Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Active Learning of Model Discrepancy with Bayesian Experimental Design

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read The paper claims that sequential Bayesian experimental design for low-dimensional physics parameters can, with a gating indicator, actively supply the data needed to learn a neural-network correction of model discrepancy, making full…

desk verdict A plausible alternating BED/gradient-discrepancy framework whose load-bearing assumption—that θG-optimal designs inform θNN—is only validated retrospectively; worth refereeing but needs a prospective test. read the letter →

arxiv 2502.05372 v2 pith:KN3HH6XE submitted 2025-02-07 cs.LG

Huchen Yang , Chuanqi Chen , Jin-Long Wu This is my paper · ORCID
classification cs.LG MSC 62K0562F1568T07
keywords modeldiscrepancyBayesianexperimentaldesignsequentialactivelearningensembleKalmaninversionneuralnetworkcorrectiondigitaltwininformationgain
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

This paper sets out to solve a practical circularity: Bayesian experimental design (BED) chooses the most informative measurements for a model's parameters, but the model is wrong, and data-driven corrections are usually trained on whatever data happen to be available. The authors propose an alternating loop in which each stage of sequential BED selects a measurement for the low-dimensional physics parameters, and that same measurement is then used to update a neural network that approximates the model discrepancy. A cheap ensemble-Kalman indicator of information gain decides whether the new data are worth using for the discrepancy update, preventing the loop from being corrupted by uninformative or misleading measurements. If the scheme works as claimed, digital-twin models can actively gather their own correction data without paying the cost of full Bayesian inference over high-dimensional network parameters. The paper validates the claim on a convection-diffusion source-inversion problem, including a high-dimensional structural-error case where full BED is no longer feasible.

What carries the argument

The load-bearing mechanism is the alternating block-coordinate loop: at each stage, (1) a greedy sequential BED problem selects a design $d$ by maximizing expected information gain for the physics parameters $\theta_G$, with the current discrepancy network fixed; (2) a MAP estimate or an average of top-$m$ posterior modes estimates $\theta_G$; (3) a gradient step (via adjoint or automatic differentiation) updates the discrepancy parameters $\theta_{NN}$ by maximizing the likelihood at the fixed $\theta_G$. The second mechanism is the ensemble-based KLD indicator: running a few EKI iterations on $\theta_{NN}$ yields Gaussian ensembles whose prior-posterior KL divergence, Eq. (18), approximates the information gain of the candidate data, and the update is applied only if that value is large. Together they convert the high-dimensional Bayesian inference problem over $\theta_{NN}$ into a sequence of low-dimensional BED problems plus cheap gradient updates.

What would settle it

Run the hybrid loop on a problem where the discrepancy term is deliberately placed so it vanishes at the BED-optimal sensor locations but is large elsewhere; the claim predicts the KLD indicator will stay near zero and the network will fail to correct the error, and a competing method that chooses designs directly for discrepancy informativeness would outperform. A cheaper check is to rank a batch of random candidate designs by the EKI-KLD indicator and compare that ranking against the actual reduction in discrepancy error achieved by each design; the claim stands only if the indicator ranking correlates positively with true correction quality.

Watch

Extended reading notes

Core claim

The central claim is that optimal designs for the physics parameters $\theta_G$ are, to a large extent, also informative for calibrating the discrepancy parameters $\theta_{NN}$, so the sequential BED loop can double as an active learning scheme for model error. The paper develops a hybrid framework that at each stage (i) solves the BED problem for $\theta_G$ with the current correction term fixed, (ii) forms a robust point estimate of $\theta_G$ by averaging top posterior modes, and (iii) updates $\theta_{NN}$ by gradient-based maximization of the likelihood at that point. To guard against bad data, an ensemble Kalman inversion (EKI) approximation of the KL divergence between posterior and prior ensembles is used as an indicator; updates are only performed when the indicator is large. The numerical results show that the corrected model keeps the posterior of the source location close to the truth, that informative designs reduce the mean squared error of the predicted field from 0.142 (baseline) to 0.005, and that the same loop also works in an ill-posed case with an unknown convection coefficient, where the network compensates the wrong convection velocity yet the source-location posterior remains accurate.

Load-bearing premise

The loop depends on the assumption that measurements chosen to be informative about the physics parameters are also informative about the discrepancy network; the paper states this can occasionally fail and produce misleading updates that the indicator can only filter after the design is fixed.

Editorial extensions

If this is right

  • Sequential BED can be run with a high-dimensional neural-network correction term without MCMC over the network weights.
  • The EKI-based KLD indicator provides a cheap gate: designs whose data would bias the discrepancy update can be detected and skipped.
  • The same alternating loop works with classical PDE solvers (adjoint method) and with auto-differentiable solvers, so it ports across solver ecosystems.
  • In ill-posed joint estimation settings, the corrected model can produce accurate posteriors for the physics parameters even when an auxiliary physics parameter (the convection growth rate) is not identifiable.
  • The approach transfers to at least one other inverse problem (acoustic source amplitude with medium-property discrepancy), suggesting it is not tied to the convection-diffusion example.

