Learning task-specific predictive models for scientific computing

Jianyuan Yin; Qianxiao Li

REVIEW 4 major objections 5 minor 41 references

Learning task-specific predictive models for scientific computing

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A path-focused loss beats plain MSE for scientific computing tasks.

desk verdict Useful reformulation with a real theory gap: the bound is proven for direct simulation but not for the optimization and string-method examples where support is defined too narrowly. read the letter →

arxiv 2506.03835 v1 pith:HX7DYQBF submitted 2025-06-04 cs.LG

Jianyuan Yin , Qianxiao Li This is my paper

classification cs.LG MSC 68T0762J0265K1065G9990C30

keywords task-specificlearningsurrogatemodelsscientificcomputingdistributionshiftalgorithmsupportreweightedempiricalriskmaximumpredictionerrormeansquared

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that when a learned surrogate model is used inside a scientific computing algorithm, the usual mean-squared-error training is not necessarily the right objective. The true goal is to make the algorithm's output accurate, which is controlled by the maximum prediction error on the set of points where the algorithm evaluates the surrogate. The authors propose to minimize a reweighted empirical risk that approximates this maximum error, using a change of measure to the training data distribution. They show numerically on trajectory prediction, optimal control, and minimum energy path problems that this task-specific training reduces the algorithmic output error compared to MSE-trained models.

What carries the argument

The central object is the algorithm support S(f) = [s1(f), …, sJ(f)]: the finite sequence of points at which the downstream algorithm evaluates a surrogate f. The key identity is the Lipschitz-based bound JA(θ) ≲ max_{x ∈ S(fθ)} ||fθ(x) − f*(x)||², which reduces the downstream task error to a maximum prediction error on this support. To minimize this bound without resampling data, the paper constructs a reweighted empirical risk RN(θ) = (1/N)Σ_i mθ(xi) ||fθ(xi) − yi||², where the weights mθ estimate the Radon–Nikodym derivative dμθ/dρ through kernel density estimation, a stratification factor for uneven support points, and a softmax weighting that approximates the supremum. The mechanism is the iterative update of these weights as the surrogate's support evolves, allowing the training distribution to shift toward the task distribution.

What would settle it

Construct a one-dimensional Euler integration example where the ground-truth algorithm support lies entirely in a region with zero training data (e.g., the initial point and trajectory are far from the training data support). If the iterative reweighting algorithm fails to improve the trajectory error beyond MSE, or if the reweighting weights all collapse to near zero, then Assumption 2.4 is violated and the method's central premise does not hold.

Watch

Extended reading notes

Core claim

The central discovery is that the error of a downstream scientific computing algorithm that uses a surrogate model can be bounded, up to a constant, by the maximum prediction error of the surrogate on the algorithm's own support — the set of input points the algorithm queries. This bound holds by Lipschitz continuity of the algorithm and the ground-truth function. Consequently, minimizing this maximum error, rather than the mean squared error over the training distribution, is the correct surrogate for the task. The paper realizes this by formulating a task-specific population risk with a Radon–Nikodym derivative that reweights training samples toward the algorithm support, and by developing an iterative algorithm that alternates between updating the surrogate and recomputing the reweighting coefficients via kernel density estimation. Numerical experiments on three illustrative problems show that the task-specific surrogate consistently achieves lower algorithm output error than the MSE-trained surrogate, with larger gains when the training distribution and the algorithm support are more mismatched.

Load-bearing premise

The entire method assumes that the algorithm support of every surrogate model encountered during training stays inside the region where training data are available; otherwise the reweighted risk estimate becomes meaningless extrapolation.

Editorial extensions

If this is right

MSE-trained surrogates are suboptimal for any downstream task whose algorithm does not evaluate the model uniformly over the training distribution; task-specific training should be preferred when the algorithm support is known.
The method applies broadly to scientific computing tasks such as trajectory prediction, optimal control, and minimum energy path calculations, where the algorithm support is determined by the physics and can be simulated.
When the training data already concentrate near the algorithm support, the advantage of task-specific training over MSE shrinks, but it grows with distribution shift.
Larger hypothesis spaces can already achieve low MSE, reducing the benefit of task-specific reweighting, while constrained hypothesis spaces benefit substantially.
The proposed iterative algorithm can be implemented with standard optimization libraries and does not require differentiating through the downstream algorithm.

