Pith. sign in

REVIEW 1 major objections 1 minor 27 references

Verified residual-specific explicit derivative kernels for physics-informed learning and discretized PDE adjoints

T0 review · 1 major / 1 minor · reviewed 2026-06-30 · grok-4.3

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

Pith's one-line read Residual-specific explicit derivative kernels achieve floating-point agreement with nested AD while accelerating PINN training by 2-4x.

desk verdict The paper shows residual-specific explicit kernels can match AD to float precision with 2-4x speedups in PINN training and CFD adjoints via agent generation plus checks, but the checks are case-limited. read the letter →

arxiv 2606.29702 v1 pith:DW66TTDX submitted 2026-06-29 physics.comp-ph

Wenbo Cao , Zhe Lu , Weiwei Zhang This is my paper · ORCID
classification physics.comp-ph
keywords physics-informedneuralnetworksexplicitdifferentiationautomaticdiscreteadjointscomputationalfluiddynamicsPINNtrainingPDEresidualsadjointmethods
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 establishes that explicit derivative kernels tailored to specific PDE residuals can replace general-purpose automatic differentiation for derivative computations in physics-informed neural networks and discretized PDE adjoints. The method generates residual-specific kernels via agent-assisted implementation and verifies them with numerical checks such as Taylor-remainder and inner-product tests. For PINNs the kernels realize the derivative-state closure through specialized layerwise operations, while for CFD they supply tangent and transpose actions that integrate into adjoint workflows. If these kernels are correct they deliver the same accuracy as nested AD but with lower runtime and memory costs for high-order residuals.

What carries the argument

residual-specific partial-jet propagation that renders the derivative-state closure explicit via specialized layerwise kernels

What would settle it

A new residual or discretization where the explicit kernel's residual or gradient output differs from the nested-AD result by more than machine epsilon, or where the adjoint sensitivities fail a finite-difference consistency test.

Watch

Extended reading notes

Core claim

The paper claims that residual-specific partial-jet propagation makes the derivative-state closure of a target PDE residual explicit and realizes it through specialized layerwise kernels rather than nested AD, while the same verification-driven strategy produces tangent-action and transpose-action kernels for a finite-volume CFD residual that pass Taylor-remainder, inner-product, and reduced-gradient consistency checks and embed into a GPU-resident discrete-adjoint workflow.

Load-bearing premise

Numerical verification checks plus agent-assisted implementation suffice to guarantee correctness of the generated explicit kernels for the full range of residuals and discretizations the method targets.

Editorial extensions

If this is right

  • ED kernels achieve floating-point-level agreement in residual and parameter-gradient evaluations relative to nested AD.
  • Complete PINN training accelerates by 2-4x while reducing peak GPU memory in most cases.
  • Generated tangent-action and transpose-action kernels pass Taylor-remainder, inner-product, and reduced-gradient consistency checks.
  • Kernels embed into a GPU-resident discrete-adjoint workflow for freestream Mach-number and angle-of-attack inversion.

Reading between the lines

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

  • The same generation-plus-verification pattern could be applied to derivative computations in other scientific machine-learning settings that rely on high-order PDE residuals.
  • Agent-assisted kernel creation paired with these checks may reduce the manual effort needed to obtain fast custom derivatives for novel discretizations.
  • Wider use might lower overall compute and memory demands in large-scale physics simulations that repeatedly evaluate residuals and adjoints.
Share X Bluesky LinkedIn Reddit HN

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

1 major / 1 minor

Summary. The paper claims that residual-specific explicit derivative kernels, generated via agent-assisted implementation for PINNs using partial-jet propagation and for finite-volume CFD operators, achieve floating-point agreement with nested AD, deliver 2-4x speedups and memory reductions in PINN training, and pass Taylor-remainder, inner-product, and reduced-gradient consistency checks, enabling their use in discrete-adjoint workflows for parameter inversion.

