{"id":"46adf1f8-43c7-4241-b0d3-2db53549ac1a","arxiv_id":"2506.18247","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Integrating Bayesian neural networks into differentiable hybrid PIML architectures yields uncertainty estimates with accuracy slightly worse than or equal to deterministic baselines.","lead":"This paper replaces the neural networks inside hybrid physics-informed machine learning models with Bayesian neural networks to estimate prediction uncertainty, then tests the idea on a math benchmark and on real flight data for a small aircraft. The authors find the Bayesian versions are about as accurate as the original models, and that sampling the weights directly is the most reliable way to propagate uncertainty.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's efficient Taylor uncertainty propagation (Eq. 3) fails on its own benchmark (Fig. 7) and is then replaced by end-to-end Monte Carlo without any cost/calibration comparison, leaving the central 'efficient quantification' claim unsubstantiated.","rationale":"The reader's weakest-assumption analysis identifies Eq. (3) as the load-bearing point, and the paper's own figures support that concern. In Fig. 7(b), the Taylor-propagated bounds are not merely slightly over- or under-covering; they deviate dramatically from the true function, and the authors then switch to end-to-end Monte Carlo for the results shown in Fig. 8. This is an internal inconsistency in the efficiency narrative: the method advertised as efficient is demonstrated to be inaccurate on the simplest case, while the method that works is the one the paper says becomes impractical for complex problems. Because the abstract and title emphasize 'efficient quantification,' the absence of any runtime or coverage comparison between Taylor and MC is a central evidentiary gap, not a minor omission. The paper is honest about reporting the degraded Taylor performance, and the accuracy claim may survive with additional evidence, so REJECT is too strong. CONDITIONAL remains appropriate; the authors should either validate Taylor in settings where it works or reframe the contribution around MC-based UQ with cost measurements. The post-hoc selection of 10 hidden units for the MC comparison is a secondary concern that also needs addressing, but the Taylor failure is the more fundamental issue because it directly undermines the claimed efficiency contribution.","tokens_in":12454,"tokens_out":3890,"duration_ms":44996,"concrete_test":"On the Gramacy-Lee benchmark, fix the trained 10-hidden-unit BNN transfer network and propagate the same 20 weight samples through the partial physics model two ways: (a) Taylor expansion of Eq. 3 using the Jacobian at the mean transfer parameters, and (b) full end-to-end Monte Carlo. Report coverage (fraction of 100 test points inside the 68% bound), mean interval width, and wall-clock propagation time for both. If Taylor coverage is far below 68% or its runtime advantage over MC is not measured, the efficient-quantification claim fails; also repeat the Taylor variant on the fixed-wing aircraft case to test generalizability.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central efficiency claim rests on Section II.C Eq. (3): a first-order Taylor expansion propagates transfer-parameter uncertainties through the partial physics model, replacing Monte Carlo and thereby improving computational efficiency. The paper's own analytical results contradict this assumption. Fig. 7(b) shows the 68% confidence bounds from Taylor propagation deviating dramatically from the true Gramacy-Lee function, and the text states that Monte Carlo simulations were employed 'to address this.' The final comparison (Fig. 8) therefore uses end-to-end MC with a 10-hidden-unit transfer network, not the proposed Taylor method. No wall-clock time, sample count, or coverage metric is reported for either propagation strategy, so 'efficient' is asserted rather than demonstrated. Since the Taylor approximation demonstrably fails on a smooth one-dimensional benchmark with a known transfer map, its adequacy for the nonlinear VLM/propeller aircraft model is unsupported; the aircraft case reports only MC-based uncertainty plots (Figs. 10-11) and no Taylor variant. Without a quantitative accuracy-versus-cost comparison, the paper does not establish that BNN-integrated PIML architectures provide efficient uncertainty quantification. The accuracy claim is also entangled with the post-hoc choice of 10 hidden units for the MC models, but the primary gap is the missing validation of Eq. 3.