Reading between the lines

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

  • The indicator as presented gates updates after a design is fixed; a natural next step the paper names is to invert the flow and use the indicator itself as the design objective, selecting points that maximize $\theta_{NN}$ informativeness rather than $\theta_G$ informativeness.
  • The success of using $\theta_G$-optimal designs for $\theta_{NN}$ suggests a testable transfer principle: in problems where $\theta_G$ and $\theta_{NN}$ are strongly correlated through the forward map, the EIG landscape for $\theta_G$ may serve as a cheap surrogate for the intractable EIG landscape of $\theta_{NN}$.
  • The proposed post-training decomposition (re-attributing explained parts of the learned right-hand side back to the physics model) could turn the framework from a correction tool into an interpretability tool for identifying which parts of the physics are wrong.
  • Because the indicator is computed under a Gaussian ensemble assumption, its ranking of designs could be validated against exact MCMC-based information gain on small problems; a mismatch there would reveal where the Gaussian closure loses fidelity.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes a hybrid framework that alternates sequential Bayesian experimental design (BED) for low-dimensional physics-based parameters θG with gradient-based optimization of a neural-network correction term θNN that models model discrepancy. A Gaussian-ensemble approximation of the Kullback-Leibler divergence, computed via ensemble Kalman inversion (EKI), is introduced as an indicator of whether the data collected at a BED-chosen design are informative for updating θNN. The method is demonstrated on a contaminant-source inversion problem governed by a convection-diffusion equation, covering a parametric model error case, a structural (high-dimensional) model error case, and an ill-posed case with an unknown velocity coefficient, plus a Helmholtz acoustic inverse problem in an appendix.

Significance. If the central claim holds, the framework would be a practical way to repurpose BED-optimal designs for physics parameters as informative data for learning high-dimensional model discrepancy, avoiding full Bayesian inference over network parameters. The paper has several concrete strengths: the code and data are released; the parametric-error case is benchmarked against a full-grid BED; the method is demonstrated with both a classical solver and an auto-differentiable solver; and the authors explicitly acknowledge several limitations, including the assumption in Section 2.3 that BED-optimal designs for θG are informative for θNN, the occasional misleading data, the use of pre-training data generated from the true discrepancy form, and the restriction to spatial-only discrepancy. These candor points are valuable, but they also delineate precisely where the evidence falls short of the abstract's 'efficient and robust' claim.

major comments (4)
  1. [Section 2.3 and Algorithm 1] The load-bearing assumption that 'the data from an optimal design for θG is assumed to be informative in calibrating θNN' is not validated prospectively. The EKI-based KLD indicator in Eq. (18) is computed after the design and data are fixed and only decides whether to update θNN; it cannot select or repair a design. The supporting evidence in Sections 3.1.2 and 3.2 is retrospective: a 'poor' design is identified only after it is observed to underperform, and a 'good' design is a nearby point that was 'randomly checked' (Section 3.1.2) or a 'modified design' chosen in hindsight (Section 3.2). No threshold for 'if DKL is large' is specified in Algorithm 1, no distribution over feasible designs is examined, and no false-positive/false-negative analysis of the indicator is reported. Since the abstract claims that the method efficiently and robustly corrects high-dimensional discrepancy 'using data suggested by the sequential BED', this gap is load-bearing for the paper's central claim.
  2. [Section 3.2, Figure 6 and surrounding text] The paper's own single-run demonstration shows that the proposed algorithm, as written, can degrade: at stages 4 and 5 the high-probability posterior region deviates from the true source location because the stage-3 design and data, selected by BED for θG, were uninformative and biased the calibrated model. The improved results in Figure 7 are obtained by manually substituting a 'good design' selected in hindsight, not by any criterion in Algorithm 1. This directly contradicts the abstract's characterization of the method as robust, and it shows that the gating indicator, which is only applied after the design is fixed, does not protect against poor designs. A prospective demonstration is needed: run the full algorithm with an explicitly defined KLD threshold (or design-selection rule) on multiple trajectories and report the distribution of outcomes.
  3. [Section 3.2, neural-network initialization] The high-dimensional structural-error experiment relies on initializing the discrepancy network 'on data generated from the true target decaying source form' to enforce the center-decay structural constraint. The paper itself states that such pre-training data 'would not guarantee the desired constraint in the subsequently trained model and may not be available in real-world applications.' Because the true discrepancy function is exactly what the method is supposed to learn, this initialization weakens the demonstration that the framework can learn high-dimensional discrepancy from BED-suggested data alone. The authors mention an alternative penalty-based constraint that does not require true-form data; a demonstration using that approach, or at least an ablation without the true-form pre-training, would substantially strengthen the claim.
  4. [Appendix A] The 'sketch of proof' for the existence of a globally optimal solution lists conditions that include 'the alignment of the fields ensures that the modeled system G+NN captures the dynamics of the true system' and 'θ†G can be approximated by θ∗G from its posterior by BED.' These conditions are essentially restatements of the convergence result that the appendix is supposed to support, and the universal-approximation argument shows only that a network exists, not that the alternating BED/gradient-descent loop converges to it. The appendix should be re-labeled as heuristic motivation, and the convergence of the coupled iteration should be either formally analyzed or explicitly left as an open problem rather than presented as a proof.
minor comments (6)
  1. [Section 2.2, text near Eq. (8)] There are typos in this passage: 'd hiden' should be 'd hidden' and 'independant' should be 'independent'.
  2. [Section 3.1.2, Figure 5] The description of the 'good design' is underspecified: the authors state that they 'randomly checked several nearby points' but do not report how many points were checked, how they were sampled, or the seed used. This matters because the comparison in Figure 5b is between two hand-picked designs.
  3. [Algorithm 1] The condition 'if DKL is large' is not operationalized. Since the numerical examples compare only two designs per case, a reader cannot determine what threshold was used, whether the threshold was fixed in advance, or how sensitive the results are to that choice.
  4. [Table 1 and Figures 6-9] All numerical results are single-trajectory with no error bars, no repeated runs, and no sensitivity analysis with respect to the EKI ensemble size J, the number of EKI iterations K, or the number of gradient steps for θNN. Given that the paper emphasizes robustness, a small multi-seed study would materially improve the evidence.
  5. [Section 2.3, Eq. (18)] The Gaussian assumption underlying the ensemble-based KLD approximation is stated but not validated in the high-dimensional setting. In the parametric example the comparison is only qualitative, and for the structural-error case no comparison against a full Bayesian (e.g., MCMC) posterior for θNN is provided. A brief validation study would make the indicator's reliability more credible.
  6. [Section 4, multimodal posterior discussion] The discussion of the top-m averaging strategy is candid but also a limitation that deserves more prominence in the main text: the method is stated to be effective only when the posterior is unimodal or dominated by one mode, and the proposed multimodal extension is left for future work.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the central derivation is self-contained and benchmarked externally, with a mild retrospective validation limitation that is not definitional.