Significance. If the verification is sufficient to establish correctness across the intended range of residuals and discretizations, this method could serve as an efficient, structure-exploiting complement to general-purpose AD in physics-informed learning and CFD adjoint computations.

major comments (1)
  1. [Verification of the generated kernels] The Taylor-remainder, inner-product, and reduced-gradient consistency checks are presented as verification for the tangent-action and transpose-action kernels, but these checks are applied only to the specific PINN residuals and finite-volume CFD operator examined. The manuscript does not demonstrate that these checks are exhaustive or that they would detect localized sign or indexing errors that might cancel under the tested conditions, which is load-bearing for the claim that the kernels are 'verified' and correct for the broader class of residuals.
minor comments (1)
  1. The abstract asserts specific performance numbers (2-4x speedups, floating-point agreement) without referencing the corresponding figures or tables in the manuscript.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the careful reading and for identifying the need to clarify the scope of our numerical verification. We respond to the single major comment below.

read point-by-point responses

  1. Referee: [Verification of the generated kernels] The Taylor-remainder, inner-product, and reduced-gradient consistency checks are presented as verification for the tangent-action and transpose-action kernels, but these checks are applied only to the specific PINN residuals and finite-volume CFD operator examined. The manuscript does not demonstrate that these checks are exhaustive or that they would detect localized sign or indexing errors that might cancel under the tested conditions, which is load-bearing for the claim that the kernels are 'verified' and correct for the broader class of residuals.

    Authors: We agree that the verification is performed on the specific residuals and operators presented and that the manuscript does not contain a general proof that the chosen checks are exhaustive for every conceivable residual. The Taylor-remainder test is sensitive to first-order errors (including sign flips), the inner-product test directly probes transpose consistency, and the reduced-gradient test checks the overall adjoint action; in the reported cases these tests reach floating-point agreement with nested AD. We have revised the text to state explicitly that 'verified' refers to the examined instances, to note that analogous checks are required for new residuals, and to remove any implication of automatic generality beyond the demonstrated cases. revision: yes

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity

full rationale

The paper proposes a method for generating residual-specific explicit derivative kernels via agent-assisted implementation, then reports empirical outcomes (floating-point agreement with nested AD, 2-4x speedups, and passage of Taylor-remainder / inner-product / reduced-gradient checks) from direct execution on specific PINN residuals and a finite-volume CFD operator. No equations, fitted parameters, or predictions are presented that reduce by construction to the inputs; the central claims are implementation results rather than derivations. No self-citation chains or uniqueness theorems are invoked as load-bearing premises. The work is therefore self-contained against external benchmarks.

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

Abstract-only review supplies no information on free parameters, background axioms, or new postulated entities.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Verified residual-specific explicit derivative kernels for physics-informed learning and discretized PDE adjoints." pith.science (2026). https://pith.science/paper/DW66TTDX

@misc{pith2026260629702,
  author       = {Pith},
  title        = {Pith review of: Verified residual-specific explicit derivative kernels for physics-informed learning and discretized PDE adjoints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DW66TTDX}},
  note         = {Machine review of arXiv:2606.29702}
}
read the original abstract

Derivative computation is central to scientific computing, from space-time derivatives in physics-informed neural networks (PINNs) to residual Jacobian actions and discrete-adjoint operators in computational fluid dynamics (CFD). General-purpose automatic differentiation (AD) reduces implementation effort, but can incur substantial runtime and memory overhead for high-order residuals and complex discretized operators. Explicit derivative kernels can exploit problem-specific structure and provide efficient, controllable evaluations, but their use has been limited by derivation and implementation costs. This work revisits explicit differentiation (ED) as a residual-specific and verifiable route enabled by agent-assisted implementation and stringent numerical verification. For PINNs, we propose residual-specific partial-jet propagation, which makes the derivative-state closure of the target PDE residual explicit and realizes it through specialized layerwise kernels, rather than relying only on nested AD or a generic Taylor-mode transform. Relative to nested AD, the resulting ED kernels achieve floating-point-level agreement in residual and parameter-gradient evaluations and accelerate complete PINN training, often reaching 2-4x speedups while reducing peak GPU memory in most cases. For discretized PDE adjoints, we apply the same verification-driven strategy to a finite-volume CFD residual. The generated tangent-action and transpose-action kernels pass Taylor-remainder, inner-product, and reduced-gradient consistency checks, and are embedded into a GPU-resident discrete-adjoint workflow for freestream Mach-number and angle-of-attack inversion. These results suggest that verified explicit derivative kernels, supported by agent-assisted implementation, can serve as a practical, structure-aware complement to general-purpose AD for derivative-intensive scientific computing.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