Reading between the lines

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

This perspective suggests a new criterion for active learning in scientific computing: acquire new training points near the current algorithm support, mirroring the reweighting weights, which could accelerate convergence of iterative algorithms.
The bound JA ≲ RS could be tightened by exploiting more structure of the algorithm, e.g., contraction rates for stable integrators, possibly leading to sharper task-specific objectives.
For stochastic algorithms or noisy observations, the deterministic bound may extend to an expected-error bound, but the paper does not address this; testing such an extension on a control problem with measurement noise would be a natural next step.
The method's reliance on kernel density estimation implies a dimension-dependent sample complexity; for high-dimensional inputs, alternative density estimators or learned weight mappings may be needed.

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.

Desk Editor's Note

read the letter

The core idea is sensible and the Euler/trajectory experiment is solid, but the two inverse-problem examples have a load-bearing mismatch between the theory and the actual algorithm support. I would conditionally accept if that is fixed or the claims are scoped down.

What is new: showing MSE is suboptimal for downstream tasks is not new (decision-focused learning, end-to-end learning), but the specific construction—bounding task error by max prediction error on the algorithm support and then doing KDE-based reweighting without differentiating through the algorithm—is a distinctive, practical combination. The upper bound in Prop 3.1 is a straightforward Lipschitz argument and it is correct for algorithms whose output is a direct function of evaluations on a well-defined support, like the explicit Euler trajectory example. The experiments are honestly done: 20 seeds, ablations on distribution shift and hypothesis space, and code is on GitHub. Credit is due for that.

Now the soft spots. The biggest one is real. In §4.2 the tracking problem defines the support as the single trajectory under the optimal control uθ, and in §4.3 as the final convergent string. But the control optimizer and the string iteration evaluate the surrogate at many other points before arriving at that output. Those off-support values influence uθ and therefore influence A(fθ). Changing fθ off the claimed support can change the output, so the bound in Prop 3.1, which controls JA by the max error on S(fθ), does not cover these two experiments. The paper's own remark that 'the states visited by the optimization process need not be included in the algorithm support' is easy to read as wishful thinking. The experiments may still show real improvement, but not because of the stated theorem.

Second, Assumption 2.4 (algorithm support of every surrogate stays inside supp ρ) is load-bearing and no mechanism detects or recovers from drift. For iterative algorithms, drift is plausible. The reweighting weights degrade to noise if the support leaves the data. The paper acknowledges the need for a 'not fundamentally incorrect' initialization but nothing more.

Third, the empirical risk has several free kernels, softmax scale, regularizer, early stopping tolerance, without sensitivity analysis or convergence guarantees. That is minor for a methods paper, but it is worth saying.

The paper does not cite decision-focused learning (e.g., learning to optimize, end-to-end surrogates). That is an omission, not fraud, and the specific reweighting idea is still theirs.

Who is this for: people building surrogate models for simulation and control, especially ML force fields. A serious referee should see it. The core flag is the support mismatch in the inverse-problem sections. My recommendation: send to peer review, but the authors should either extend the bound to include all queried points, or present the tracking/string results as heuristic with scoped claims.

Referee Report

4 major / 5 minor

Summary. The paper considers the problem of learning a surrogate model fθ to be used inside a downstream scientific computing algorithm A, where the goal is to minimize the algorithm output error JA(θ) rather than the prediction MSE. The authors prove an upper bound (Proposition 3.1) showing that, under an assumption that A only depends on fθ through evaluations on a finite algorithm support S(fθ), JA is controlled by the maximum prediction error on that support. They then convert this bound into a reweighted empirical risk, using kernel density estimates and softmax weights, and propose an iterative training algorithm. Numerical experiments on Lorenz trajectory prediction, a cart-pendulum optimal control tracking problem, and minimum energy path (MEP) computation report improved JA relative to MSE-trained surrogates, together with ablations on distribution shift and hypothesis space size.

