Pith. sign in

REVIEW 4 major objections 5 minor 36 references

This paper establishes that conformal prediction can be applied at three granularities—case level, point level, and vehicle level—to turn deterministic neural-operator aerodynamic surrogates into calibrated reliability indicators.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:24 UTC pith:MOPU6WKV

load-bearing objection A solid, honest empirical study of conformal calibration for automotive aerodynamic surrogates; the OOF protocol breaks exchangeability so the headline coverage numbers overstate the formal guarantee, but the split-protocol results and the multi-granularity comparison are genuinely useful. the 4 major comments →

arxiv 2607.17297 v1 pith:MOPU6WKV submitted 2026-07-19 physics.flu-dyn

Multi-Granularity Conformal Prediction for Reliable Neural-Operator Automotive Aerodynamic Surrogates

Chundong Jia , Chao Xia , Alexey Vdovin , Qing Jia , Simone Sebben , Zhigang Yang This is my paper
classification physics.flu-dyn
keywords conformal predictionneural operatorsuncertainty quantificationaerodynamic surrogatesautomotive aerodynamicsdrag coefficientwall shear stressDrivAerML
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that a single conformal-prediction pipeline can be attached to deterministic neural-operator surrogates to produce calibrated reliability intervals for both scalar drag coefficient and surface pressure/wall-shear-stress fields. For drag, it shows that asymmetric conformalized quantile regression corrects the raw quantile intervals' under-coverage, and that pooling out-of-fold scores reduces the Monte Carlo coverage standard deviation from 10.41 to 3.10 percentage points. For surface fields, it shows that normalizing residuals by a learned pointwise residual-scale field yields spatially adaptive intervals that hold near 90% coverage while cutting mean width by 22.68% for pressure and 25.35–27.09% for wall shear stress under the out-of-fold protocol. If this is right, engineers can use the calibrated intervals to decide which geometries and which surface regions deserve follow-up CFD verification, instead of relying on uncalibrated point predictions.

Core claim

The central discovery is that the same conformal calibration machinery works at three granularities: case-level for drag coefficient, point-level for surface fields, and vehicle-level for reliability screening. For C_d, a center-width quantile head is trained with pinball loss, and asymmetric lower/upper expansion offsets are calibrated on held-out scores — under under-coverage of raw intervals, the offsets expand the interval to reach ~90% coverage. For surface fields, the model is trained to output both the mean field and a residual-scale field; the conformal score is the absolute residual divided by this scale. Calibrating a single multiplier per channel over all points (point-adaptive),

What carries the argument

The key objects are the residual-normalized conformal score s = |y − ŷ| / (σ̂ + ε) and, for the scalar task, asymmetric positive-part quantile-expansion scores. The residual-scale field σ̂ is trained alongside the mean predictor under a smooth L1 log loss with local-neighborhood smoothness regularization. These scores are what make a single calibrated multiplier per output channel act on local uncertainty: intervals widen exactly where the learned scale is large. The out-of-fold (OOF) protocol pools scores from K-fold models to compute the multiplier, and the final model trained on all data inherits that multiplier.

Load-bearing premise

The coverage guarantee relies on the calibration and deployment geometries being exchangeable draws from the same distribution — the fixed official split and the out-of-fold pooling both assume the test morphs are statistically similar to the training/calibration morphs; if a new geometry falls outside that range, the reported coverage need not hold.

What would settle it

Evaluate the calibrated intervals on a set of vehicle geometries deliberately shifted outside the morphing range used for training/calibration (e.g., longer wheelbase, different roof slope), or re-split the official data into many different calibration/test partitions and recompute coverage; if coverage drops well below 90% or varies by more than the reported standard deviations, the central claim of calibrated reliability under the stated protocols is falsified. A second direct check: compare the OOF multiplier against the multiplier obtained from a single model trained on all 400 training ca

