notes

Core claim

Models trained on paired temperature and heat-flux records learn to output heat flux when supplied only the two air temperatures. Once trained, the network can stand in for direct sensor readings at new locations, so the physical heat-flux sensor need not remain in place for the full duration required by the ISO 9869-1 procedure.

Load-bearing premise

That the error introduced by the network's flux predictions stays small enough not to push the final calculated U-value outside acceptable limits.

Editorial extensions

If this is right

  • One heat-flux sensor can serve multiple measurement locations after an initial training interval.
  • Total time to obtain in-situ U-values for several building elements drops because sensors no longer stay fixed for the full standard duration.
  • Temperature data alone become sufficient to generate the flux time series required for transmittance calculations.
  • The approach directly supports the ISO 9869-1 heat-flux method by filling gaps when the sensor is relocated.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Retraining or fine-tuning on each new wall type may still be required before the method can be used without direct measurements.
  • Any systematic bias in the temperature-to-flux mapping would affect the accuracy of derived U-values for renovation planning.
  • Collecting temperature-flux pairs from a wider range of wall constructions could test whether a single model generalizes across building stock.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 1 minor

Summary. The paper proposes applying artificial neural networks (MLP with 3 hidden neurons, LSTM with 100 units, GRU with 100 units, and a 50/50 LSTM-GRU hybrid) to heat-flux-method (HFM) data to predict heat flux rates from interior/exterior air temperatures. The central claim is that satisfactory predictions on one multi-layer wall would allow relocation of the single heat-flux sensor to additional locations, thereby shortening total measurement time while still satisfying ISO 9869-1 requirements for in-situ U-value determination; results are described as promising on the training wall but subject to possible limitations on a second wall.

Significance. If the temperature-to-flux mapping can be shown to generalize across walls with flux errors that propagate to U-values remaining inside the 5–10 % tolerance band required by ISO 9869-1, the approach would materially reduce the practical barrier of long measurement durations and increase adoption of HFM for pre-renovation assessments. The work is entirely empirical and data-driven; no parameter-free derivations or machine-checked proofs are claimed.

major comments (2)
  1. [Abstract] Abstract: the statement that the analysis 'gave promising results' on one wall is not accompanied by any quantitative prediction metrics (MAE, RMSE, R²), cross-validation procedure, or error-propagation analysis from point-wise flux predictions into the time-averaged transmittance value; without these, the operational claim that the method permits reliable sensor relocation cannot be evaluated.
  2. [Abstract] Abstract / implied results section: the generalization assumption—that an ANN trained on temperature-flux pairs from one wall will yield heat-flux estimates accurate enough on a second wall for the integrated U-value to stay within ISO 9869-1 uncertainty—is flagged as having 'possible limitations' yet is not tested quantitatively; the manuscript must supply a direct comparison of ANN-derived versus measured U-values (or at minimum a Monte-Carlo propagation of observed flux residuals) on the second wall.
minor comments (1)
  1. [Abstract] Abstract: specify the exact training/validation split, optimizer, loss function, and early-stopping criteria used for each of the four ANN architectures.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on strengthening the quantitative support for our claims. We address each major comment below.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the statement that the analysis 'gave promising results' on one wall is not accompanied by any quantitative prediction metrics (MAE, RMSE, R²), cross-validation procedure, or error-propagation analysis from point-wise flux predictions into the time-averaged transmittance value; without these, the operational claim that the method permits reliable sensor relocation cannot be evaluated.

    Authors: We agree that the abstract should include these quantitative elements to allow evaluation of the sensor-relocation claim. The manuscript body reports MAE, RMSE and R² for the four architectures on the primary wall together with the cross-validation procedure; we will condense the key figures into the abstract and add a concise statement on propagation of point-wise flux residuals into the time-averaged U-value. revision: yes

  2. Referee: [Abstract] Abstract / implied results section: the generalization assumption—that an ANN trained on temperature-flux pairs from one wall will yield heat-flux estimates accurate enough on a second wall for the integrated U-value to stay within ISO 9869-1 uncertainty—is flagged as having 'possible limitations' yet is not tested quantitatively; the manuscript must supply a direct comparison of ANN-derived versus measured U-values (or at minimum a Monte-Carlo propagation of observed flux residuals) on the second wall.

