pith. machine review for the scientific record. sign in

arxiv: 2605.00820 · v1 · submitted 2026-05-01 · 💻 cs.CE · cs.LG· cs.NA· math.NA

Recognition: unknown

HyCOP: Hybrid Composition Operators for Interpretable Learning of PDEs

Jinpai Zhao , Nishant Panda , Yen Ting Lin , Eirik Valseth , Diane Oyen , Clint Dawson
Authors on Pith no claims yet

Pith reviewed 2026-05-09 17:58 UTC · model grok-4.3

classification 💻 cs.CE cs.LGcs.NAmath.NA
keywords hybrid operatorsPDE learninginterpretable modelsmodular compositionneural operatorspolicy learningscientific machine learningoperator learning
0
0 comments X

The pith

HyCOP learns a policy to compose simple modules into short programs that solve parametric PDEs more accurately outside the training distribution.

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

The paper introduces HyCOP as a framework that replaces a single large neural operator with a learned policy deciding which basic module to apply next, such as advection or diffusion steps, and for how long, based on the current state and query features. Modules can be either numerical sub-solvers or learned closures, allowing the resulting program to be evaluated at any desired time without step-by-step rollout. A reader would care because the approach aims to deliver both higher accuracy on unseen regimes and human-readable sequences of operations that can be updated by swapping individual modules like boundary conditions. The supporting theory decomposes total error into separate composition and module contributions to diagnose where the method succeeds or fails.

Core claim

HyCOP learns parametric PDE solution operators by composing simple modules (advection, diffusion, learned closures, boundary handling) in a query-conditioned way, producing interpretable programs rather than monolithic maps, with order-of-magnitude out-of-distribution improvements over monolithic neural operators and support for modular transfer through dictionary updates, while theory characterizes expressivity and supplies an error decomposition separating composition error from module error.

What carries the argument

A policy over short programs that selects and sequences modules conditioned on regime features and state statistics, enabling hybrid numerical-learned surrogates evaluable at arbitrary query times.

If this is right

  • The resulting programs are human-readable sequences of module applications rather than opaque weight matrices.
  • Out-of-distribution accuracy improves by an order of magnitude compared with monolithic neural operators on the tested benchmarks.
  • Boundary conditions or residual terms can be changed by updating entries in the module dictionary without retraining the full model.
  • Solutions can be queried at arbitrary times without autoregressive rollout of intermediate steps.
  • The error decomposition isolates whether failures stem from poor module choice or from individual module inaccuracy.

Where Pith is reading between the lines

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

  • The modular structure may make it easier to enforce physical constraints such as conservation by restricting the allowed module dictionary.
  • Dictionary-based transfer could reduce the cost of adapting a model to new geometries or forcing terms that share most but not all physics.
  • Inspecting the policy-chosen program on a failing query could serve as a diagnostic tool to decide whether to add a new module type.

Load-bearing premise

A learned policy can reliably select and compose simple modules to capture full PDE dynamics without substantial composition error or loss of accuracy across regimes.

What would settle it

Running the method on a held-out PDE regime where the automatically chosen module sequences either match or exceed monolithic neural operator accuracy while remaining short and interpretable, or conversely where the programs become long or inaccurate due to composition failures.

Figures

Figures reproduced from arXiv: 2605.00820 by Clint Dawson, Diane Oyen, Eirik Valseth, Jinpai Zhao, Nishant Panda, Yen Ting Lin.