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim holds, calibrated Cd intervals tell designers when a candidate geometry's drag target cannot be confidently assessed without a CFD check.
  • Spatially adaptive surface intervals provide local uncertainty maps that flag wheels, sharp edges, underbody, and wake-influenced regions for targeted verification.
  • Case-wise normalized calibration reduces the number of vehicles with below-90% point coverage, giving a vehicle-level reliability signal that point-pooled coverage misses.
  • Smoothness regularization makes residual-scale fields spatially coherent and narrows intervals without coverage loss, which should make the uncertainty maps interpretable in practice.
  • Because the pipeline needs only point predictions plus a residual scale, it can be added to any neural-operator surrogate after training, at negligible calibration cost.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's OOF protocol is not a strict jackknife+ procedure; the coverage guarantee transferred to the final full-data model is approximate. A stricter variant would use cross-conformal or jackknife+ multipliers to retain a finite-sample guarantee under the same data budget.
  • The stated coverage numbers presuppose exchangeability between calibration and test geometries; on a distribution-shifted set of morphing parameters, coverage would likely degrade. Combining the intervals with a distance-to-training score, as the paper suggests for future work, would extend their validity to shifted geometries.
  • The residual-scale branch is architecture-agnostic; one could retrofit this calibration to any existing aerodynamic surrogate that outputs point predictions, including point-cloud or mesh-based models, provided the calibration data are recorded. This may make the method immediately useful for industrial CFD screening pipelines.
  • A block-bootstrap over vehicles (resampling whole cars rather than points) would give a more honest estimate of case-level coverage variability than point-level pooling, since surface points within a vehicle are spatially correlated.

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

4 major / 5 minor

Summary. The paper develops a multi-granularity conformal-prediction framework for neural-operator surrogates in automotive aerodynamics. For scalar drag coefficient, it uses conformalized quantile regression with asymmetric positive-part scores. For surface pressure and wall shear stress, it predicts a residual-scale field and applies residual-normalized conformal calibration at global, point-adaptive, and case-wise granularities, optionally with a smoothness regularizer. The method is evaluated on the DrivAerML dataset with GeoTransolver and Transolver backbones under both ordinary split and cross-validation-assisted out-of-fold protocols. The main reported results are that conformal calibration restores near-90% coverage, the OOF protocol reduces Monte Carlo coverage standard deviation from 10.41 to 3.10 percentage points, point-adaptive normalized calibration gives the narrowest near-nominal intervals (22.68% width reduction for pressure, 25.35--27.09% for WSS), and the pipeline transfers across backbones.

Significance. If the headline claims are supported, the paper makes a useful contribution to reliable use of neural-operator surrogates in engineering: it adapts conformal prediction to high-dimensional, spatially heterogeneous surface fields, introduces a learned residual-scale normalization, and reports vehicle-level coverage statistics rather than only pooled point coverage. The split-conformal part is mathematically sound, the code is publicly available, and the experimental reporting is unusually transparent (case counts, Monte Carlo stability, case-level coverage distributions). The main gap is that the OOF protocol used for several headline numbers does not carry a distribution-free coverage guarantee, and the small official test set makes the reported coverage differences noisy. These issues are fixable by reframing claims, adding uncertainty intervals, or replacing/justifying the OOF procedure.

