{"id":"f060fa34-c8a4-4c9b-b3a4-63a5a68195e9","arxiv_id":"2508.20633","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid deep-learning and Gappy POD reduced-order model reconstructs refrigerator temperature fields from sparse sensors with reported error under 1°C and 5,000x speed-up.","lead":"This paper combines neural networks with proper orthogonal decomposition to reconstruct the full temperature field inside a domestic refrigerator from a few sensor readings. The authors report under one degree Celsius error and a 5,000 times speed-up over full CFD simulation.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Validation leakage in POD basis: SVD is computed from all 88 configurations, so the leave-one-out target field lies in the basis and the claimed sparse-data accuracy is not yet established.","rationale":"The reader's weakest-assumption analysis identifies the same core issue: the POD basis is constructed from the full 88-configuration dataset before leave-one-out validation, and the paper never demonstrates that a training-only basis is used. This concern is load-bearing because the central claim is about predicting the full temperature field at an unseen parameter value from sparse sensors. If the validation full field is already in the basis, the gappy reconstruction is effectively recall, not prediction, and the reported errors are artificially low. The concern is concrete and testable: rerunning the sparse-training experiments with a fresh SVD on training data only would settle it. I agree with the reader's rejection as the appropriate posture: the FOM validation and the proposed loss formulation are useful, but the ROM evidence as presented does not support the headline accuracy claim. I would not change the verdict; the paper should be revised to perform strict training-only basis construction and report the resulting errors before any acceptance can be considered.","tokens_in":21221,"tokens_out":5303,"duration_ms":60730,"concrete_test":"Re-run the leave-one-out experiments with a fresh POD basis computed only from the training snapshots. For the most diagnostic case, take the 13% training / 5-sensor ANNGPOD experiment (Fig. 19): build the truncated basis using SVD of only the 12 training temperature fields, keep the same energy threshold (99.7%), then compute the MAE over all validation planes with the same network and sensor data. Compare this MAE to the reported Fig. 19 values. If the MAE rises above the claimed 1 °C limit, or if ANNGPOD no longer consistently beats conventional GPOD, the sparse-data accuracy is an artifact of the validation field being present in the basis. The same training-only SVD should also be applied to the POD-RBF experiments in Fig. 13 to see whether the <1 °C figure survives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central accuracy claim rests on a possible data leak in the POD basis construction. Section 5.1 states explicitly, \"Singular Value Decomposition (SVD) is first applied to the dataset containing 88 parametric locations to compute the POD modes.\" The subsequent validation procedures then leave out one parameter value at a time for POD-RBF, GPOD, and ANNGPOD, but no statement anywhere says that a fresh, training-only SVD is recomputed for each split. If the same 88-snapshot SVD is used, the validation configuration's full temperature field is contained in the POD subspace (to truncation accuracy). For GPOD and ANNGPOD, this is not a benign detail: when the target field is in the basis and the masked sensor system has full column rank, the least-squares / GPOD loss can recover near-exact modal coefficients from sparse sensor readings. The reported MAEs then measure how well the basis represents a known snapshot, not how well the method generalizes to an unseen parameter value. The ANN's GPOD loss term can exploit the same in-basis target to drive the reconstruction error toward zero. Thus the central '1 degree Celsius even at very sparse training data' claim and the ANNGPOD-vs-GPOD comparison may be inflated. The FOM validation against experiments is genuine supporting evidence, but it validates the CFD model, not the ROM generalization claim. The paper also omits code, data, and hyperparameters, but the basis-leakage issue is the load-bearing technical concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a non-intrusive reduced-order model (ROM) for conjugate heat transfer in a domestic refrigerator. A full-order CHT RANS model is validated against experiments, and an 88-case parametric database is generated over evaporator temperature, ambient temperature, and fan speed. The authors compare three ROM approaches: POD-RBF interpolation, conventional Gappy POD (GPOD), and a proposed ANNGPOD method that augments a POD-ANN regression with a GPOD loss term based on sparse sensor data. The central claim is a prediction error below 1°C and a speed-up of about 5e3 even with very sparse training data, including a demonstration using five experimental sensors as input.","tokens_in":21602,"tokens_out":11043,"duration_ms":117859,"significance":"If the accuracy claims are reproducible under a proper train/validation split, the ANNGPOD idea is an interesting industrial ROM contribution: it combines a parameter-to-coefficient neural network with sensor-based feedback, and the paper includes a genuine full-order CHT validation and a real experimental sensor test. The paper is also transparent about some limitations, such as non-monotonic training-set behavior and the need to rerun the algorithm for each new parameter value. However, the central accuracy claim currently rests on a validation protocol that includes validation snapshots in the POD basis construction, and the GPOD baseline is handicapped by a mode-count restriction; both issues must be corrected before the claimed superiority is established.","major_comments":[{"comment":"The POD basis is computed once from all 88 parametric configurations: \"Singular Value Decomposition (SVD) is first applied to the dataset containing 88 parametric locations to compute the POD modes.