full rationale

Reviewed the full derivation chain. The core claims are (i) that alternating sequential BED over low-dimensional physics parameters theta_G with gradient-based updates of a neural-network discrepancy term theta_NN corrects model discrepancy, and (ii) that the EKI-based KLD approximation (Eq. 18) can gate or rank data informativeness. Neither claim reduces to its inputs. The BED objective (Eqs. 3-7) is a standard expected information gain over theta_G, and the discrepancy update (Eqs. 8-9) is a MAP/likelihood optimization, not a definition of the success metric. The KLD indicator is computed from ensemble covariances via Eq. 18 and evaluated against externally defined MSE, RE, and posterior accuracy. The Section 2.3 statement that theta_G-optimal designs are assumed informative for theta_NN is an explicit assumption, and the paper acknowledges that data can occasionally be less informative or even misleading; an assumption, even a potentially weak one, is not circular. The validation of the indicator in Figs. 5b and 8b is retrospective, in that a 'good' design is chosen after seeing improved calibration outcomes and then shown to have higher approximated KLD; this is a statistical-validation limitation rather than a definitional reduction, because the KLD values are computed independently and could in principle have contradicted the selection. Self-citations (e.g., [13], [54], [57] and other works by the same group) are used as background and are not load-bearing for the main claims; no uniqueness theorem from the authors' prior work is invoked to force the method. No equation in the paper is equivalent to its output by construction, and the hybrid framework is tested against full BED and standard solvers rather than being fitted to its own conclusions.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The method rests on additive neural-network discrepancy, Gaussian noise and Gaussian ensemble approximations, an informativeness correlation between theta_G-optimal designs and theta_NN calibration, a hand-crafted pre-training constraint, and an assumption of near-unimodal posteriors. The ledger shows no new physical entities are postulated; the main fragility is in the domain assumptions rather than in equations.

free parameters (3)
  • Initial model-error parameter for parametric-error case = theta_s = 3
    Section 3.1.2 starts the gradient-based correction from theta_s = 3 while the true value is 2; the demonstration of convergence depends on this starting point and no sensitivity to it is reported.
  • EKI ensemble size J = not reported
    The proposed indicator in Section 2.3 depends on J; the paper never states J or how choices affect the KLD approximation.
  • EKI iteration count K = 10 in the structural-error example
    Section 3.2 uses 10 EKI iterations to compute the indicator; no sensitivity study is provided.
assumptions (7)
  • domain assumption The true dynamics can be decomposed as G† = G + NN, i.e., the model discrepancy is additive and representable by a neural network.
    Eq. (2) in Section 2; all experiments assume this form, and the network is restricted to time-invariant spatial corrections.
  • domain assumption Measurement noise is zero-mean Gaussian, eta ~ N(0, Gamma).
    Section 2 and Eq. (A.1); the likelihood and EKI update both rely on Gaussian noise.
  • domain assumption EKI ensembles approximate the posterior and prior as Gaussian for the KLD indicator.
    Eq. (18) in Section 2.3 explicitly treats both ensembles as Gaussian; the paper notes this is an approximation.
  • ad hoc to paper The data from a design that is optimal for theta_G remains informative for theta_NN.
    Stated as the key idea in Section 2.3 and acknowledged to fail occasionally; this is the load-bearing premise of the active learning loop.
  • ad hoc to paper The discrepancy network is pre-trained on data from the true target source form to enforce a center-decay structural constraint.
    Section 3.2 describes this pre-training; the paper admits such data may not be available in real-world applications.
  • standard math Universal approximation of neural networks guarantees existence of a global optimum in Appendix A.
    Appendix A uses universal approximation to argue that some theta_NN can represent G† - G, but the argument is a sketch and does not prove convergence of the alternating algorithm.
  • domain assumption The posterior of theta_G is approximately unimodal so averaging top-m MAP points is a robust point estimate.
    Section 2.2 uses the mean of top-m posterior peaks; Section 4 notes this can fail for genuinely multimodal posteriors.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Active Learning of Model Discrepancy with Bayesian Experimental Design." pith.science (2026). https://pith.science/paper/KN3HH6XE