major comments (4)
  1. [§3.7, Tables 4, 9, 10] The OOF-to-final-model transfer does not inherit a conformal coverage guarantee. In §3.7, OOF scores are computed from fold models trained on 320 or 300 cases, pooled, and then the resulting quantile is applied to predictions of a final model retrained on all 400 cases. Standard conformal coverage requires exchangeability between the calibration scores and the test score; here the final model has a different (systematically smaller) error distribution, as the paper itself reports in Tables 3 and 8. The pooled OOF quantile is therefore not an order statistic of exchangeable scores with the test score. The paper should either (i) provide a theoretical justification (e.g., bounds in the spirit of jackknife+), (ii) explicitly label the OOF coverage as empirical rather than guaranteed and report uncertainty bounds on the coverage, or (iii) use a split-conformal protocol for the central claims
  2. [Tables 4, 5, 9, 10] The official test set has only 50 cases. A reported 90% coverage has a 95% confidence interval of roughly ±8.3 percentage points (n=50); for example, 94.00% coverage in Table 4 has an approximate 95% CI of [83.5%, 98.7%]. The observed differences between split and OOF coverages, and among calibration strategies, may be within sampling error. Additionally, the Monte Carlo stability check in Table 5 resamples calibration/evaluation splits from cached OOF predictions, so it measures variability of OOF score quantiles, not the coverage of the final model on the official test set. Please report confidence intervals (binomial or case-level bootstrap) for all coverage estimates and clarify what the Monte Carlo experiment does and does not establish.
  3. [§5.5] The limitation paragraph correctly acknowledges exchangeability between calibration and deployment distributions, but it does not mention the separate mismatch introduced by the OOF protocol: the calibration scores come from subset-trained fold models, while the test predictions come from a full-data model. This is a distinct invalidation of the conformity guarantee and should be stated explicitly. Moreover, the paper does not test distribution shift between the official train/calibration and test geometries. Please add a discussion of this additional limitation and its consequence for the OOF coverage numbers.
  4. [§4.3, Table 4] The split-CP baseline uses each of the five 80-case calibration folds separately, while the OOF protocol pools 400 OOF scores and also refits the model on all 400 cases. The comparison in Tables 4 and 5 therefore conflates calibration size with the protocol itself. This is not a fatal flaw, but it should be stated explicitly so that readers do not attribute all of the OOF improvement to score pooling. If a matched calibration-size comparison is feasible, it would strengthen the claim that the OOF protocol itself, rather than training set size, is responsible for the improved stability.
minor comments (5)
  1. [§5.1] In the text following Table 4, the statement 'the calibrated OOF interval is slightly wider than the calibrated split interval' is correct numerically, but the subsequent comparison of interval scores is confounded by the different coverage levels; consider adding a sentence noting that interval score comparisons are most meaningful at equal coverage.
  2. [§5.2] The smoothness ablation in Table 6 reports 'Coverage' and 'Width' in units of 10^-3, but the header also says 'Coverage and width reduction are reported in percent.' Please separate units for coverage changes and width reductions in the table caption for clarity.
  3. [§3.5] In Eq. (34), the within-case aggregation uses the empirical 1-α quantile of point-level normalized scores. Since the number of surface points N is large (8192), this statistic is close to the 90th percentile of a continuous distribution, which is fine, but the sensitivity of the case-level score to N is not discussed; a short remark would help.
  4. [§4.1] The separate 80/10/10 random split for the smoothness ablation yields a 49-case test subset, rather than 40 as a literal 10% of 484 might suggest. Please clarify the number of cases assigned to each split in that ablation.
  5. [References] Reference 'Stephen et al., 2021' is a common placeholder for the Gentle Introduction paper; please update to include the full author list to avoid ambiguity.

Circularity Check

1 steps flagged

Conformal calibration chain is self-contained; only the auxiliary smoothness score is defined by its own training loss.

specific steps
  1. self definitional [Section 3.6 Eq. (37) vs. Section 5.2 Eq. (50) and Table 7]
    "a channel-wise smoothness score is computed using the same local-neighborhood criterion as the training regularizer: S_c = 1/|E| sum_{(i,j) in E} [log(sigma_{i,c}+epsilon)-log(sigma_{j,c}+epsilon)]^2 ... L_smooth = 1/|E| sum_{(i,j) in E} ||log(sigma_i+epsilon)-log(sigma_j+epsilon)||_2^2"

    Eq. (50) is the squared form of Eq. (37). With lambda_smooth=0.01 the training loss directly minimizes this score on the same kNN-8 edge set, so the reported 74.29% reduction in Table 7 is the regularizer lowering its own objective rather than an independent validation or first-principles prediction. This is auxiliary: the main conformal intervals are calibrated from held-out/OOF scores (Eqs. 20, 29, 31, 35, 39), so the tautology does not propagate to the coverage or width claims.

full rationale

The central conformal-prediction derivation is not circular. Scalar Cd intervals are built from quantile-regression outputs plus asymmetric CQR offsets computed from held-out calibration scores (Eqs. 19-21). Surface-field intervals use residual-normalized scores (Eq. 27) with conformal quantiles computed separately per channel from calibration data (Eqs. 29, 31, 35) or pooled OOF scores (Eqs. 38-44). The learned residual scale is trained on absolute residuals (Eq. 25), but the conformal multiplier is not re-fit to test coverage; it is obtained from calibration samples. Thus the reported coverage and width reductions are empirical, not enforced by construction. The backbones (GeoTransolver, Transolver) and DrivAerML dataset are external, and no load-bearing self-citation or imported uniqueness claim is used. The acknowledged exchangeability limitation (Sec. 5.5) and the separate OOF full-model calibration-score mismatch (Secs. 3.7 and 4.3) are statistical-validity concerns rather than circularity: the OOF intervals are not definitionally equal to their calibration inputs. Only the smoothness-score evaluation is self-referential, as flagged above, so the overall circularity score is low.