    Authors: We acknowledge that the preliminary analysis on the second wall only indicated limitations without a quantitative U-value comparison. We will add a Monte-Carlo propagation of the observed flux residuals to assess whether the resulting U-value uncertainty remains inside the ISO 9869-1 5–10 % band, and will report the outcome directly in the revised manuscript. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: data-driven ANN training on measured pairs

full rationale

The paper trains standard ANN architectures (MLP, LSTM, GRU) on temperature-flux pairs collected via HFM on one wall and evaluates point-wise prediction error on held-out data and a second wall. No equations, fitted parameters, or self-citations are invoked to derive the claimed prediction; the mapping is learned empirically from external sensor data. The central claim therefore remains a statistical generalization task rather than a reduction to its own inputs by construction.

Assumptions & free parameters 0 free parameters · 0 assumptions · 0 invented entities

Abstract-only review supplies no explicit free parameters, axioms, or invented entities; standard ANN training assumptions (i.i.d. data, sufficient training epochs, etc.) are implicit but unstated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Thermal transmittance prediction based on the application of artificial neural networks on heat flux method results." pith.science (2026). https://pith.science/paper/K7A3XSE2

@misc{pith2026210314995,
  author       = {Pith},
  title        = {Pith review of: Thermal transmittance prediction based on the application of artificial neural networks on heat flux method results},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/K7A3XSE2}},
  note         = {Machine review of arXiv:2103.14995}
}
read the original abstract

Deep energy renovation of building stock came more into focus in the European Union due to energy efficiency related directives. Many buildings that must undergo deep energy renovation are old and may lack design/renovation documentation, or possible degradation of materials might have occurred in building elements over time. Thermal transmittance (i.e. U-value) is one of the most important parameters for determining the transmission heat losses through building envelope elements. It depends on the thickness and thermal properties of all the materials that form a building element. In-situ U-value can be determined by ISO 9869-1 standard (Heat Flux Method - HFM). Still, measurement duration is one of the reasons why HFM is not widely used in field testing before the renovation design process commences. This paper analyzes the possibility of reducing the measurement time by conducting parallel measurements with one heat-flux sensor. This parallelization could be achieved by applying a specific class of the Artificial Neural Network (ANN) on HFM results to predict unknown heat flux based on collected interior and exterior air temperatures. After the satisfying prediction is achieved, HFM sensor can be relocated to another measuring location. Paper shows a comparison of four ANN cases applied to HFM results for a measurement held on one multi-layer wall - multilayer perceptron with three neurons in one hidden layer, long short-term memory with 100 units, gated recurrent unit with 100 units and combination of 50 long short-term memory units and 50 gated recurrent units. The analysis gave promising results in term of predicting the heat flux rate based on the two input temperatures. Additional analysis on another wall showed possible limitations of the method that serves as a direction for further research on this topic.

Figures

Figures reproduced from arXiv: 2103.14995 by the authors.

Figure 1
Figure 1. Schematic cross section through heat flux meter. 2.2. Artificial neural networks Area of artificial intelligence and machine learning (ML) started to be used significantly in the last decade due to development of faster computer machines and engagement of new learning algorithms [20]. Large part of ML models depend on the concept of ANN. ANN is briefly a systematical set of weighting coefficients, biases and activat… view at source ↗
Figure 2
Figure 2. Multilayer perceptron class of artificial neural network with two inputs, three neurons in hidden layer, and one output. W is matrix of weighting coefficients and biases. The empirical loss (5) can than be defined as total loss over the entire dataset. This is performed iteratively using the gradient method (often optimized) by computing the gradient ∂J(W) ∂W and by updating the weights matrix W until convergence is… view at source ↗
Figure 3
Figure 3. Schematic view of long-short term memory unit with ReLU activation function. GRU has one gate less than LSTM (it does not have forget gate) and schematic view of it can be seen in figure 4. It consists of reset gate rt , update gate zt and candidate activation gate eht . Although they look slightly different, RNN follow the same principles described for MLP [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Schematic view of gated recurrent unit with ReLU activation function. 2.3. Research design HFM experiment on a wall with unknown U-value was carried out in the period between 22 February 2019 and 26 February 2019 (figure 5). Experimental data that is analyzed consists …
Figure 5
Figure 5. Figure 5: Experimental setup for HFM data acquisition for analyzed wall. Four types of ANN architectures were analyzed: MLP3 – multilayer perceptron with 3 neurons in hidden layer, LSTM100 – architecture with 100 LSTM cells in hidden layer, GRU100 – architecture with 100 GRU cel…
Figure 6
Figure 6. Figure 6: Results for training/validation ratio 1/4 [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Predicted vs. actual heat flux for MLP3 for training/validation ratio 1/4 [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Results for training/validation ratio 1/2 [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]
Figure 9
Figure 9. Figure 9: Results for training/validation ratio 2/3. 4. Discussion and conclusion The results are promising in terms of prediction of heat flux based on measured heat flux with two input temperatures in training period, and after the training period by measuring just input tempe…
Figure 10
Figure 10. Figure 10: Analysis on additional wall with red-mark describing possible cause of non-stable prediction. The most noticeable limitation of the method could be an influence by a change in the [PITH_FULL_IMAGE:figures/full_fig_p007_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