@misc{pith2026250205372,
  author       = {Pith},
  title        = {Pith review of: Active Learning of Model Discrepancy with Bayesian Experimental Design},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KN3HH6XE}},
  note         = {Machine review of arXiv:2502.05372}
}
read the original abstract

Digital twins have been actively explored in many engineering applications, such as manufacturing and autonomous systems. However, model discrepancy is ubiquitous in most digital twin models and has significant impacts on the performance of using those models. In recent years, data-driven modeling techniques have been demonstrated promising in characterizing the model discrepancy in existing models, while the training data for the learning of model discrepancy is often obtained in an empirical way and an active approach of gathering informative data can potentially benefit the learning of model discrepancy. On the other hand, Bayesian experimental design (BED) provides a systematic approach to gathering the most informative data, but its performance is often negatively impacted by the model discrepancy. In this work, we build on sequential BED and propose an efficient approach to iteratively learn the model discrepancy based on the data from the BED. The performance of the proposed method is validated by a classical numerical example governed by a convection-diffusion equation, for which full BED is still feasible. The proposed method is then further studied in the same numerical example with a high-dimensional model discrepancy, which serves as a demonstration for the scenarios where full BED is not practical anymore. An ensemble-based approximation of information gain is further utilized to assess the data informativeness and to enhance learning model discrepancy. The results show that the proposed method is efficient and robust to the active learning of high-dimensional model discrepancy, using data suggested by the sequential BED. We also demonstrate that the proposed method is compatible with both classical numerical solvers and modern auto-differentiable solvers.

Figures

Figures reproduced from arXiv: 2502.05372 by the authors.

Figure 1
Figure 1. Schematic diagram of the hybrid framework for sequential BED and active learning of model discrepancy. The prior and posterior distributions refer to the probability distribution of the parameter θG in Eq. (2). 3. Numerical Results To demonstrate the performance of our proposed method, we study the example of source inversion of a contaminant in the convection-diffusion field, which is a classical test example for B… view at source ↗
Figure 2
Figure 2. Concentration value at different times in a convection-diffusion field. The numerical simulation is performed in a larger domain ([−2, 3]2 ) and presented in a smaller one ([0, 1]2 ) to emphasize the areas close to the source location. Stages 1-5 correspond to 0.05, 0.10, 0.15, 0.20, 0.25 time units. The design d = {dx, dy, dt} in this problem refers to the spatiotemporal coordinate to measure the concentration valu… view at source ↗
Figure 3
Figure 3. Results of learning parametric model error via standard sequential BED. The top row shows the marginal distribution of θs at each stage, and the bottom row presents the two-dimensional conditional distribution of θx and θy, conditioned on the value of θs with the highest marginal probability density. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: shows the evolution of the posterior distribution for {θx, θy} across five stages. The high probability regions gradually converge to the true source location, suggesting an effective correction of parametric model error θs. The comparison between Figs 3 and 4 confirms…
Figure 5
Figure 5. Figure 5: b confirms that the performance of the ensemble-based approximated KLD as an indicator to evaluate data informativeness for the calibration of parametric model error. (a) Original model correction (b) EKI updating [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Posterior distributions of inferred source location {θx, θy} based on (a) a randomly initialized structural error term and (b) the calibrated structural error term [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]
Figure 7
Figure 7. Figure 7: Posterior distributions of inferred source location {θx, θy} based on the calibrated structural error using more informative data at stage 3. The blue triangular indicates the better design for structural error correction. (a) Network performance (b) EKI results [PITH…
Figure 8
Figure 8. Figure 8: Results of calibrating the structural error via the hybrid approach. Panel (a) presents the comparison of the true source term and the modeled ones. Panel (b) presents the ensemble-based approximated KLD for the model error calibration with good and poor designs. With …
Figure 9
Figure 9. Figure 9: Solution fields of the true system and the modeled ones. Panel (a) shows a visual comparison of the solutions fields. Panel (b) presents the mismatches between the true solution field and the solution fields from three different modeled systems [PITH_FULL_IMAGE:figure…
Figure 10
Figure 10. Figure 10: Posterior distributions of inferred source location {θx, θy} with (a) both ky and θNN fixed, (b) ky fixed and θNN updated , and (c) ky and θNN jointly updated. In panel (a), no model correction is applied; only one design point is selected per stage using BED, resulti…
Figure 11
Figure 11. Figure 11: Comparison between the true data, modeled approximation, and residuals across five measurement stages for RHS and solution field of the PDE: (a) Results for the RHS of the PDE, and (b) Results for the solution field of the PDE. Each panel consists of three rows: the f…

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Modeling Partially Observed Nonlinear Dynamical Systems and Efficient Data Assimilation via Discrete-Time Conditional Gaussian Koopman Network

    cs.LG 2025-07 conditional novelty 5.0 of 10

    A discrete-time conditional Gaussian Koopman network with analytical data-assimilation formulas is introduced and shown to match neural operator forecasts and ensemble Kalman filter assimilation on three PDE benchmarks.