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes integrating Bayesian neural networks (BNNs) into differentiable hybrid physics-informed machine learning (PIML) architectures by replacing the ANN component with a BNN whose final layer is Bayesian, trained via a two-stage procedure (deterministic pretraining followed by an MSE plus ELBO objective). Uncertainty in the BNN transfer parameters is propagated through the partial physics model by either a first-order Taylor expansion (Eq. 3) or by end-to-end Monte Carlo sampling. The approach is evaluated on a multimodal Gramacy & Lee benchmark and on flight-test data from a fixed-wing RC aircraft, with prediction accuracy compared against ANN, BNN, PIML-ANN, and partial-physics baselines.","tokens_in":12683,"tokens_out":5017,"duration_ms":47866,"significance":"If the central efficiency claim were substantiated, the work would offer a practical recipe for adding uncertainty quantification to hybrid PIML surrogates without a large accuracy penalty. The two-stage training and last-layer Bayesian treatment are pragmatic and address known BNN training difficulties. However, the paper's own results show the proposed Taylor propagation producing poor confidence bounds on the analytical benchmark, and no quantitative reliability or runtime comparison is provided; the central contribution is therefore not yet demonstrated. The configuration tables and convergence plots are useful for reproducibility, but the evidence currently supports only a qualitative exploration, not a validated efficient-UQ method.","major_comments":[{"comment":"The central efficiency claim rests on the first-order Taylor expansion in Eq. (3), but Fig. 7(b) shows that on the Gramacy & Lee benchmark the resulting 68% confidence bounds deviate dramatically from the true function, and the text then replaces this approach with end-to-end Monte Carlo simulations (Fig. 8). No coverage statistic, interval score, or wall-clock comparison is reported for the two propagation strategies. As a result, the paper does not establish that the Taylor approach is an efficient or adequate way to propagate uncertainty in hybrid PIML models; please add a quantitative accuracy-versus-cost comparison for both propagation methods on both case studies.","section":"II.C, Eq. (3), Figs. 7–8"},{"comment":"The accuracy comparison in Fig. 6 and Table I uses PIML-BNN and BNN architectures with 200 nodes per layer, whereas the uncertainty results in Fig. 8 are obtained with 10 hidden units and 20 Monte Carlo runs. The final RMSE comparison is therefore not on the same architecture as the uncertainty evaluation, and the baselines (ANN/PIML-ANN) remain at 200 nodes. Please report results for matched architectures or explicitly present the hyperparameter search as a separate exploratory result.","section":"Table I vs. Fig. 8"},{"comment":"The uncertainty estimates are only visualized as confidence bands; no quantitative calibration metric (e.g., empirical coverage, negative log-likelihood, interval score) is computed for any model. Without such a metric the claim that BNN-integrated PIML architectures successfully provision uncertainty propagation is not supported.","section":"IV, Figs. 7–11"},{"comment":"For the aircraft case study only Monte Carlo based uncertainty plots are shown; no Taylor-propagation results or runtime figures are presented for this more realistic problem. This leaves open whether the efficiency advantage claimed for Eq. (3) holds for the nonlinear VLM/propeller model. Please include Taylor-based results for this case or clearly state that the efficiency claim is restricted to the analytical benchmark.","section":"IV.B, Figs. 10–11"}],"minor_comments":[{"comment":"The denominator in the printed expression for f_FP(x) appears to be 2(π(x−0.5)/4), which is inconsistent with the substitution f_FP(x)=f_PP(0.5+2 sin(π(x−0.5)/4)); it should likely be 2(0.5+2 sin(π(x−0.5)/4)).","section":"Eq. (4)"},{"comment":"Equation (1) uses p(ŷ(x)|D), but the surrounding text refers to p(y|x,θ) and θ inconsistently; please unify the notation for inputs, weights, and predictive distributions.","section":"Eq. (1) and surrounding text"},{"comment":"The phrase 'Monty Flight' in the convergence-history discussion is a typo; it should refer to the fixed-wing aircraft case study.","section":"Appendix B"},{"comment":"The abstract's statement that Monte Carlo sampling was 'found to be most effective' is not supported by any quantitative uncertainty-quality comparison; please either add supporting metrics or soften the claim.","section":"Abstract and Section V"}],"recommendation":"major_revision","confidential_remarks":"The paper's central efficiency claim is not supported by the reported experiments because the proposed Taylor method fails on the analytical benchmark and is replaced by Monte Carlo without a cost/coverage comparison. The use of the authors' own BLOFI tool and prior PIML papers is appropriate for this line of work, but the manuscript would benefit from a clear statement of data and code availability. The accuracy comparison is further complicated by the hidden-unit mismatch between Table I and Fig. 8, which should be resolved in revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe headline: this is an honest, modest empirical study that reports a slightly negative result—BNNs dropped into their auto-differentiable hybrid PIML architectures give accuracy at or slightly below deterministic baselines, and the cheap Taylor uncertainty propagation they propose fails on their own analytical benchmark, so they fall back to end-to-end Monte Carlo. If you read one thing, read the discussion around Figs. 7 and 8: the paper's central efficiency claim is quietly walked back there.\n\nWhat is new: the specific combination of BNNs (last-layer Bayesian) with their input-transformation/adaptive-parameter PIML architectures, plus a direct MC-vs-Taylor comparison, is not in the cited literature. The two-stage training (deterministic pretraining then Bayesian retraining with learned priors) is sensible and clearly described. Credit where due: the authors do not oversell. The abstract states the accuracy is slightly worse or at par, and the conclusions admit MC outperforms Taylor. The flight-data case with a VLM/propeller model is a realistic test.