\" The validation points are then taken from these same 88 cases, and no statement indicates that a fresh, training-only SVD is recomputed for each split. For GPOD (Eqs. 28–29) and ANNGPOD (Eqs. 30–32), the sparse-sensor data are processed with a basis that has already been trained on the very configuration being reconstructed; at the 99.7% energy truncation used for ANNGPOD the target field is essentially a training member, so the reconstruction measures how well the method fits a known snapshot, not how it generalizes to an unseen parameter. This undermines the central '1°C at very sparse data' claim and the ANNGPOD-vs-GPOD comparison. The same issue affects the experimental test in §5.4, sin","section":"§5.1, Figs. 13–19, Table 2"},{"comment":"The GPOD baseline is restricted to \"the number of modes equal to the number of sensor locations\" in order to avoid the pseudo-inverse, whereas ANNGPOD is allowed 99.7% energy, corresponding to about 25 modes. In Fig. 16, GPOD_5sensors thus uses 5 modes while ANNGPOD_5sensors uses 25 modes. The reported improvement of ANNGPOD over GPOD is therefore confounded with the number of retained modes. A fair comparison should let GPOD use the same number of modes (via the pseudo-inverse or a regularized least-squares solve), or should restrict ANNGPOD to the same number of modes as GPOD.","section":"§5.2–5.3, Fig. 16"},{"comment":"The loss weights λ1 and λ2 in Eq. (32) are not assigned numerical values anywhere, and Algorithm 1 (line 8) writes the total loss as L ← (LPOD-ANN + LGPOD), omitting them. Since these weights determine the balance between the POD-ANN regression loss and the sparse-sensor GPOD loss, and since the text itself suggests they can be tuned to improve accuracy, the reported ANNGPOD results are not reproducible without these values. Please report the values and a sensitivity study.","section":"Eq. (32) and Algorithm 1"},{"comment":"The authors state that \"with the 48% data, the ANNGPOD method has shown poor performance than the 13% and 30% training data at most of the planes\" for both 5 and 26 sensors. This non-monotonic behavior is difficult to reconcile with the abstract's claim of a \"stable\" method and with the introduction's claim of accurate prediction \"irrespective of the number or placement of sensor points.\" At minimum, the robustness claim must be qualified; ideally, an explanation of why a larger training set degrades performance should be provided.","section":"§5.3, Fig. 19"}],"minor_comments":[{"comment":"The training proportion is given as 13% and 98% in the text but as 11% in the table caption and in the concluding remarks; please harmonize.","section":"Table 2 and §5.4"},{"comment":"The dimension of C is stated as R^{N_r^T × N_dof}; it should be R^{N_r^T × M}.","section":"Eq. (17)"},{"comment":"Line 8 should explicitly include λ1 and λ2; as written it also omits the subscript on the GPOD loss.","section":"Algorithm 1"},{"comment":"The POD-ANN method of Section 3.2 is never evaluated alone. A POD-ANN-only result would help isolate the contribution of the GPOD loss term in ANNGPOD.","section":"§3.2 / §5.3"},{"comment":"The advertised 'speed-up of 5e3' for ANNGPOD is per query, because the network is retrained for each new parameter value. This should be stated in the abstract/conclusions where the speed-up is advertised.","section":"§5.5"},{"comment":"There are several typos, e.g., 'Form Fig. 16b' (should be 'From'), 'M AEis' (missing space), and the abstract's '1 ( oC)'.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's full-order validation and the use of real experimental sensor data are commendable, and the all-88 SVD is stated openly rather than hidden. My main concern is methodological: the validation protocol must be repaired before the accuracy claims can be assessed. I would be willing to review a revision with training-only SVD recomputation and a fair GPOD baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take on the refrigerator ROM paper. The new thing is the ANNGPOD loss: a POD-ANN training loss plus a gappy-POD sensor mismatch term. That's a legitimate and useful extension, and the paper demonstrates it on a real domestic fridge with a full CHT model that is validated against experiments in two configurations. The sensitivity analysis and the computational cost reporting are also done properly. So there is real work here.\n\nThe problem is the validation. Section 5.1 states that \"SVD is first applied to the dataset containing 88 parametric locations to compute the POD modes.