Axiom & Free-Parameter Ledger

5 free parameters · 3 axioms · 1 invented entities

No physical constants or new forces are invented; the residual-scale network is a learned auxiliary head, not a new entity. The chief load-bearing additions are model/config hyperparameters (a floor, two loss weights, knn size, numerical epsilons) and the OOF-calibration domain assumption. The paper's headline claims should be read as conditional on these choices.

free parameters (5)
  • residual-scale target floor r_min = not quantified
    Introduced in Eq. 24 to avoid ill-conditioned log-scale targets; affects residual-scale magnitudes and hence normalized interval widths. No sensitivity analysis.
  • scale loss weight lambda_scale = 0.05
    Chosen in Section 4.2 for the surface-field objective (Eq. 26); no ablation over lambda_scale is reported.
  • smoothness loss weight lambda_smooth = 0.01
    Selected for the smoothness ablation in Section 4.3; the paper compares only 0.01 vs 0, not the optimality of the value.
  • knn graph size k_nn = 8
    Used for the smoothness regularizer and smoothness score; no sensitivity analysis.
  • epsilon numerical constants = not quantified
    Softplus scale floor, residual floor, log-stability epsilon, and denominator epsilon in Eq. 25/27/37 affect relative interval widths and are unquantified.
axioms (3)
  • domain assumption Exchangeability of calibration and test samples
    Invoked for the finite-sample conformal coverage statement in Section 3.1, applied to the fixed DrivAerML official split; the paper itself acknowledges in Section 5.5 that this breaks under distribution shift.
  • domain assumption OOF calibration of a refit final model inherits conformal coverage
    Section 3.7 pools out-of-fold scores from fold models, then applies the quantile to final model predictions on the official test set. Standard conformal guarantees apply to the models that produced the scores, not necessarily to a new model refit on all 400 cases.
  • domain assumption Surface-field model outputs a well-calibrated local residual scale
    Section 3.4's normalized score assumes sigma captures spatial error heterogeneity; if sigma is not coherent, the normalized intervals inherit its bias.
invented entities (1)
  • Case-wise normalized conformal aggregation independent evidence
    purpose: Aggregates point-level normalized scores per vehicle before taking the conformal quantile, to stabilize vehicle-level coverage.
    It is a defensible calibration variant with measurable case-level coverage numbers, so it carries a falsifiable handle through the reported tables; it postulates no new physical object.

pith-pipeline@v1.3.0-alltime-deepseek · 20301 in / 12261 out tokens · 108988 ms · 2026-08-01T18:24:33.412761+00:00 · methodology

0 comments
read the original abstract

High-fidelity computational fluid dynamics (CFD) provides detailed aerodynamic data for vehicle design, but its cost limits design iteration. Neural-operator surrogates reduce this cost, yet their deterministic predictions do not indicate when a geometry or surface region is reliable. This study develops a conformal-prediction framework for reliability-aware automotive aerodynamic surrogate modeling on the DrivAerML dataset. GeoTransolver is the main backbone, while Transolver assesses transfer across neural-operator architectures. For drag coefficient prediction, conformalized quantile regression constructs calibrated case-level intervals. For surface pressure and wall shear stress (WSS), point prediction is combined with residual-scale estimation and residual-normalized conformal calibration to obtain spatially adaptive intervals. Global absolute, point-adaptive normalized, and case-wise normalized calibration are compared under split and cross-validation-assisted out-of-fold protocols. All experiments target 90% nominal coverage. Conformal calibration corrects the under-coverage of raw drag-coefficient quantile intervals, while out-of-fold score aggregation reduces the Monte Carlo coverage standard deviation from 10.41 to 3.10 percentage points. For surface fields, point-adaptive normalized calibration yields the narrowest near-nominal intervals, reducing mean width by 22.68% for pressure and 25.35%--27.09% for WSS under the out-of-fold protocol. Case-wise normalized calibration is more conservative but improves vehicle-level reliability. Smoothness regularization reduces the residual-scale local-variation score by 74.29% and lowers interval widths without material coverage loss. The framework converts deterministic neural-operator outputs into calibrated reliability indicators for prioritizing uncertain vehicle geometries and surface regions in follow-up CFD verification.