Figure 1
Figure 1. Figure 1: Compose–diagnose– enrich. HyCOP’s workflow for incomplete or hybrid physics (§5.2). Regime A: surrogate fitting (physics is known). When the governing equations are fully specified, the scientist wants a fast surrogate for downstream tasks that remains accurate as initial con￾ditions, boundary conditions, parameters, or resolution shift at test time. HyCOP encodes each known process as a dictionary primi￾t… view at source ↗
Figure 2
Figure 2. Figure 2: HyCOP at a glance. (a) A small learned policy πθ (∼50–100 parameters) maps a query x = (u0, µ, T) to a program of (primitive, duration) pairs from dictionary D, composed over [0, T] to produce uˆ(T); programs adapt to the regime (walkthrough in §3.1). (b) Each primitive can be numerical, learned, or both, enabling hybrid dictionaries (§5.2.1). (c) πθ smoothly reallocates time with regime numbers, consisten… view at source ↗
Figure 3
Figure 3. Figure 3: Expressivity (Thm. A.7) and error decomposition (Thm. 4.1) provide the approximation view at source ↗
Figure 4
Figure 4. Figure 4: Dam-break transfer (zero-shot). Left: reference SWE height surface with a highlighted cross-section. Right: 1D slice at y = 50m comparing baselines vs. HyCOP. Swapping only the boundary primitive (periodic → wall) yields a sharp improvement and resolves the shock region (gray box), indicating boundary physics as the dominant shift view at source ↗
Figure 5
Figure 5. Figure 5: HyCOP training with Evolution Strategies (ES). HyCOP conditions the policy on low-dimensional physics-based features f(x) (e.g., Péclet/Damköhler/Froude numbers and state statistics) to predict a split program (primitive choices and durations). We optimize policy parameters via ES using black-box evaluations of the composed operator loss. before computing updates, making optimization robust to outliers and… view at source ↗
Figure 6
Figure 6. Figure 6: 1D advection–diffusion (ID): qualitative comparison at view at source ↗
Figure 7
Figure 7. Figure 7: 1D advection–diffusion (OOD): extrapolation in view at source ↗
Figure 8
Figure 8. Figure 8: 1D SWE qualitative example (ID) view at source ↗
Figure 9
Figure 9. Figure 9: 1D SWE qualitative example (OOD). C.2.3 Viscous Burgers We consider the 1D viscous Burgers equation on x ∈ [0, 2], ∂u ∂t + u ∂u ∂x = ν ∂ 2u ∂x2 , (6) with the canonical split into nonlinear advection and viscous diffusion primitives. Data generation. Training trajectories are generated with viscosity ν ∈ [0.005, 0.1]. Initial conditions are sampled from six families: step functions, sinusoidal waves, Gauss… view at source ↗
Figure 10
Figure 10. Figure 10: 1D viscous Burgers (ID): qualitative comparison at view at source ↗
Figure 11
Figure 11. Figure 11: 1D viscous Burgers (OOD): extrapolation in view at source ↗
Figure 12
Figure 12. Figure 12: 2D ADR qualitative example (ID) view at source ↗
Figure 13
Figure 13. Figure 13: 2D ADR qualitative example (OOD). Note: cRMSE is not applicable for ADR (no conserved mass constraint under Fisher–KPP kinetics). 29 view at source ↗
Figure 14
Figure 14. Figure 14: 2D SWE qualitative example (ID) view at source ↗
Figure 15
Figure 15. Figure 15: 2D SWE qualitative example (OOD) view at source ↗
Figure 16
Figure 16. Figure 16: 2D ADR trajectory (ID). Example long-horizon prediction at horizons of 1/5/10/20 steps for the 2D ADR system (Section C.4.1) view at source ↗
Figure 17
Figure 17. Figure 17: 2D ADR trajectory (OOD). Same visualization as view at source ↗
Figure 18
Figure 18. Figure 18: 2D SWE trajectory (ID). Example long-horizon prediction at horizons of 1/5/10/20 steps for the 2D SWE system (Section C.4.2) view at source ↗
Figure 19
Figure 19. Figure 19: 2D SWE trajectory (OOD). Same visualization as view at source ↗
Figure 20
Figure 20. Figure 20: AD→ADR via dictionary enrichment (single illustrative sample). HyCOP pretrained on advection–diffusion (AD) misses reaction-driven changes in ADR. Adding a residual primitive and relearning the policy localizes corrections to regions with high reaction intensity, reducing error. Per-sample L 2 values shown; test-set averages are reported in view at source ↗
Figure 21
Figure 21. Figure 21: 1D KS (ID). HyCOP reproduces the spatiotemporal structure (top), energy spectrum (bottom-left), and long-time state distribution (bottom-right) within the training regime (W ∈ [24, 40], T ∈ [5, 8]). learned. This isolates the primitive-error term: the policy space is identical to HyCOP on ADR, so any gap traces to primitive quality, not to compositional structure. C.6 HyCOP on chaotic/multiscale PDEs We u… view at source ↗
Figure 22
Figure 22. Figure 22: 1D KS (OOD: domain shift). HyCOP generalizes to W = 50 (1.25× training maximum) while preserving attractor statistics view at source ↗
Figure 23
Figure 23. Figure 23: 1D KS (OOD: horizon shift). HyCOP generalizes to T = 20 (2.5× training maximum) while maintaining stable energy spectrum and state distribution view at source ↗
Figure 24
Figure 24. Figure 24: 1D KS (OOD: combined domain–horizon shift). HyCOP generalizes to the joint extrapolation (W = 50, T = 20) while preserving attractor statistics: spatiotemporal structure (top), energy spectrum (bottom-left), and long-time state distribution (bottom-right). Pointwise errors are uninformative under Lyapunov divergence; HyCOP matches the reference solver at the level of invariant statistics. 39 view at source ↗
read the original abstract