\" Then the ROMs are tested at leave-one-out validation points without ever recomputing the basis on training-only snapshots. For POD-RBF that is already a concern; for GPOD and ANNGPOD it is load-bearing. When the target field lives in the POD subspace, the sparse sensor reconstruction can recover modal coefficients well enough to make the MAE look far better than it would for a genuinely unseen parameter value. The headline claim of \"one degree centigrade even at very sparse training data\" is exactly the claim this leakage inflates.\n\nThere are smaller issues: the loss weights lambda_1 and lambda_2 are never reported, the RBF width is not given, and no code or data are provided. The authors' own Figure 19 shows ANNGPOD at 48% training data doing worse than at 13% and 30%, which undercuts the \"stable\" label.\n\nI'd still send this to peer review: the method is worth discussing, and the FOM validation is genuine. But the fixed basis is a serious flaw that needs to be addressed. A referee should ask for a training-only SVD and a re-run of the comparisons. As it stands, I would not cite it for the accuracy claims.","headline":"ANNGPOD is a plausible incremental extension, but the POD basis built from all 88 configurations leaks the validation targets into the representation space, so the sparse-data accuracy claims are not yet established.","tokens_in":22078,"tokens_out":2951,"would_cite":false,"duration_ms":30958,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper introduces ANNGPOD, a deep-learning-enhanced Gappy Proper Orthogonal Decomposition method that reconstructs the full refrigerator temperature field from a few sensor measurements, claiming under 1°C error and a 5000x speed-up even","keywords":["Conjugate Heat Transfer","Reduced Order Model","Gappy Proper Orthogonal Decomposition","Artificial Neural Network","POD-ANN","Temperature field reconstruction","Sparse sensor data","Domestic refrigerator"],"falsifier":"Recompute the POD basis using only the training parameter snapshots (excluding the validation configuration), then repeat the leave-one-out temperature reconstruction with ANNGPOD and classical GPOD; if the mean absolute errors still stay below 1°C on all 11 planes and for 5 to 26 sensors, the claim survives, but if errors degrade substantially the reported accuracy is an artifact of basis leakage.","tokens_in":21103,"feed_emoji":"🧊","tokens_out":5267,"duration_ms":59487,"temperature":0.7,"pith_summary":"The paper tries to establish that adding a Gappy-POD sensor-mismatch term to the loss function of a POD-ANN neural network produces a stable non-intrusive reduced-order model for temperature-field reconstruction. On a 20-million-cell conjugate heat transfer model of a domestic refrigerator, the authors report mean absolute errors below 1°C on 11 reconstruction planes and a computational speed-up of about 5000 over the full-order simulation, even when the training snapshot set is only 13% of 88 parametric configurations and the sensor count is as low as 5. The method is designed to exploit both high-fidelity numerical data and sparse experimental data, which the data-only POD-RBF surrogate cannot do. The paper also states that ANNGPOD is not real-time, requiring about 193 seconds per reconstruction, so the speed-up claim is relative to the full-order simulation cost.","feed_headline":"A neural Gappy POD rebuilds fridge temperatures to one degree","feed_subtitle":"Hybrid reduced-order model stays accurate with as few as 5 sensors, beating classical Gappy POD and running 5000x faster than CFD.","key_machinery":"The key object is the ANNGPOD total loss function, Eq. (32): L = λ1 times the POD-ANN mean-squared error over training modal coefficients plus λ2 times a Gappy-POD residual that evaluates, at the target parameter µ*, the squared difference between the sparse sensor measurements and the temperature field reconstructed from the network's predicted modal coefficients. This second term injects pointwise physical data into the neural-network training, which removes the classical GPOD constraint that the number of modes must equal the number of sensors and is the mechanism the paper credits for the method's stability.","core_discovery":"ANNGPOD trains a feedforward neural network to map the three operating parameters (evaporator temperature, ambient temperature, fan velocity) to POD modal coefficients, but the loss function combines the standard POD-ANN mean squared error with an additional term measuring the discrepancy between the reconstructed field and sparse sensor readings at the target parameter value. The Gappy-POD term is computed every epoch and included in backpropagation, so the network is simultaneously fitted to the numerical solution manifold and to the actual sensor data. The paper claims this makes the reconstruction stable with respect to sensor number and placement: unlike classical GPOD, whose accuracy d","pith_inferences":["The reported tests build the POD basis from the full 88-configuration snapshot set before forming training/validation splits and never state that a fresh training-only SVD is recomputed; therefore a fair comparison that excludes the validation field from the basis could change the ranking between ANNGPOD and POD-RBF at sparse training percentages.","The loss-coupling recipe is physics-agnostic and transferable: any parameterized PDE where pointwise sensor data are available could use the same two-term loss to stabilize field reconstruction, not just heat-transfer problems.","The non-monotone behavior with training-set