Reference graph

Works this paper leans on

75 extracted references · 62 canonical work pages · cited by 1 Pith paper

  1. [1]

    Turbulence modeling in the age of data

    Karthik Duraisamy, Gianluca Iaccarino, and Heng Xiao. Turbulence modeling in the age of data. Annual review of fluid mechanics, 51(1):357–377, 2019

    work page 2019

  2. [2]

    Mont ´ans, Francisco Chinesta, Rafael G ´omez-Bombarelli, and J

    Francisco J. Mont ´ans, Francisco Chinesta, Rafael G ´omez-Bombarelli, and J. Nathan Kutz. Data-driven modeling and learning in science and engineering. Comptes Rendus M´ecanique, 347(11):845–855, November 2019

    work page 2019

  3. [3]

    Machine learning for fluid mechanics

    Steven L Brunton, Bernd R Noack, and Petros Koumoutsakos. Machine learning for fluid mechanics. Annual review of fluid mechanics, 52(1):477–508, 2020

    2020

  4. [4]

    Dynamic mode decomposition: data-driven modeling of complex systems

    J Nathan Kutz, Steven L Brunton, Bingni W Brunton, and Joshua L Proctor. Dynamic mode decomposition: data-driven modeling of complex systems. SIAM, 2016

    2016

  5. [5]

    Data-driven operator inference for nonintrusive projection-based model reduction

    Benjamin Peherstorfer and Karen Willcox. Data-driven operator inference for nonintrusive projection-based model reduction. Computer Methods in Applied Mechanics and Engineer- ing, 306:196–215, 2016

    work page 2016

  6. [6]

    Physics-informed machine learning approach for reconstructing reynolds stress modeling discrepancies based on dns data.Physical Review Fluids, 2(3):034603, 2017

    Jian-Xun Wang, Jin-Long Wu, and Heng Xiao. Physics-informed machine learning approach for reconstructing reynolds stress modeling discrepancies based on dns data.Physical Review Fluids, 2(3):034603, 2017

    work page 2017

  7. [7]

    Physics-informed machine learning approach for augmenting turbulence models: A comprehensive framework

    Jin-Long Wu, Heng Xiao, and Eric Paterson. Physics-informed machine learning approach for augmenting turbulence models: A comprehensive framework. Physical Review Fluids, 3(7):074602, 2018

    work page 2018

  8. [8]

    Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving non- linear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural net- works: A deep learning framework for solving forward and inverse problems involving non- linear partial differential equations. Journal of Computational physics, 378:686–707, 2019

    2019

Show all 75 references
  1. [9]

    Fourier neural operator for parametric partial dif- ferential equations

    Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, Andrew Stuart, and Anima Anandkumar. Fourier neural operator for parametric partial dif- ferential equations. arXiv preprint arXiv:2010.08895, 2020

    2010 arXiv

  2. [10]

    Learn- ing nonlinear operators via deeponet based on the universal approximation theorem of oper- ators

    Lu Lu, Pengzhan Jin, Guofei Pang, Zhongqiang Zhang, and George Em Karniadakis. Learn- ing nonlinear operators via deeponet based on the universal approximation theorem of oper- ators. Nature machine intelligence, 3(3):218–229, 2021

    2021

  3. [11]

    Machine learning–accelerated computational fluid dynamics

    Dmitrii Kochkov, Jamie A Smith, Ayya Alieva, Qing Wang, Michael P Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21):e2101784118, 2021

    2021

  4. [12]

    A causality-based learning approach for discovering the un- derlying dynamics of complex systems from partial observations with stochastic parameteri- zation

    Nan Chen and Yinling Zhang. A causality-based learning approach for discovering the un- derlying dynamics of complex systems from partial observations with stochastic parameteri- zation. Physica D: Nonlinear Phenomena, 449:133743, 2023

    2023

  5. [13]

    CEBoosting: Online sparse identification of dynamical systems with regime switching by causation entropy boosting

    Chuanqi Chen, Nan Chen, and Jin-Long Wu. CEBoosting: Online sparse identification of dynamical systems with regime switching by causation entropy boosting. Chaos: An Inter- disciplinary Journal of Nonlinear Science, 33(8), 2023. 26

    2023

  6. [14]

    CGNSDE: Conditional Gaussian neural stochastic differential equation for modeling complex systems and data assimilation

    Chuanqi Chen, Nan Chen, and Jin-Long Wu. CGNSDE: Conditional Gaussian neural stochastic differential equation for modeling complex systems and data assimilation. Com- puter Physics Communications, 304:109302, 2024

    2024

  7. [15]

    CGKN: A deep learning frame- work for modeling complex dynamical systems and efficient data assimilation

    Chuanqi Chen, Nan Chen, Yinling Zhang, and Jin-Long Wu. CGKN: A deep learning frame- work for modeling complex dynamical systems and efficient data assimilation. Journal of Computational Physics, page 113950, 2025

    2025

  8. [16]

    Modeling partially observed nonlinear dynamical systems and efficient data assimilation via discrete-time conditional gaussian koopman network

    Chuanqi Chen, Zhongrui Wang, Nan Chen, and Jin-Long Wu. Modeling partially observed nonlinear dynamical systems and efficient data assimilation via discrete-time conditional gaussian koopman network. Computer Methods in Applied Mechanics and Engineering , 445:118189, 2025

    2025

  9. [17]