23 extracted references · 23 canonical work pages

  1. [1]

    Laaroussi Y, Bahrar M, Zavrl E, Mankibi M E and Stritih U 2020 Sustainable Cities and Society 63 102452

    work page 2020

  2. [2]

    Jensen P A, Maslesa E, Berg J B and Thuesen C 2018 Building and Environment 143 130–137

    work page 2018

  3. [3]

    Bienvenido-Huertas D, Moyano J, Mar´ ın D and Fresco-Contreras R 2019Renewable and Sustainable Energy Reviews 102 356–371

    work page

  4. [4]

    Ficco G, Iannetta F, Ianniello E, d’Ambrosio Alfano F R and Dell’Isola M 2015 Energy and Buildings 104 108–121

    work page 2015

  5. [5]

    Gomes M G, Flores-Colen I, Manga L, Soares A and de Brito J 2017 Construction and Building Materials 135 279–286

    work page 2017

  6. [6]

    Lucchi E 2017 Energy and Buildings 151 393–405

    work page 2017

  7. [7]

    Building elements

    International Organization of Standardization ISO 9869-1:2014 – Thermal insulation. Building elements. In-situ measurement of thermal resistance and thermal transmittance. Heat flow meter method Tech. rep

    work page 2014

  8. [8]

    Soares N, Martins C, Gon¸ calves M, Santos P, da Silva L S and Costa J J 2019 Energy and Buildings 182 88–110

    work page 2019

Show all 23 references
  1. [9]

    Bienvenido-Huertas D 2020 Sustainability 12 6247

    2020

  2. [10]

    Tejedor B, Gaspar K, Casals M and Gangolells M 2020 Applied Sciences 10 8337

    2020

  3. [11]

    Evangelisti L, Guattari C and Asdrubali F 2019 Energy and Buildings 203 109441

    2019

  4. [12]

    Rezvani F, and Bribi´ an I Z 2019Revista de la construcci´ on18 311–322

  5. [13]

    Meng X, Luo T, Gao Y, Zhang L, Shen Q and Long E 2017 Applied Thermal Engineering 122 747–757

    2017

  6. [14]

    Roque E, Vicente R, Almeida R M, da Silva J M and Ferreira A V 2020 Applied Thermal Engineering 169 114935

    2020

  7. [15]

    Gaˇ si M, Milovanovi´ c B and Gumbarevi´ c S 2019Buildings 9 132

  8. [16]

    Choi D S and Ko M J 2017 Energies 10 1019

    2017

  9. [17]

    Gaspar K, Casals M and Gangolells M 2018 Energy and Buildings 176 360–370

    2018

  10. [18]

    M´ arquez J M A, Boh´ orquez M A M and Melgar S G 2017Sensors 17 2017

    2017

  11. [19]

    Gumbarevi´ c S, Milovanovi´ c B, Gaˇ si M and Bagari´ c M 2020Proceedings of 7th International Electronic Conference on Sensors and Applications (MDPI)

  12. [20]

    Jordan M I and Mitchell T M 2015 Science 349 255–260

    2015

  13. [21]

    License: Creative Commons BY-NC-SA

    Amini A and Soleimany A 2021 Massachusetts Institute of Technology: MIT OpenCourseWare, https://ocw.mit.edu. License: Creative Commons BY-NC-SA

    2021

  14. [22]

    Hochreiter S and Schmidhuber J 1997 Neural Computation 9 1735–1780

    1997

  15. [23]

    Cho K, van Merrienboer B, Gulcehre C, Bahdanau D, Bougares F, Schwenk H and Bengio Y 2014 ( Preprint 1406.1078)

    2014 arXiv

Pith tools

Reviewed May 25, 2026 · model on record in the stance chip above.