We introduce HyCOP, a modular framework that learns parametric PDE solution operators by composing simple modules (advection, diffusion, learned closures, boundary handling) in a query-conditioned way. Rather than learning a monolithic map, HyCOP learns a policy over short programs - which module to apply and for how long - conditioned on regime features and state statistics. Modules may be numerical sub-solvers or learned components, enabling hybrid surrogates evaluated at arbitrary query times without autoregressive rollout. Across diverse PDE benchmarks, HyCOP produces interpretable programs, delivers order-of-magnitude OOD improvements over monolithic neural operators, and supports modular transfer through dictionary updates (e.g., boundary swaps, residual enrichment). Our theory characterizes expressivity and gives an error decomposition that separates composition error from module error and doubles as a process-level diagnostic.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

0 major / 3 minor

Summary. The paper introduces HyCOP, a modular framework for learning parametric PDE solution operators by composing simple modules (advection, diffusion, learned closures, boundary handling) via a query-conditioned policy over short programs. Modules may be numerical sub-solvers or learned components, enabling hybrid surrogates evaluable at arbitrary query times without autoregressive rollout. The manuscript claims that HyCOP yields interpretable programs, order-of-magnitude OOD gains over monolithic neural operators, and supports modular transfer via dictionary updates (e.g., boundary swaps, residual enrichment). It also provides a theory characterizing expressivity together with an error decomposition that separates composition error from module error and serves as a process-level diagnostic.

Significance. If the empirical OOD gains and modular-transfer results hold, the work would be significant for scientific machine learning and computational engineering. It offers a hybrid, interpretable alternative to black-box neural operators, with built-in support for regime adaptation and a diagnostic decomposition that can surface composition failures. The provision of modular-transfer experiments and an explicit error decomposition are concrete strengths that go beyond typical neural-operator papers.

minor comments (3)
  1. [Abstract] Abstract: the phrase 'order-of-magnitude OOD improvements' is strong; the main text should report the precise factors, baselines, and statistical variability from the benchmarks to allow direct assessment of the claim.
  2. [Theory] Theory section: while the error decomposition is a useful diagnostic, the manuscript should explicitly show how composition error is isolated and measured in the reported experiments so that readers can verify the separation.
  3. [Experiments] Experiments: clarify the exact conditioning features (regime features and state statistics) used by the policy and provide pseudocode or a small example program to illustrate interpretability.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for the positive assessment of HyCOP, the accurate summary of its contributions, and the recommendation for minor revision. The referee's description correctly highlights the modular composition, query-conditioned policy, hybrid numerical-learned modules, OOD gains, modular transfer, and the error decomposition.