Figures

Figures reproduced from arXiv: 2607.17297 by Alexey Vdovin, Chao Xia, Chundong Jia, Qing Jia, Simone Sebben, Zhigang Yang.

Figure 1
Figure 1. Figure 1: Scalar drag coefficient interval prediction with connected readout and calibration [PITH_FULL_IMAGE:figures/full_fig_p010_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Surface-field prediction and calibration workflow. The model estimates both the [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Predicted residual-scale fields for a representative vehicle. The left and right [PITH_FULL_IMAGE:figures/full_fig_p026_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Representative deterministic surface-field predictions on a test vehicle. The [PITH_FULL_IMAGE:figures/full_fig_p027_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Spatial distribution of conformal interval width under the CV-assisted OOF [PITH_FULL_IMAGE:figures/full_fig_p030_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

36 extracted references · 8 linked inside Pith

  1. [1]

    Journal of Mechanical Design , volume=

    Drivaernet: A parametric car dataset for data-driven aerodynamic design and prediction , author=. Journal of Mechanical Design , volume=. 2025 , publisher=

    2025

  2. [2]

    Advances in Neural Information Processing Systems , volume=

    Drivaernet++: A large-scale multimodal car dataset with computational fluid dynamics simulations and deep learning benchmarks , author=. Advances in Neural Information Processing Systems , volume=

  3. [3]

    arXiv preprint arXiv:2408.11969 , year=

    DrivAerML: High-fidelity computational fluid dynamics dataset for road-car external aerodynamics , author=. arXiv preprint arXiv:2408.11969 , year=

    Pith/arXiv arXiv

  4. [4]

    Nature machine intelligence , volume=

    Learning nonlinear operators via DeepONet based on the universal approximation theorem of operators , author=. Nature machine intelligence , volume=. 2021 , publisher=

    2021

  5. [5]

    arXiv preprint arXiv:2010.08895 , year=

    Fourier neural operator for parametric partial differential equations , author=. arXiv preprint arXiv:2010.08895 , year=

    Pith/arXiv arXiv 2010

  6. [6]

    Engineering Applications of Artificial Intelligence , volume=

    Deep neural operators as accurate surrogates for shape optimization , author=. Engineering Applications of Artificial Intelligence , volume=

  7. [7]

    Sequential Deep Operator Networks (

    He, Junyan and Kushwaha, Shashank and Park, Jaewan and Koric, Seid and Abueidda, Diab and Jasiuk, Iwona , journal=. Sequential Deep Operator Networks (

  8. [8]

    Engineering Applications of Artificial Intelligence , volume=

    A geometry-adaptive physics-informed operator framework generalized for arbitrary geometries , author=. Engineering Applications of Artificial Intelligence , volume=

  9. [9]

    Advances in Neural Information Processing Systems , volume=

    Geometry-informed neural operator for large-scale 3d pdes , author=. Advances in Neural Information Processing Systems , volume=

  10. [10]

    arXiv preprint arXiv:2501.13350 , year=

    Domino: A decomposable multi-scale iterative neural operator for modeling large scale engineering simulations , author=. arXiv preprint arXiv:2501.13350 , year=

    Pith/arXiv arXiv

  11. [11]

    arXiv preprint arXiv:2402.02366 , year=

    Transolver: A fast transformer solver for pdes on general geometries , author=. arXiv preprint arXiv:2402.02366 , year=

    Pith/arXiv arXiv

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Geometry aware operator transformer as an efficient and accurate neural surrogate for pdes on arbitrary domains , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    arXiv preprint arXiv:2502.09692 , year=

    AB-UPT: Scaling neural CFD surrogates for high-fidelity automotive aerodynamics simulations via anchored-branched universal physics transformers , author=. arXiv preprint arXiv:2502.09692 , year=

    arXiv

  14. [14]

    Journal of Computational Physics , pages=

    Spatially-aware transformer operator for real-time aerodynamic evaluations of arbitrary three-dimensional vehicles , author=. Journal of Computational Physics , pages=. 2025 , publisher=

    2025

  15. [15]