27 extracted references · 2 canonical work pages

  1. [1]

    Raissi, P

    M. Raissi, P. Perdikaris, 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) 686-707

    2019

  2. [2]

    Y . Du, J.A. Ekaterinaris, Time -marching schemes for spatially high order accurate discretizations of the Euler and Navier–Stokes equations, Progress in Aerospace Sciences, 130 (2022) 100795

    2022

  3. [3]

    Knoll, D.E

    D.A. Knoll, D.E. Keyes, Jacobian-free Newton–Krylov methods: a survey of approaches and applications, J. Comput. Phys., 193 (2004) 357-397

    2004

  4. [4]

    Cacuci, Sensitivity theory for nonlinear systems

    D.G. Cacuci, Sensitivity theory for nonlinear systems. I. Nonlinear functional analysis approach, Journal of Mathematical Physics, 22 (1981) 2794-2802

    1981

  5. [5]

    McKeon, A.S

    B.J. McKeon, A.S. Sharma, A critical-layer framework for turbulent pipe flow, J. Fluid Mech., 658 (2010) 336-382

    2010

  6. [6]

    Kenway, C.A

    G.K.W. Kenway, C.A. Mader, P. He, J.R.R.A. Martins, Effective adjoint approaches for computational fluid dynamics, Progress in Aerospace Sciences, 110 (2019) 100542

    2019

  7. [7]

    Jameson, Aerodynamic design via control theory, Journal of scientific computing, 3 (1988) 233-260

    A. Jameson, Aerodynamic design via control theory, Journal of scientific computing, 3 (1988) 233-260

    1988

  8. [8]

    Baydin, B.A

    A.G. Baydin, B.A. Pearlmutter, A.A. Radul, J.M. Siskind, Automatic differentiation in machine learning: a survey, Journal of Machine Learning Research, 18 (2018) 1-43

    2018

Show all 27 references
  1. [9]

    L. Lu, X.H. Meng, Z.P. Mao, G.E. Karniadakis, DeepXDE: A Deep Learning Library for Solving Differential Equations, SIAM Rev., 63 (2021) 208-228

    2021

  2. [10]

    Haghighat, R

    E. Haghighat, R. Juanes, SciANN: A Keras/TensorFlow wrapper for scientific computations and physics-informed deep learning using artificial neural networks, Computer Methods in Applied Mechanics and Engineering, 373 (2021) 113552

    2021

  3. [11]

    Hennigh, S

    O. Hennigh, S. Narasimhan, M.A. Nabian, A. Subramaniam, K. Tangsali, Z. Fang, M. Rietmann, W. Byeon, S. Choudhry, NVIDIA SimNet™: An AI-accelerated multi-physics simulation framework, in: International conference on computational science, Springer, 2021, pp. 447 -461

    2021

  4. [12]

    Albring, M

    T.A. Albring, M. Sagebaum, N.R. Gauger, Efficient aerodynamic design using the discrete adjoint method in SU2, in: 17th AIAA/ISSMO multidisciplinary analysis and optimization conference, 2016, pp. 3518

    2016

  5. [13]

    Mader, G.K

    C.A. Mader, G.K. Kenway, A. Yildirim, J.R. Martins, ADflow: An open -source 21 computational fluid dynamics solver for aerodynamic and multidisciplinary optimization, Journal of Aerospace Information Systems, 17 (2020) 508-527