Significance. If the proposed bound and algorithm were valid for all three examples, the paper would provide a principled alternative to MSE training in machine-learning-augmented scientific computing. The formal Lipschitz estimate in Proposition 3.1 is correct under Assumption 2.1, and the numerical evaluation is substantial: 20 random seeds per experiment, comparison against external downstream metrics, and ablation studies. The method's empirical promise is real and the paper ships code. However, the tracking and MEP examples violate the support assumption on which the central bound rests, so the current manuscript does not actually establish the stated theoretical justification for two of its three main experiments. The heuristic derivation from the population risk to the finite-sample reweighted objective also lacks any consistency or error analysis. The work is therefore best viewed as proposing a plausible and partially validated heuristic, rather than as a fully supported general framework as currently claimed.

major comments (4)

[Sec. 4.2, Eqs. (4.7)-(4.9)] The tracking problem does not satisfy Assumption 2.1. The output uθ is defined as the minimizer of Cτ(u; fθ), so the optimization process evaluates fθ along many candidate trajectories; these evaluations determine uθ and therefore the reported JA. The algorithm support in Eq. (4.8) includes only the final Euler trajectory, and the text explicitly states that the states visited by the optimization process need not be included. Changing fθ off that support can change uθ and JA without changing RS, so Proposition 3.1 does not control JA(θ) for this example. The improvements in Fig. 3 are therefore not backed by the paper's core bound.
[Sec. 4.3, MEP calculation] The same support mismatch occurs in the string method example. A(fθ) is the limit of an iterative string evolution, during which fθ is evaluated at intermediate strings; these evaluations determine the final nodes. Setting S(fθ) to the final convergent string only excludes exactly the evaluations that determine the output. Consequently, the MEP experiment lies outside the scope of Proposition 3.1, and the connection between Eq. (3.3) and Fig. 4 is missing.
[Assumption 2.4 and Sec. 3.3] The change of measure in Eq. (3.6) and the KDE reweighting weights in Eq. (3.13) require S(fθ)⊂supp ρ. This is assumed, not ensured. The algorithm starts from an MSE-initialized model and iterates; if the support drifts outside the training data support, the kernel weights vanish and the empirical risk (3.14) becomes uninformative even after normalization. The paper notes the vanishing-weight failure mode but provides no detection or recovery mechanism. Given that the tracking and string iterations can explore regions away from the initial support, this is a practical gap in the proposed method.
[Sec. 3.3, Eqs. (3.5)-(3.14)] The derivation relating the population target RS to the finite-sample objective is heuristic: the supremum in (3.4) is replaced by a softmax-weighted KDE mixture, the error lθ in (3.11) is itself estimated from training data, and four kernel widths plus M and ω0 are selected without convergence statements. There is no finite-sample result connecting a minimizer of RN to a small RS. The paper should either provide a consistency analysis or explicitly label this part as a heuristic approximation; the current presentation of (3.14) as the discretized version of (3.7) overstates the formal support.

minor comments (5)

[Throughout] Cross-references are inconsistent: 'Theorem 2.3', 'Theorem 2.5', 'Theorem 2.4', and 'Theorem 2.1' should refer to Example 2.3, Example 2.5, Assumption 2.4, and Assumption 2.1, respectively.
[Supplementary Materials] The sentence 'For example. with an overly large M' contains a punctuation typo (period after 'example').
[Fig. 2(B) and 2(C)] The caption does not define M1 and M2; the definitions appear only in the text. A sentence in the caption would help readability.
[Sec. 3.4, Algorithm 3.1] Algorithm 3.1's output line refers to RS,N, but the definition of RS,N appears after the algorithm box in Eq. (3.17). Consider moving Eq. (3.17) before the algorithm.
[Sec. 3.3, Eq. (3.10)] The expression dκ(·,x̃j)/dρ is used before discussing the existence of the density; a brief note that the kernel is assumed absolutely continuous with respect to ρ on the relevant region would avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central bound is a genuine Lipschitz estimate and the reported gains are checked against ground-truth task errors.