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper's central claims rest on a modular policy-learning framework, empirical OOD benchmarks across PDEs, and an independent theoretical characterization of expressivity plus an error decomposition that separates composition from module error. No load-bearing step reduces by construction to a fitted parameter, self-citation, or renamed input; the derivation chain is self-contained against external benchmarks and does not invoke uniqueness theorems or ansatzes from prior author work as the sole justification.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Only the abstract is available, preventing identification of specific free parameters, axioms, or invented entities. The approach likely relies on learned policy parameters and module definitions, but no details are provided.

pith-pipeline@v0.9.0 · 5455 in / 1164 out tokens · 31004 ms · 2026-05-09T17:58:13.632951+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

64 extracted references · 5 canonical work pages · 1 internal anchor

  1. [1]

    Journal of Computational Physics , volume =

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations , author =. Journal of Computational Physics , volume =. 2019 , doi =

    2019

  2. [2]

    Learning nonlinear operators via

    Lu, Lu and Jin, Pengzhan and Pang, Guofei and Zhang, Zhongqiang and Karniadakis, George Em , journal =. Learning nonlinear operators via. 2021 , doi =

    2021

  3. [3]

    Neural Operator: Learning Maps Between Function Spaces with Applications to

    Kovachki, Nikola and Li, Zongyi and Liu, Burigede and Azizzadenesheli, Kamyar and Bhattacharya, Kaushik and Stuart, Andrew and Anandkumar, Anima , journal =. Neural Operator: Learning Maps Between Function Spaces with Applications to. 2023 , url =

    2023

  4. [4]

    International Conference on Learning Representations (ICLR) , year =

    Fourier Neural Operator for Parametric Partial Differential Equations , author =. International Conference on Learning Representations (ICLR) , year =

  5. [5]

    Convolutional Neural Operators for Robust and Accurate Learning of

    Raoni. Convolutional Neural Operators for Robust and Accurate Learning of. Advances in Neural Information Processing Systems (NeurIPS) , year =

  6. [6]

    ACM/IMS Journal of Data Science , volume =

    Physics-Informed Neural Operator for Learning Partial Differential Equations , author =. ACM/IMS Journal of Data Science , volume =. 2024 , doi =

    2024

  7. [7]

    Poseidon: Efficient Foundation Models for

    Herde, Maximilian and Raoni. Poseidon: Efficient Foundation Models for. Advances in Neural Information Processing Systems (NeurIPS) , year =

  8. [8]

    SIAM Journal on Scientific Computing , volume =

    Reduced Operator Inference for Nonlinear Partial Differential Equations , author =. SIAM Journal on Scientific Computing , volume =. 2022 , doi =

    2022

  9. [9]

    SIAM Journal on Scientific Computing , volume =

    Nonintrusive Reduced-Order Models for Parametric Partial Differential Equations via Data-Driven Operator Inference , author =. SIAM Journal on Scientific Computing , volume =. 2023 , doi =

    2023

  10. [10]

    Annual Review of Fluid Mechanics , volume =

    Learning Nonlinear Reduced Models from Data with Operator Inference , author =. Annual Review of Fluid Mechanics , volume =. 2024 , doi =

    2024

  11. [11]

    , journal =

    Chen, Yifan and Hosseini, Bamdad and Owhadi, Houman and Stuart, Andrew M. , journal =. Solving and Learning Nonlinear. 2021 , doi =

    2021

  12. [12]

    Journal of Computational Physics , volume =

    Kernel Methods are Competitive for Operator Learning , author =. Journal of Computational Physics , volume =. 2024 , doi =

    2024

  13. [13]

    Annual Reviews in Control , volume =

    Koopman Operator Dynamical Models: Learning, Analysis and Control , author =. Annual Reviews in Control , volume =. 2021 , doi =