    2020

  6. [14]

    P. He, C.A. Mader, J.R. Martins, K.J. Maki, Dafoam: An open-source adjoint framework for multidisciplinary design optimization with openfoam, AIAA J., 58 (2020) 1304-1319

    2020

  7. [15]

    Sharma, V

    R. Sharma, V . Shankar, Accelerated training of physics -informed neural networks (pinns) using meshless discretizations, Adv. Neural Inf. Process. Syst., 35 (2022) 1034-1046

    2022

  8. [16]

    Z. Hu, Z. Shi, G.E. Karniadakis, K. Kawaguchi, Hutchinson trace estimation for high - dimensional and high -order physics -informed neural networks, Computer Methods in Applied Mechanics and Engineering, 424 (2024) 116883

    2024

  9. [17]

    Dangel, T

    F. Dangel, T. Siebert, M. Zeinhofer, A. Walther, Collapsing Taylor mode automatic differentiation, Adv. Neural Inf. Process. Syst., 38 (2025)

    2025

  10. [18]

    Hückelheim, N

    J. Hückelheim, N. Kukreja, S.H.K. Narayanan, F. Luporini, G. Gorman, P. Hovland, Automatic differentiation for adjoint stencil loops, in: Proceedings of the 48th International Conference on Parallel Processing, 2019, pp. 1-10

    2019

  11. [19]

    Müller, P

    J.D. Müller, P. Cusdin, On the performance of discrete adjoint CFD codes using automatic differentiation, International journal for numerical methods in fluids, 47 (2005) 939-945

    2005

  12. [20]

    Nytko, A

    N. Nytko, A. Taghibakhshi, T.U. Zaman, S. MacLachlan, L.N. Olson, M. West, Optimized sparse matrix operations for reverse mode automatic differentiation, SIAM J. Sci. Comput., 47 (2025) C1115-C1143

    2025

  13. [21]

    Hückelheim, L

    J. Hückelheim, L. Hascoët, Source -to-source automatic differentiation of openmp parallel loops, ACM Transactions on Mathematical Software (TOMS), 48 (2022) 1-32

    2022

  14. [22]

    M. Chen, J. Tworek, H. Jun, Q. Yuan, H.P.D.O. Pinto, J. Kaplan, H. Edwards, Y . Burda, N. Joseph, G. Brockman, Evaluating large language models trained on code, arXiv preprint arXiv:2107.03374, (2021)

    2021 arXiv

  15. [23]

    J. Yang, C. Jimenez, A. Wettig, K. Lieret, S. Yao, K. Narasimhan, O. Press, Swe -agent: Agent-computer interfaces enable automated software engineering, Adv. Neural Inf. Process. Syst., 37 (2024) 50528-50652

    2024

  16. [24]

    Madaan, N

    A. Madaan, N. Tandon, P. Gupta, S. Hallinan, L. Gao, S. Wiegreffe, U. Alon, N. Dziri, S. Prabhumoye, Y . Yang, Self-refine: Iterative refinement with self -feedback, Adv. Neural Inf. Process. Syst., 36 (2023) 46534-46594

    2023

  17. [25]

    Shinn, F

    N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, S. Yao, Reflexion: Language agents with verbal reinforcement learning, Adv. Neural Inf. Process. Syst., 36 (2023) 8634-8652

    2023

  18. [26]

    H. Le, Y . Wang, A.D. Gotmare, S. Savarese, S.C.H. Hoi, Coderl: Mastering code generation through pretrained models and deep reinforcement learning, Adv. Neural Inf. Process. Syst., 35 (2022) 22 21314-21328

    2022

  19. [27]

    S. Wang, S. Sankaran, H. Wang, P. Perdikaris, An expert's guide to training physics-informed neural networks, arXiv preprint arXiv:2308.08468, (2023)

    2023

Pith tools

Reviewed June 30, 2026 · model on record in the stance chip above.