full rationale

The derivation chain is self-contained. Proposition 3.1 derives Eq. 3.2, an inequality bounding the algorithm-support deviation and prediction error in terms of max_{x in S(f)} ||f(x)-f*(x)||, using Lipschitz continuity of S_j, f*, and A under Assumption 2.1. This is an upper bound, not an identity, so minimizing RS in Eq. 3.3 is not equivalent by construction to minimizing JA. The empirical risk RN in Eq. 3.14 reweights training data through KDE estimates in Eqs. 3.8-3.13 of the Radon-Nikodym derivative dμθ/dρ; this is an approximation whose validity depends on Assumption 2.4 but does not redefine the reported quantity. The numerical claims are evaluated externally: JA(θ) is computed with ground-truth f* in Eqs. 4.2, 4.9 and 4.11, so the reported ratios JA(θTS)/JA(θMSE) are independent benchmarks rather than fitted values. The only self-referential element is that both the reweighting coefficients mθ in Eq. 3.13 and the early-stopping statistic RS,N in Eq. 3.17 are built from the same kernel error estimates lθ in Eq. 3.11; this is internal algorithmic consistency, not a reduction of the paper's conclusion to its inputs. Section 4.2 explicitly states that "the states visited by the optimization process need not be included in the algorithm support"; this is a potential validity gap for applying Proposition 3.1 to the tracking and MEP experiments, but it is a correctness concern about the scope of the bound, not a circularity. No load-bearing self-citation was found: the borrowed parameterization in Eq. 4.14 is attributed to [35], but the central task-specific learning framework and experiments do not reduce to that citation. Accordingly, no circular step is identified.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim does not introduce new physical or mathematical entities. The free parameters are the kernel variances and reweighting hyperparameters, which are chosen per example and partly fit to the training data. The axioms are the standard Lipschitz regularity assumptions and the critical support-inclusion assumption (2.4), which is load-bearing for the change of measure.

free parameters (7)

Kernel variance in (3.8) for approximating sampling measure = 1.0 (Lorenz), 0.01 (tracking), 0.01 (MEP)
Chosen per example; affects the estimated density ρN and hence the reweighting coefficients. The paper states the method is not sensitive to this choice.
Kernel variance in (3.9) for stratifying algorithm support = 2.0 (Lorenz), 0.01 (tracking), 0.001 (MEP)
Chosen per example; affects stratification factors αθ.
Kernel variance in (3.11) for estimating prediction errors = 1.0 (Lorenz), 0.01 (tracking), 0.001 (MEP); chosen by minimizing (S.2) on training data
Directly shapes the early stopping criterion RS,N and the weights ωθ; tuned on the training data itself, so it is effectively a fitted parameter.
Kernel variance in (3.13) for approximating RN derivatives = 1.0 (Lorenz), 0.05 (tracking), 0.01 (MEP)
Determines the neighborhood around the algorithm support where the surrogate is trained to be accurate; chosen per example.
Softmax scale M = 10
Balances mean and supremum error in weights ωθ; paper states not sensitive but values affect conditioning.
Regularizer ω0 = 0.5
Minimal weight for each support point; paper states not sensitive to M or ω0.
Early stopping tolerance = 1e-6 (standard error of RS,N over last 10 iterations)
Termination criterion; chosen for all experiments.

assumptions (5)

domain assumption The downstream task algorithm A evaluates f only at a finite sequence S(f) of points, and A, Sj, and f* are Lipschitz continuous.
Assumption 2.1; needed for the upper bound in Proposition 3.1.
domain assumption The support of the ground truth S(f*) and the support of every surrogate encountered during training are contained in supp ρ, the support of the training measure.
Assumption 2.4; required for the Radon-Nikodym change of measure in (3.6) and for the KDE-based weights mθ to be meaningful.
standard math The kernel κ has bounded support and small variance, so the Radon-Nikodym derivative dµθ/dρ exists.
Invoked before (3.6); relies on the Radon-Nikodym theorem.
domain assumption The hypothesis space F and optimizer can minimize the reweighted empirical risk LN(θ; m) sufficiently well.
The iterative algorithm assumes each inner optimization step makes progress; no convergence guarantee is given.
domain assumption Training labels are noiseless, y_i = f*(x_i).
Assumed in Section 2; the method is not analyzed for noisy data.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning task-specific predictive models for scientific computing." pith.science (2026). https://pith.science/paper/HX7DYQBF

@misc{pith2026250603835,
  author       = {Pith},
  title        = {Pith review of: Learning task-specific predictive models for scientific computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HX7DYQBF}},
  note         = {Machine review of arXiv:2506.03835}
}

read the original abstract

We consider learning a predictive model to be subsequently used for a given downstream task (described by an algorithm) that requires access to the model evaluation. This task need not be prediction, and this situation is frequently encountered in machine-learning-augmented scientific computing. We show that this setting differs from classical supervised learning, and in general it cannot be solved by minimizing the mean square error of the model predictions as is frequently performed in the literature. Instead, we find that the maximum prediction error on the support of the downstream task algorithm can serve as an effective estimate for the subsequent task performance. With this insight, we formulate a task-specific supervised learning problem based on the given sampling measure, whose solution serves as a reliable surrogate model for the downstream task. Then, we discretize the empirical risk based on training data, and develop an iterative algorithm to solve the task-specific supervised learning problem. Three illustrative numerical examples on trajectory prediction, optimal control and minimum energy path computation demonstrate the effectiveness of the approach.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 37 canonical work pages