    2021

  14. [14]

    and Budi

    Brunton, Steven L. and Budi. Modern. SIAM Review , volume =. 2022 , doi =

    2022

  15. [15]

    and Azizzadenesheli, Kamyar , journal =

    Rahman, Md Ashiqur and Ross, Zachary E. and Azizzadenesheli, Kamyar , journal =. U-. 2023 , issn =

    2023

  16. [16]

    Proceedings of the 41st International Conference on Machine Learning (ICML) , pages =

    Neural Operators with Localized Integral and Differential Kernels , author =. Proceedings of the 41st International Conference on Machine Learning (ICML) , pages =. 2024 , volume =

    2024

  17. [17]

    2015 , volume =

    Ronneberger, Olaf and Fischer, Philipp and Brox, Thomas , booktitle =. 2015 , volume =

    2015

  18. [18]

    2025 , eprint=

    Scale-Consistent Learning for Partial Differential Equations , author=. 2025 , eprint=

    2025

  19. [19]

    2025 , eprint=

    Physics Steering: Causal Control of Cross-Domain Concepts in a Physics Foundation Model , author=. 2025 , eprint=

    2025

  20. [20]

    Factorized

    Tran, Alasdair and Mathews, Alexander and Xie, Lexing and Ong, Cheng Soon , booktitle =. Factorized. 2023 , url =

    2023

  21. [21]

    Advances in Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks , year =

    Takamoto, Makoto and Praditia, Timothy and Leiteritz, Raphael and MacKinlay, Dan and Alesiani, Francesco and Pfl. Advances in Neural Information Processing Systems (NeurIPS) Track on Datasets and Benchmarks , year =

  22. [22]

    Wilhelm, Maite J. C. and Portegies Zwart, Simon , journal =. 2024 , doi =

    2024

  23. [23]

    Fractal decomposition of exponential operators with applications to many-body theories and

    Suzuki, Masuo , journal =. Fractal decomposition of exponential operators with applications to many-body theories and. 1990 , doi =

    1990

  24. [24]

    and Su, Yuan and Tran, Minh C

    Childs, Andrew M. and Su, Yuan and Tran, Minh C. and Wiebe, Nathan and Zhu, Shuchen , journal =. Theory of. 2021 , doi =

    2021

  25. [25]

    and Childs, Andrew M

    Berry, Dominic W. and Childs, Andrew M. and Cleve, Richard and Kothari, Robin and Somma, Rolando D. , journal =. Simulating. 2015 , doi =

    2015

  26. [26]

    arXiv preprint arXiv:2602.00884 , year=

    Test-time Generalization for Physics through Neural Operator Splitting , author=. arXiv preprint arXiv:2602.00884 , year=

    work page arXiv

  27. [27]

    Zhang, Jiahao and Wang, Yueqi and Lin, Guang , journal =

  28. [28]

    arXiv preprint arXiv:2403.12938 , year=

    Learning Neural Differential Algebraic Equations via Operator Splitting , author=. arXiv preprint arXiv:2403.12938 , year=

    work page arXiv

  29. [29]

    Learning Physical Operators using Neural Operators

    Learning Physical Operators using Neural Operators , author=. arXiv preprint arXiv:2602.23113 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2410.23889 , year=

    Kassa. arXiv preprint arXiv:2410.23889 , year=

    work page arXiv

  31. [31]

    Morel, Rudy and Han, Jiequn and Oyallon, Edouard , booktitle=

  32. [32]

    and Perdikaris, Paris and Turner, Richard E

    Lippe, Phillip and Veeling, Bastiaan S. and Perdikaris, Paris and Turner, Richard E. and Brandstetter, Johannes , booktitle=

  33. [33]

    Blending Neural Operators and Relaxation Methods in

    Zhang, Enrui and Kahana, Adar and Kopani. Blending Neural Operators and Relaxation Methods in. Nature Machine Intelligence , volume=. 2024 , doi=