    Forty-third International Conference on Machine Learning , year=

    SMART: Scalable Mesh-free Aerodynamic Simulations from Raw Geometries using a Transformer-based Surrogate Model , author=. Forty-third International Conference on Machine Learning , year=

  16. [16]

    arXiv preprint arXiv:2512.20399 , year=

    GeoTransolver: Learning Physics on Irregular Domains Using Multi-scale Geometry Aware Physics Attention Transformer , author=. arXiv preprint arXiv:2512.20399 , year=

    Pith/arXiv arXiv

  17. [17]

    2005 , publisher=

    Algorithmic learning in a random world , author=. 2005 , publisher=

    2005

  18. [18]

    , author=

    A tutorial on conformal prediction. , author=. Journal of machine learning research , volume=

  19. [19]

    arXiv preprint arXiv: 2107.07511 , year=

    A gentle introduction to conformal prediction and distribution-free uncertainty quantification , author=. arXiv preprint arXiv: 2107.07511 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2402.01960 , year=

    Calibrated uncertainty quantification for operator learning via conformal prediction , author=. arXiv preprint arXiv:2402.01960 , year=

    Pith/arXiv arXiv

  21. [21]

    Physica D: Nonlinear Phenomena , volume=

    Conformalized-deeponet: A distribution-free framework for uncertainty quantification in deep operator networks , author=. Physica D: Nonlinear Phenomena , volume=. 2025 , publisher=

    2025

  22. [22]

    Machine Learning: Science and Technology , volume=

    Uncertainty quantification of surrogate models using conformal prediction , author=. Machine Learning: Science and Technology , volume=. 2026 , publisher=

    2026

  23. [23]

    arXiv preprint arXiv:2412.10459 , year=

    Conformal Prediction on Quantifying Uncertainty of Dynamic Systems , author=. arXiv preprint arXiv:2412.10459 , year=

    Pith/arXiv arXiv

  24. [24]

    2012 , institution=

    Introduction of a new realistic generic car model for aerodynamic investigations , author=. 2012 , institution=

    2012

  25. [25]

    International conference on machine learning , pages=

    Weight uncertainty in neural network , author=. International conference on machine learning , pages=. 2015 , organization=

    2015

  26. [26]

    international conference on machine learning , pages=

    Dropout as a bayesian approximation: Representing model uncertainty in deep learning , author=. international conference on machine learning , pages=. 2016 , organization=

    2016

  27. [27]

    Advances in neural information processing systems , volume=

    Simple and scalable predictive uncertainty estimation using deep ensembles , author=. Advances in neural information processing systems , volume=

  28. [28]

    Garg, Shailesh and Chakraborty, Souvik , journal=

  29. [29]

    Advances in neural information processing systems , volume=

    Conformalized quantile regression , author=. Advances in neural information processing systems , volume=

  30. [30]

    Engineering Applications of Artificial Intelligence , volume=

    Prediction intervals to account for uncertainties in neural network predictions: Methodology and application in bus travel time prediction , author=. Engineering Applications of Artificial Intelligence , volume=

  31. [31]

    Engineering Applications of Artificial Intelligence , volume=

    Uncertainty quantification in sequential hybrid deep transfer learning for solar irradiation predictions , author=. Engineering Applications of Artificial Intelligence , volume=

  32. [32]

    Engineering Applications of Artificial Intelligence , volume=

    Towards robust ferrous scrap material classification with deep learning and conformal prediction , author=. Engineering Applications of Artificial Intelligence , volume=

  33. [33]

    Engineering Applications of Artificial Intelligence , volume=

    Integrating permutation feature importance with conformal prediction for robust Explainable Artificial Intelligence in predictive process monitoring , author=. Engineering Applications of Artificial Intelligence , volume=

  34. [34]

    arXiv preprint arXiv:2512.13069 , year=

    Multi-fidelity aerodynamic data fusion by autoencoder transfer learning , author=. arXiv preprint arXiv:2512.13069 , year=

    arXiv

  35. [35]

    The Annals of Statistics , volume=

    Predictive inference with the jackknife+ , author=. The Annals of Statistics , volume=. 2021 , publisher=

    2021

  36. [36]

    Engineering Applications of Artificial Intelligence , volume=

    A novel data generation scheme for surrogate modelling with deep operator networks , author=. Engineering Applications of Artificial Intelligence , volume=