[1]

Abadi, P

M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, et al. , TensorFlow : A system for Large-Scale machine learning , in 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), 2016, pp. 265--283

work page 2016
[2]

Agarwal, A

A. Agarwal, A. Anandkumar, P. Jain, and P. Netrapalli , Learning sparsely used overcomplete dictionaries via alternating minimization , SIAM J. Optim., 26 (2016), pp. 2775--2799

work page 2016
[3]

H. K. Aggarwal, M. P. Mani, and M. Jacob , MoDL : Model-based deep learning architecture for inverse problems , IEEE Trans. Med. Imaging, 38 (2018), pp. 394--405

work page 2018
[4]

J. An, L. Ying, and Y. Zhu , Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients , in International Conference on Learning Representations, 2021

work page 2021
[5]

D. Bash, Y. Cai, V. Chellappan, S. L. Wong, X. Yang, P. Kumar, J. D. Tan, A. Abutaha, J. J. Cheng, Y.-F. Lim, S. I. P. Tian, Z. Ren, F. Mekki-Berrada , W. K. Wong, J. Xie, J. Kumar, S. A. Khan, Q. Li, T. Buonassisi, and K. Hippalgaonkar , Multi- Fidelity High-Throughput Optimization of Electrical Conductivity in P3HT-CNT Composites , Advanced Functional M...

work page 2021
[6]

Chang, A

T. Chang, A. Gillette, and R. Maulik , Leveraging interpolation models and error bounds for verifiable scientific machine learning , J. Comput. Phys., 524 (2025), p. 113726

work page 2025
[7]

u tt, and K.-R. M \

S. Chmiela, A. Tkatchenko, H. E. Sauceda, I. Poltavsky, K. T. Sch \"u tt, and K.-R. M \"u ller , Machine learning of accurate energy-conserving molecular force fields , Sci. Adv., 3 (2017), p. e1603015

work page 2017
[8]

Clevert, T

D.-A. Clevert, T. Unterthiner, and S. Hochreiter , Fast and accurate deep network learning by exponential linear units ( ELUs ) , arXiv preprint arXiv:1511.07289, (2015)

arXiv 2015

Show all 41 references

[9]

De Florio, I

M. De Florio, I. G. Kevrekidis, and G. E. Karniadakis , AI-Lorenz : A physics-data-driven framework for black-box and gray-box identification of chaotic systems with symbolic regression , Chaos, Soliton Fract., 188 (2024), p. 115538

2024
[10]

Doerr, M

S. Doerr, M. Majewski, A. P \'e rez, A. Kramer, C. Clementi, F. Noe, T. Giorgino, and G. De Fabritiis , TorchMD : A deep learning framework for molecular simulations , J. Chem. Theory Comput., 17 (2021), pp. 2355--2363

2021
[11]

W. E, W. Ren, and E. Vanden-Eijnden , String method for the study of rare events , Phys. Rev. B, 66 (2002), p. 052301

2002
[12]

height 2pt depth -1.6pt width 23pt, Simplified and improved string method for computing the minimum energy paths in barrier-crossing events , J. Chem. Phys., 126 (2007), p. 164103

2007
[13]

W. E and E. Vanden-Eijnden , Transition-path theory and path-finding algorithms for the study of rare events , Annu. Rev. Phys. Chem., 61 (2010), pp. 391--420

2010
[14]

D. G. Giovanis, D. Loukrezis, I. G. Kevrekidis, and M. D. Shields , Polynomial chaos expansions on principal geodesic grassmannian submanifolds for surrogate modeling and uncertainty quantification , J. Comput. Phys., 519 (2024), p. 113443

2024
[15]

Gr \"u ne and J

L. Gr \"u ne and J. Pannek , Nonlinear Model Predictive Control: Theory and Algorithms , Springer, 2017

2017
[16]

Hairer and G

E. Hairer and G. Wanner , Solving Ordinary Differential Equations II: Stiff and Differential-Algebraic Problems , vol. 375, Springer, Berlin, Heidelberg, 1996

1996
[17]

K. He, X. Zhang, S. Ren, and J. Sun , Deep residual learning for image recognition , in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 770--778

2016
[18]

G. E. Karniadakis, I. G. Kevrekidis, L. Lu, P. Perdikaris, S. Wang, and L. Yang , Physics-informed machine learning , Nat. Rev. Phys., 3 (2021), pp. 422--440

2021
[19]

B. Lin, Q. Li, and W. Ren , A data driven method for computing quasipotentials , in Proceedings of the 2nd Mathematical and Scientific Machine Learning Conference, J. Bruna, J. Hesthaven, and L. Zdeborova, eds., vol. 145 of Proceedings of Machine Learning Research, PMLR, 2022,...