    Atkinson and William G

    Anthony C. Atkinson and William G. Hunter. The design of experiments for parameter estimation. 10(2):271–289. Publisher: ASA Website eprint: https://www.tandfonline.com/doi/pdf/10.1080/00401706.1968.10490560

    1968

  10. [18]

    A. C. (Anthony Curtis) Atkinson. Optimum experimental designs . Oxford [England] : Clarendon Press ; New York : Oxford University Press, 1992

    1992

  11. [19]

    Roger E. Kirk. Experimental design. pages 23–45. Publisher: Sage London

  12. [20]

    Howard J. Seltman. Experimental design and analysis

  13. [21]

    Ryan, Christopher C

    Elizabeth G. Ryan, Christopher C. Drovandi, James M. McGree, and Anthony N. Pettitt. A Review of Modern Computational Algorithms for Bayesian Optimal Design. International Statistical Review, 84(1):128–154, April 2016

    2016

  14. [22]

    Bayes linear analysis for Bayesian optimal experimental design

    Matthew Jones, Michael Goldstein, Philip Jonathan, and David Randell. Bayes linear analysis for Bayesian optimal experimental design. Journal of Statistical Planning and Inference , 171:115–129, April 2016

    2016

  15. [23]

    Xun Huan and Youssef M. Marzouk. Simulation-based optimal Bayesian experimental design for nonlinear systems. Journal of Computational Physics, 232(1):288–317, January 2013

    2013

  16. [24]

    Ivanova, and Freddie Bickford Smith

    Tom Rainforth, Adam Foster, Desi R. Ivanova, and Freddie Bickford Smith. Modern Bayesian Experimental Design, November 2023. arXiv:2302.14545 [cs, stat] version: 2

    2023 arXiv

  17. [25]

    Applied Statistical Decision Theory

    Howard Raiffa and Robert Schlaifer. Applied Statistical Decision Theory . John Wiley & Sons, June 2000. Google-Books-ID: dJjyEAAAQBAJ

    2000

  18. [26]

    Morris H. DeGroot. Optimal Statistical Decisions . John Wiley & Sons, January 2005. Google-Books-ID: dtVieJ245z0C

    2005

  19. [27]

    D. V . (Dennis Victor) Lindley. Bayesian statistics : a review . Philadelphia : Society for Industrial and Applied Mathematics, 1972

    1972

  20. [28]

    Bayesian Experimental Design: A Review

    Kathryn Chaloner and Isabella Verdinelli. Bayesian Experimental Design: A Review. Statis- tical Science, 10(3):273–304, August 1995. Publisher: Institute of Mathematical Statistics. 27

    1995

  21. [29]

    Bayesian Decision Procedures for Dose De- termining Experiments

    John Whitehead and Hazel Brunier. Bayesian Decision Procedures for Dose De- termining Experiments. Statistics in Medicine , 14(9):885–893, 1995. eprint: https://onlinelibrary.wiley.com/doi/pdf/10.1002/sim.4780140904

    1995 doi

  22. [30]

    Cavagnaro, Jay I

    Daniel R. Cavagnaro, Jay I. Myung, Mark A. Pitt, and Janne V . Kujala. Adaptive Design Optimization: A Mutual Information-Based Approach to Model Discrimination in Cogni- tive Science. Neural Computation, 22(4):887–905, April 2010. Conference Name: Neural Computation

    2010

  23. [31]

    Drovandi, James M

    Christopher C. Drovandi, James M. McGree, and Anthony N. Pettitt. Sequential Monte Carlo for Bayesian sequentially designed experiments for discrete data.Computational Statistics & Data Analysis, 57(1):320–335, January 2013

    2013

  24. [32]

    Ivanova, Adam Foster, Steven Kleinegesse, Michael U

    Desi R. Ivanova, Adam Foster, Steven Kleinegesse, Michael U. Gutmann, and Tom Rainforth. Implicit Deep Adaptive Design: Policy-Based Experimental Design without Likelihoods, November 2021. arXiv:2111.02329 [cs, stat]

    2021 arXiv

  25. [33]

    Ivanova, Ilyas Malik, and Tom Rainforth

    Adam Foster, Desi R. Ivanova, Ilyas Malik, and Tom Rainforth. Deep Adaptive Design: Amortizing Sequential Bayesian Experimental Design, June 2021. arXiv:2103.02438 [cs, stat]

    2021 arXiv

  26. [34]

    Xun Huan and Youssef M. Marzouk. Gradient-based stochastic optimization methods in Bayesian experimental design. International Journal for Uncertainty Quantification , 4(6):479–510, 2014. arXiv:1212.2228 [math, stat]

    2014 arXiv

  27. [35]

    Wanggang Shen and Xun Huan. Bayesian Sequential Optimal Experimental Design for Non- linear Models Using Policy Gradient Reinforcement Learning.Computer Methods in Applied Mechanics and Engineering, 416:116304, November 2023. arXiv:2110.15335 [cs, stat]

    2023 arXiv

  28. [36]

    Optimal experimental design: Formula- tions and computations

    Xun Huan, Jayanth Jagalur, and Youssef Marzouk. Optimal experimental design: Formula- tions and computations. arXiv preprint arXiv:2407.16212, 2024

    2024 arXiv

  29. [37]

    Berry, Andy P

    Peter M ¨uller, Don A. Berry, Andy P. Grieve, Michael Smith, and Michael Krams. Simulation-based sequential Bayesian design. Journal of Statistical Planning and Inference, 137(10):3140–3150, October 2007