\n\nThe soft spots are real and proportionate. First, the analytical comparison in Fig. 8 uses 10 hidden units for the PIML-BNN and BNN while the baselines use 200 nodes per layer; that post-hoc choice makes the accuracy comparison hard to interpret. Second, there is no quantitative calibration check: no coverage, reliability, or sharpness metric for any uncertainty band. Third, the efficiency claim for Eq. (3) Taylor propagation is unmeasured: no wall-clock time, no sample count, and their own Fig. 7 shows the linearized bounds deviate dramatically from the true function on a smooth 1-D benchmark. That last point is the load-bearing one, because 'efficient quantification' is the paper's stated goal.\n\nOn the plus side, the paper's own text flags these limitations in the results section—they explicitly say MC was used 'to address this'—so it is not hiding the problem. The missing piece is a quantitative accuracy-versus-cost comparison and a coverage evaluation; both are addressable in revision.\n\nBottom line: this is a useful preliminary study for people building UQ into hybrid physics-ML surrogates, but it does not yet demonstrate the efficiency it claims. I would send it to review, with a clear request to fix the hidden-unit comparison and add calibration metrics. I wouldn't cite it in my own work yet.\n\nRecommendation: engage with it as a conditional paper, not a reject.","headline":"Honest small negative result: BNNs in hybrid PIML give at-par accuracy, but the paper's cheap Taylor uncertainty propagation fails on its own benchmark and the efficiency claim is unmeasured.","tokens_in":13287,"tokens_out":2043,"would_cite":false,"duration_ms":21311,"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":"Bayesian nets add error bars to physics-informed ML","keywords":["Bayesian neural networks","physics-informed machine learning","hybrid PIML","uncertainty quantification","uncertainty propagation","Monte Carlo sampling","variational inference","aerodynamic modeling"],"falsifier":"On a nonlinear physics model with a known transfer-parameter distribution, compute 68% confidence bounds once from Eq. (3) and once from end-to-end Monte Carlo using the same trained BNN; if the Taylor bounds' coverage falls substantially below 68% while the Monte Carlo bounds are well calibrated, the efficiency claim fails.","tokens_in":12199,"feed_emoji":"📊","tokens_out":8609,"duration_ms":82075,"temperature":0.7,"pith_summary":"The paper asks whether Bayesian neural networks (BNNs) can be dropped into hybrid physics-informed machine learning (PIML) architectures — where a partial physics model is flanked by a neural network that supplies transfer parameters — to provide uncertainty estimates without sacrificing accuracy. It proposes a two-stage training scheme that first trains a deterministic PIML model, then uses the learned weights as priors for a BNN, and shows that the resulting BNN-integrated model predicts about as well as purely data-driven and original PIML models on an analytical benchmark and on flight data for a fixed-wing RC aircraft. The paper's headline finding is that Monte Carlo sampling of the probabilistic weights through the entire model is the reliable way to propagate the uncertainty, whereas the cheaper first-order Taylor expansion through the physics model fails on the nonlinear benchmark. If this holds, engineers can add calibrated confidence bounds to physics-based surrogates for robust design and control at a modest training-cost increase.","feed_headline":"Bayesian nets give PIML models uncertainty bands","feed_subtitle":"A two-stage training scheme keeps accuracy at par while Monte Carlo sampling carries the error bars through the physics.","key_machinery":"The central object is the BNN-integrated hybrid PIML architecture: a Bayesian neural network (with only its final layer probabilistic) replaces the deterministic ANN in a serial PIML configuration, producing transfer parameters (and their uncertainties) that feed into an auto-differentiable partial physics model. Uncertainty is then carried to the outputs either by a first-order Taylor expansion using the physics model's Jacobian (Eq. 3, the efficient option) or by Monte Carlo sampling of the BNN weights end-to-end through the physics model. The two-stage training scheme — deterministic pretraining with MSE loss followed by Bayesian fine-tuning with MSE + ELBO loss, using the deterministic weights as priors — is the mechanism that makes the architecture trainable.","core_discovery":"The paper's central claim is that a BNN-integrated PIML architecture — in which a Bayesian neural network outputs transfer parameters that enter a differentiable partial physics model — can predict both mean outputs and their uncertainties, with prediction performance slightly worse or at par with deterministic ANN, PIML-ANN, and BNN baselines. The mechanism is a two-stage training procedure that initializes the Bayesian layer with the weights of a previously trained deterministic PIML model, then refines with a combined MSE and ELBO loss, making the probabilistic training