2022
[20]

E. N. Lorenz , Deterministic nonperiodic flow , J. Atmos. Sci., 20 (1963), pp. 130--141

1963
[21]

K. O. Lye, S. Mishra, D. Ray, and P. Chandrashekar , Iterative surrogate model optimization (ISMO) : An active learning algorithm for PDE constrained optimization with deep neural networks , Comput. Methods Appl. Mech. Eng., 374 (2021), p. 113575

2021
[22]

Mekki-Berrada , Z

F. Mekki-Berrada , Z. Ren, T. Huang, W. K. Wong, F. Zheng, J. Xie, I. P. S. Tian, S. Jayavelu, Z. Mahfoud, D. Bash, K. Hippalgaonkar, S. Khan, T. Buonassisi, Q. Li, and X. Wang , Two-step machine learning enables optimized nanoparticle synthesis , npj Computational Materials, ...

2021
[23]

Mowlavi and S

S. Mowlavi and S. Nabi , Optimal control of PDEs using physics-informed neural networks , J. Comput. Phys., 473 (2023), p. 111731

2023
[24]

Pestourie, Y

R. Pestourie, Y. Mroueh, T. V. Nguyen, P. Das, and S. G. Johnson , Active learning of deep surrogates for PDEs : application to metasurface design , npj Comput. Mater., 6 (2020), p. 164

2020
[25]

Raissi, P

M. Raissi, P. Perdikaris, and G. E. Karniadakis , Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , J. Comput. Phys., 378 (2019), pp. 686--707

2019
[26]

M. Ren, W. Zeng, B. Yang, and R. Urtasun , Learning to reweight examples for robust deep learning , in Proceedings of the 35th International Conference on Machine Learning, J. Dy and A. Krause, eds., vol. 80 of Proceedings of Machine Learning Research, PMLR, 2018, pp. 4334--4343

2018
[27]

Seiffert, T

C. Seiffert, T. M. Khoshgoftaar, J. Van Hulse, and A. Napolitano , Resampling or reweighting: A comparison of boosting implementations , in 2008 20th IEEE International Conference on Tools with Artificial Intelligence, vol. 1, IEEE, 2008, pp. 445--451

2008
[28]

Shaham, Y

U. Shaham, Y. Yamada, and S. Negahban , Understanding adversarial training: Increasing local stability of supervised models through robust optimization , Neurocomput., 307 (2018), pp. 195--204

2018
[29]

Shankar, D

V. Shankar, D. Chakraborty, V. Viswanathan, and R. Maulik , Differentiable turbulence: Closure as a partial differential equation constrained optimization , Phys. Rev. Fluids, 10 (2025), p. 024605

2025
[30]

Z. Shen, P. Cui, T. Zhang, and K. Kunag , Stable learning via sample reweighting , Proc. AAAI Conf. Artif. Intell., 34 (2020), pp. 5692--5699

2020
[31]

Tramer and D

F. Tramer and D. Boneh , Adversarial training and robustness for multiple perturbations , in Advances in Neural Information Processing Systems, H. Wallach, H. Larochelle, A. Beygelzimer, F. d Alch\' e -Buc, E. Fox, and R. Garnett, eds., vol. 32, Curran Associates, Inc., 2019

2019
[32]

u tt, A. Tkatchenko, and K.-R. M \

O. T. Unke, S. Chmiela, H. E. Sauceda, M. Gastegger, I. Poltavsky, K. T. Sch \"u tt, A. Tkatchenko, and K.-R. M \"u ller , Machine learning force fields , Chem. Rev., 121 (2021), pp. 10142--10186

2021
[33]

J. Wang, T. Zhang, S. Liu, P.-Y. Chen, J. Xu, M. Fardad, and B. Li , Adversarial attack generation empowered by min-max optimization , in Advances in Neural Information Processing Systems, M. Ranzato, A. Beygelzimer, Y. Dauphin, P. Liang, and J. W. Vaughan, eds., vol. 34, Curr...