    2007

  30. [38]

    Bayesian inference in physics

    Udo von Toussaint. Bayesian inference in physics. Reviews of Modern Physics, 83(3):943– 999, September 2011. Publisher: American Physical Society

    2011

  31. [39]

    Numerical approaches for sequential Bayesian optimal experimental design

    Xun Huan. Numerical approaches for sequential Bayesian optimal experimental design . Thesis, Massachusetts Institute of Technology, 2015. Accepted: 2016-03-03T20:28:51Z

    2015

  32. [40]

    G.E.P. Box. Sequential Experimentation and Sequential Assembly of Designs . Report (Uni- versity of Wisconsin–Madison. Center for Quality and Productivity Improvement). Center for Quality and Productivity Improvement, University of Wisconsin-Madison, 1992

    1992

  33. [41]

    Drovandi, James M

    Christopher C. Drovandi, James M. McGree, and Anthony N. Pettitt. A Sequential Monte Carlo Algorithm to Incorporate Model Uncertainty in Bayesian Sequential Design. Journal of Computational and Graphical Statistics, January 2014. Publisher: Taylor & Francis. 28

    2014

  34. [42]

    Steven Kleinegesse, Christopher Drovandi, and Michael U. Gutmann. Sequential Bayesian Experimental Design for Implicit Models via Mutual Information, March 2020. arXiv:2003.09379 [cs, stat]

    2020 arXiv

  35. [43]

    Xun Huan and Youssef M. Marzouk. Sequential Bayesian optimal experimental design via approximate dynamic programming, April 2016. arXiv:1604.08320 [math, stat]

    2016 arXiv

  36. [44]

    Kennedy and Anthony O’Hagan

    Marc C. Kennedy and Anthony O’Hagan. Bayesian Calibration of Computer Models.Journal of the Royal Statistical Society Series B: Statistical Methodology, 63(3):425–464, September 2001

    2001

  37. [45]

    Inconsistency of Bayesian Inference for Misspeci- fied Linear Models, and a Proposal for Repairing It

    Peter Gr ¨unwald and Thijs van Ommen. Inconsistency of Bayesian Inference for Misspeci- fied Linear Models, and a Proposal for Repairing It. Bayesian Analysis, 12(4):1069–1103, December 2017. Publisher: International Society for Bayesian Analysis

    2017

  38. [46]

    Learning about physical parameters: the impor- tance of model discrepancy

    Jenn ´y Brynjarsd´ottir and Anthony O’Hagan. Learning about physical parameters: the impor- tance of model discrepancy. 30(11):114007. Publisher: IOP Publishing

  39. [47]

    Catanach and Niladri Das

    Tommie A. Catanach and Niladri Das. Metrics for Bayesian Optimal Experiment Design under Model Misspecification, April 2023. arXiv:2304.07949 [cs, eess, stat]

    2023 arXiv

  40. [48]

    Optimal Bayesian experimental design in the presence of model error

    Chi Feng. Optimal Bayesian experimental design in the presence of model error . Thesis, Massachusetts Institute of Technology, 2015. Accepted: 2015-07-17T19:46:48Z

    2015

  41. [49]

    ¨Ozge S¨urer, Matthew Plumlee, and Stefan M. Wild. Sequential Bayesian Experimental De- sign for Calibration of Expensive Simulation Models. Technometrics, 66(2):157–171, April

  42. [50]

    Brecht M. R. Donckels, Dirk J. W. De Pauw, Bernard De Baets, Jo Maertens, and Peter A. Vanrolleghem. An anticipatory approach to optimal experimental design for model discrimi- nation. Chemometrics and Intelligent Laboratory Systems, 95(1):53–63, January 2009

    2009

  43. [51]

    Stefan Streif, Felix Petzke, Ali Mesbah, Rolf Findeisen, and Richard D. Braatz. Optimal Experimental Design for Probabilistic Model Discrimination Using Polynomial Chaos.IFAC Proceedings Volumes, 47(3):4103–4109, January 2014

    2014

  44. [52]

    Iglesias, Kody J

    Marco A. Iglesias, Kody J. H. Law, and Andrew M. Stuart. Ensemble Kalman methods for inverse problems. Inverse Problems, 29(4):045001, March 2013. Publisher: IOP Publishing

    2013

  45. [53]

    Kovachki and Andrew M

    Nikola B. Kovachki and Andrew M. Stuart. Ensemble Kalman inversion: a derivative-free technique for machine learning tasks. Inverse Problems, 35(9):095005, August 2019. Pub- lisher: IOP Publishing

    2019

  46. [54]

    Learning about struc- tural errors in models of complex dynamical systems

    Jin-Long Wu, Matthew E Levine, Tapio Schneider, and Andrew Stuart. Learning about struc- tural errors in models of complex dynamical systems. Journal of Computational Physics , page 113157, 2024

    2024

  47. [55]

    Neural dynamical operator: Continuous spatial-temporal model with gradient-based and derivative-free optimization methods

    Chuanqi Chen and Jin-Long Wu. Neural dynamical operator: Continuous spatial-temporal model with gradient-based and derivative-free optimization methods. Journal of Computa- tional Physics, 520:113480, 2025. 29

    2025

  48. [56]