tractable. On the Gramacy & Lee benchmark the first-order Taylor propagation of Eq. (3) produced confidence bounds that deviated strongly from the true function, while end-to-end Monte Carlo sampling over the BNN weights recovered a meaningful uncertainty band; the same comparison on real flight data confirmed that PIML-BNN uncertainty plots track the data when Monte Carlo is used. The paper therefore establishes the BNN-as-transfer-net design as a viable route to model uncertainty in PIML, and identifies the propagation method—not the architecture—as the critical choice.","pith_inferences":["The paper does not test how the Taylor-propagated bounds degrade as the partial physics becomes more nonlinear; a testable extension is to measure their coverage against end-to-end Monte Carlo across a family of physics models with increasing curvature.","The result that 10 hidden units beat 200 on the simple benchmark suggests the transfer network's capacity affects uncertainty calibration, a relationship the paper leaves unexamined.","If the drop-in idea generalizes, other probabilistic surrogates (for instance, Gaussian processes) could occupy the transfer-parameter role, and the two-stage pretraining idea would carry over to them."],"forward_implications":["Uncertainty-aware surrogates for engineering design can be obtained by swapping the ANN in an existing hybrid PIML for a BNN, without redesigning the partial physics model.","The two-stage training recipe (deterministic pretrain, then Bayesian fine-tune with weight priors) is a practical route to training probabilistic networks inside hybrid architectures.","End-to-end Monte Carlo sampling should be the default propagation method when the partial physics is nonlinear, because the Taylor expansion did not track the true function on the benchmark.","The BNN-PIML model stays more accurate than the low-fidelity physics alone while providing confidence bounds, at a training-time premium of about a factor of seven to ten in the reported cases.","In the aircraft case, the remaining accuracy shortfall on two force components points to the low-fidelity propeller model, not the Bayesian layer, as the main error source."],"supporting_citations":[{"why":"Supplies the deterministic hybrid PIML architectures and the BLOFI low-fidelity aircraft model that the BNN replaces.","marker":"[1]"},{"why":"Documents the differentiable physics-informed ML approach whose auto-differentiable partial physics models carry the uncertainty propagation.","marker":"[2]"},{"why":"Defines Bayesian neural networks and the predictive distribution expressed in Eq. (1).","marker":"[39]"},{"why":"Provides the variational inference ELBO objective used in the second-stage Bayesian training.","marker":"[44]"}],"fun_headline_variants":["Monte Carlo sampling makes PIML uncertainty work","BNN plus PIML: uncertainty at par with deterministic","Two-stage training unlocks Bayesian PIML","Physics-informed ML gains uncertainty via Bayesian nets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The framework's efficiency claim rests on the assumption that a first-order Taylor expansion can adequately carry transfer-parameter uncertainties through the partial physics model, which the paper's own nonlinear benchmark contradicts.","fun_headline_variants_meta":{"raw":{"variants":["Monte Carlo sampling makes PIML uncertainty work","BNN plus PIML: uncertainty at par with deterministic","Two-stage training unlocks Bayesian PIML","Physics-informed ML gains uncertainty via Bayesian nets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000591,"raw_usage":{"total_tokens":2804,"prompt_tokens":1011,"completion_tokens":1793,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":627,"completion_tokens_details":{"reasoning_tokens":1732}},"tokens_in":627,"tokens_out":1793,"duration_ms":14108,"temperature":1.0,"reasoning_tokens":1732,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:22:57.798927+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a nonlinear physics model with a known transfer-parameter distribution, compute 68% confidence bounds once from Eq. (3) and once from end-to-end Monte Carlo using the same trained BNN; if the Taylor bounds' coverage falls substantially below 68% while the Monte Carlo bounds are well calibrated, the efficiency claim fails.","supporting_citations":[{"cited_title":"Physics infused machine learning based predic- tion of vtol aerodynamics with sparse datasets,","cited_arxiv_id":null,"evidence_quote":"Supplies the deterministic hybrid PIML architectures and the BLOFI low-fidelity aircraft model that the BNN replaces."},{"cited_title":"A differentiable physics-informed machine learn- ing approach to model laser-based micro-manufacturing process,","cited_arxiv_id":null,"evidence_quote":"Documents the differentiable physics-informed ML approach whose auto-differentiable partial physics models carry the uncertainty propagation."},{"cited_title":"Bayesian neural networks: An introduction and survey,","cited_arxiv_id":null,"evidence_quote":"Defines Bayesian neural networks and the predictive distribution expressed in Eq. (1)."},{"cited_title":"Stochastic variational inference,","cited_arxiv_id":null,"evidence_quote":"Provides the variational inference ELBO objective used in the second-stage Bayesian training."}],"review_version":1}