size—where 48% data sometimes performs worse than 30%—suggests the λ1/λ2 weighting and network hyperparameters are not tuned per data budget; adaptive weighting could restore monotone improvement and is a natural next step.","A testable extension is to use ANNGPOD's reconstructed field as a prior for data assimilation or as a state estimator in model-predictive control of refrigeration systems, where its 193-second inference is acceptable in an offline or periodically updated setting."],"forward_implications":["If the reported accuracy holds, refrigerator design and control studies can replace expensive full CFD sweeps with a surrogate that corrects itself using a few thermocouple readings at new operating conditions.","The method can jointly use numerical and experimental data: the demonstrated case with five experimental sensors pulls the reconstruction toward measurements and reduces ROM-vs-experiment error by about 0.44°C compared with POD-RBF.","The speed-up is offline-to-online: a full 20-million-cell RANS-CHT simulation costs about 10^6 CPU-seconds, while an ANNGPOD reconstruction costs about 193 seconds on a single core.","Classical GPOD requires the number of modes to equal the number of sensors to avoid pseudo-inverse errors; ANNGPOD decouples truncation energy from sensor availability, which is why the paper reports stable accuracy across sensor counts of 5 to 26.","The paper explicitly notes ANNGPOD is not suited to real-time prediction because each new parameter query requires a 193-second network pass, whereas the POD-RBF method predicts in 0.64 seconds."],"supporting_citations":[{"why":"Introduces the Karhunen–Loève (POD) procedure for gappy data, which is the classic GPOD formulation that ANNGPOD's second loss term extends.","marker":"[17]"},{"why":"Establishes unsteady flow sensing and estimation via gappy POD with sensor data, the baseline framework the paper compares against and must outperform.","marker":"[54]"},{"why":"Demonstrates aerodynamic data reconstruction with POD, a key earlier application that motivates the use of gappy POD for full-field reconstruction from partial data.","marker":"[9]"},{"why":"Compares gappy data reconstruction approaches and kriging, providing context for the sensor-based interpolation problem that ANNGPOD addresses.","marker":"[22]"},{"why":"Presents a conjugate heat transfer analysis of a simplified household refrigerator, supplying the CHT modeling framework used to generate the full-order temperature database.","marker":"[58]"}],"fun_headline_variants":["Neural Gappy POD hits 1°C accuracy, 5000x speedup","Fridge temperature reconstruction: 1°C error, 5000x faster","Deep-learning Gappy POD: sparse sensors, full field, 5000x speed","Neural-augmented Gappy POD: 1°C accuracy from limited sensors"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The accuracy tests build the POD basis from the full 88-configuration snapshot set before splitting into training and validation, so the target validation field may already be present in the basis; the paper never states that a fresh training-only SVD is recomputed.","fun_headline_variants_meta":{"raw":{"variants":["Neural Gappy POD hits 1°C accuracy, 5000x speedup","Fridge temperature reconstruction: 1°C error, 5000x faster","Deep-learning Gappy POD: sparse sensors, full field, 5000x speed","Neural-augmented Gappy POD: 1°C accuracy from limited sensors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000444,"raw_usage":{"total_tokens":2083,"prompt_tokens":742,"completion_tokens":1341,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":486,"completion_tokens_details":{"reasoning_tokens":1250}},"tokens_in":486,"tokens_out":1341,"duration_ms":10751,"temperature":1.0,"reasoning_tokens":1250,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T14:56:49.627749+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the POD basis using only the training parameter snapshots (excluding the validation configuration), then repeat the leave-one-out temperature reconstruction with ANNGPOD and classical GPOD; if the mean absolute errors still stay below 1°C on all 11 planes and for 5 to 26 sensors, the claim survives, but if errors degrade substantially the reported accuracy is an artifact of basis leakage.","supporting_citations":[{"cited_title":"Everson and L","cited_arxiv_id":null,"evidence_quote":"Introduces the Karhunen–Loève (POD) procedure for gappy data, which is the classic GPOD formulation that ANNGPOD's second loss term extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes unsteady flow sensing and estimation via gappy POD with sensor data, the baseline framework the paper compares against and must outperform."},{"cited_title":"Bui-Thanh, M","cited_arxiv_id":null,"evidence_quote":"Demonstrates aerodynamic data reconstruction with POD, a key earlier application that motivates the use of gappy POD for full-field reconstruction from partial data."},{"cited_title":"Gunes, S","cited_arxiv_id":null,"evidence_quote":"Compares gappy data reconstruction approaches and kriging, providing context for the sensor-based interpolation problem that ANNGPOD addresses."},{"cited_title":"Zhang and Y","cited_arxiv_id":null,"evidence_quote":"Presents a conjugate heat transfer analysis of a simplified household refrigerator, supplying the CHT modeling framework used to generate the full-order temperature database."}],"review_version":1}