    Data-driven stochastic closure modeling via conditional diffusion model and neural operator

    Xinghao Dong, Chuanqi Chen, and Jin-Long Wu. Data-driven stochastic closure modeling via conditional diffusion model and neural operator. arXiv preprint arXiv:2408.02965, 2024

    2024 arXiv

  49. [57]

    Stochastic and non-local closure modeling for nonlinear dynamical systems via latent score-based generative models

    Xinghao Dong, Huchen Yang, and Jin-Long Wu. Stochastic and non-local closure modeling for nonlinear dynamical systems via latent score-based generative models. arXiv preprint arXiv:2506.20771, 2025

    2025 arXiv

  50. [58]

    Lorentzen, and Tuhin Bhakta

    Xiaodong Luo, Rolf J. Lorentzen, and Tuhin Bhakta. Accounting for model errors of rock physics models in 4d seismic history matching problems: A perspective of machine learning. 196:107961

  51. [59]

    D. V . Lindley. On a Measure of the Information Provided by an Experiment. The Annals of Mathematical Statistics, 27(4):986–1005, December 1956. Publisher: Institute of Mathemat- ical Statistics

    1956

  52. [60]

    Michele Zemplenyi and Jeffrey W. Miller. Bayesian optimal experimental design for inferring causal structure

  53. [61]

    Closed-loop model-based design of experiments for kinetic model dis- crimination and parameter estimation: Benzoic acid esterification on a heterogeneous cata- lyst

    Conor Waldron, Arun Pankajakshan, Marco Quaglio, Enhong Cao, Federico Galvanin, and Asterios Gavriilidis. Closed-loop model-based design of experiments for kinetic model dis- crimination and parameter estimation: Benzoic acid esterification on a heterogeneous cata- lyst. 58(49...

  54. [62]

    Bernal Neira, and Alexander W

    Jialu Wang, Zedong Peng, Ryan Hughes, Debangsu Bhattacharyya, David E. Bernal Neira, and Alexander W. Dowling. Measure this, not that: Optimizing the cost and model-based information content of measurements. 189:108786

  55. [63]

    On automatic differentiation

    Andreas Griewank. On automatic differentiation. 1988

    1988

  56. [64]

    Compiling machine learning pro- grams via high-level tracing

    Roy Frostig, Matthew James Johnson, and Chris Leary. Compiling machine learning pro- grams via high-level tracing

  57. [65]

    Methods of Mathematical Physics

    David Hilbert. Methods of Mathematical Physics . CUP Archive, 1985. Google-Books-ID: hyM9AAAAIAAJ

    1985

  58. [66]

    A tutorial on the adjoint method for inverse problems

    Dan Givoli. A tutorial on the adjoint method for inverse problems. Computer Methods in Applied Mechanics and Engineering, 380:113810, July 2021

    2021

  59. [67]

    Adjoint sensitivity analysis for differential- algebraic equations: algorithms and software

    Yang Cao, Shengtai Li, and Linda Petzold. Adjoint sensitivity analysis for differential- algebraic equations: algorithms and software. 149(1):171–191

  60. [68]

    JAX: composable transformations of Python+NumPy programs, 2018

    James Bradbury, Roy Frostig, Peter Hawkins, Matthew James Johnson, Chris Leary, Dougal Maclaurin, George Necula, Adam Paszke, Jake VanderPlas, Skye Wanderman-Milne, and Qiao Zhang. JAX: composable transformations of Python+NumPy programs, 2018

    2018

  61. [69]

    Smith, Ayya Alieva, Qing Wang, Michael P

    Dmitrii Kochkov, Jamie A. Smith, Ayya Alieva, Qing Wang, Michael P. Brenner, and Stephan Hoyer. Machine learning–accelerated computational fluid dynamics. Proceedings of the National Academy of Sciences, 118(21), 2021. 30

    2021

  62. [70]

    Paola Sebastiani and Henry P. Wynn. Maximum Entropy Sampling and Optimal Bayesian Experimental Design. Journal of the Royal Statistical Society Series B: Statistical Methodol- ogy, 62(1):145–157, January 2000

    2000

  63. [71]

    Methods of Mathematical Physics, Vol

    Richard Courant and David Hilbert. Methods of Mathematical Physics, Vol. I. Physics Today, 7(5):17–17, May 1954

    1954

  64. [72]

    Atkinson

    Weimin Han and Kendall E. Atkinson. Theoretical Numerical Analysis: A Functional Analy- sis Framework, volume 39 of Texts in Applied Mathematics. Springer, New York, NY , 2009

    2009

  65. [73]

    Large-scale bayesian optimal experimental design with derivative-informed projected neural network

    Keyi Wu, Thomas O’Leary-Roseberry, Peng Chen, and Omar Ghattas. Large-scale bayesian optimal experimental design with derivative-informed projected neural network. 95(1):30

  66. [74]

    Thomas O’Leary-Roseberry, Xiaosong Du, Anirban Chaudhuri, Joaquim R. R. A. Martins, Karen Willcox, and Omar Ghattas. Learning high-dimensional parametric maps via reduced basis adaptive residual networks. 402:115730. 31 Appendix A. Theoretical global maximum The model discrepa...

  67. [2024]

    Publisher: Taylor & Francis eprint: https://doi.org/10.1080/00401706.2023.2246157

    2023

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.