2021
[34]

Yang and Y

J. Yang and Y. Zhang , Alternating direction algorithms for _1 -problems in compressive sensing , SIAM J. Sci. Comput., 33 (2011), pp. 250--278

2011
[35]

H. Yu, X. Tian, W. E, and Q. Li , OnsagerNet : Learning stable and interpretable dynamics using a generalized onsager principle , Phys. Rev. Fluids, 6 (2021), p. 114402

2021
[36]

Zhang, H

D. Zhang, H. Bi, F.-Z. Dai, W. Jiang, X. Liu, L. Zhang, and H. Wang , Pretraining of attention-based deep learning potential model for molecular simulation , npj Comput. Mater., 10 (2024), p. 94

2024
[37]

Zhang, J

L. Zhang, J. Han, H. Wang, R. Car, and W. E , Deep potential molecular dynamics: a scalable model with the accuracy of quantum mechanics , Phys. Rev. Lett., 120 (2018), p. 143001

2018
[38]

Zhang, D.-Y

L. Zhang, D.-Y. Lin, H. Wang, R. Car, and W. E , Active learning of uniformly accurate interatomic potentials for materials simulation , Phys. Rev. Mater., 3 (2019), p. 023804

2019
[39]

Zhang, J

X. Zhang, J. Long, W. Hu, W. E, and J. Han , Initial value problem enhanced sampling for closed-loop optimal control design with deep neural networks , arXiv preprint arXiv:2209.04078, (2022)

2022
[40]

Zhao and Q

J. Zhao and Q. Li , Mitigating distribution shift in machine learning--augmented hybrid simulation , SIAM J. Sci. Comput., 47 (2025), pp. C475--C500

2025
[41]

D. P. Kingma , Adam: A method for stochastic optimization , arXiv preprint arXiv:1412.6980, (2014)

2014 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Abadi, P

M. Abadi, P. Barham, J. Chen, Z. Chen, A. Davis, J. Dean, M. Devin, S. Ghemawat, et al. , TensorFlow : A system for Large-Scale machine learning , in 12th USENIX Symposium on Operating Systems Design and Implementation (OSDI 16), 2016, pp. 265--283

work page 2016

[2] [2]

Agarwal, A

A. Agarwal, A. Anandkumar, P. Jain, and P. Netrapalli , Learning sparsely used overcomplete dictionaries via alternating minimization , SIAM J. Optim., 26 (2016), pp. 2775--2799

work page 2016

[3] [3]

H. K. Aggarwal, M. P. Mani, and M. Jacob , MoDL : Model-based deep learning architecture for inverse problems , IEEE Trans. Med. Imaging, 38 (2018), pp. 394--405

work page 2018

[4] [4]

J. An, L. Ying, and Y. Zhu , Why resampling outperforms reweighting for correcting sampling bias with stochastic gradients , in International Conference on Learning Representations, 2021

work page 2021

[5] [5]

D. Bash, Y. Cai, V. Chellappan, S. L. Wong, X. Yang, P. Kumar, J. D. Tan, A. Abutaha, J. J. Cheng, Y.-F. Lim, S. I. P. Tian, Z. Ren, F. Mekki-Berrada , W. K. Wong, J. Xie, J. Kumar, S. A. Khan, Q. Li, T. Buonassisi, and K. Hippalgaonkar , Multi- Fidelity High-Throughput Optimization of Electrical Conductivity in P3HT-CNT Composites , Advanced Functional M...

work page 2021

[6] [6]

Chang, A

T. Chang, A. Gillette, and R. Maulik , Leveraging interpolation models and error bounds for verifiable scientific machine learning , J. Comput. Phys., 524 (2025), p. 113726

work page 2025

[7] [7]

u tt, and K.-R. M \

S. Chmiela, A. Tkatchenko, H. E. Sauceda, I. Poltavsky, K. T. Sch \"u tt, and K.-R. M \"u ller , Machine learning of accurate energy-conserving molecular force fields , Sci. Adv., 3 (2017), p. e1603015

work page 2017

[8] [8]

Clevert, T

D.-A. Clevert, T. Unterthiner, and S. Hochreiter , Fast and accurate deep network learning by exponential linear units ( ELUs ) , arXiv preprint arXiv:1511.07289, (2015)

arXiv 2015