    2024

  34. [34]

    Machine Learning: Science and Technology , volume=

    Verification and Validation for Trustworthy Scientific Machine Learning , author=. Machine Learning: Science and Technology , volume=. 2026 , publisher=

    2026

  35. [35]

    55th AIAA Aerospace Sciences Meeting , year=

    Singh, Anand Pratap and Duraisamy, Karthikeyan and Zhang, Ze Jia , title=. 55th AIAA Aerospace Sciences Meeting , year=

  36. [36]

    Perez, Ethan and Strub, Florian and de Vries, Harm and Dumoulin, Vincent and Courville, Aaron , title =. Proceedings of the Thirty-Second AAAI Conference on Artificial Intelligence and Thirtieth Innovative Applications of Artificial Intelligence Conference and Eighth AAAI Symposium on Educational Advances in Artificial Intelligence , articleno =. 2018 , isbn =

    2018

  37. [37]

    and Fournier, John J

    Adams, Robert A. and Fournier, John J. F. , title =

  38. [38]

    Berge, Claude , title =

  39. [39]

    Acta Numerica , volume =

    Blanes, Sergio and Casas, Fernando and Murua, Ander , title =. Acta Numerica , volume =

  40. [40]

    Mathematics of Control, Signals and Systems , volume =

    Cybenko, George , title =. Mathematics of Control, Signals and Systems , volume =

  41. [41]

    and Kaper, T

    Goldman, D. and Kaper, T. J. , title =. SIAM Journal on Numerical Analysis , volume =

  42. [42]

    Hairer, Ernst and Lubich, Christian and Wanner, Gerhard , title =

  43. [43]

    Neural Networks , volume =

    Hornik, Kurt and Stinchcombe, Maxwell and White, Halbert , title =. Neural Networks , volume =

  44. [44]

    BIT Numerical Mathematics , volume =

    Jahnke, Tobias and Lubich, Christian , title =. BIT Numerical Mathematics , volume =

  45. [45]

    and Quispel, G

    McLachlan, Robert I. and Quispel, G. Reinout W. , title =. Acta Numerica , volume =

  46. [46]

    Annals of Mathematics , volume =

    Michael, Ernest , title =. Annals of Mathematics , volume =

  47. [47]

    Salimans, J

    Salimans, Tim and Ho, Jonathan and Chen, Xi and Sidor, Szymon and Sutskever, Ilya , title =. arXiv preprint arXiv:1703.03864 , year =

    work page arXiv

  48. [48]

    IMA Journal of Numerical Analysis , volume =

    Sheng, Qin , title =. IMA Journal of Numerical Analysis , volume =

  49. [49]

    Journal of Mathematical Physics , volume =

    Suzuki, Masuo , title =. Journal of Mathematical Physics , volume =

  50. [50]

    SIAM Journal on Numerical Analysis , volume =

    Thalhammer, Mechthild , title =. SIAM Journal on Numerical Analysis , volume =

  51. [51]

    Solving Ordinary Differential Equations

    Hairer, Ernst and N. Solving Ordinary Differential Equations

  52. [52]

    2017 , eprint=

    The Concrete Distribution: A Continuous Relaxation of Discrete Random Variables , author=. 2017 , eprint=

    2017

  53. [53]

    SIAM journal on numerical analysis , volume=

    On the construction and comparison of difference schemes , author=. SIAM journal on numerical analysis , volume=. 1968 , publisher=

    1968

  54. [54]

    Splitting Methods in Communication, Imaging, Science, and Engineering , editor=

    Operator Splitting , author=. Splitting Methods in Communication, Imaging, Science, and Engineering , editor=. 2016 , publisher=

    2016

  55. [55]

    Proceedings of the American Mathematical Society , volume=

    On the product of semi-groups of operators , author=. Proceedings of the American Mathematical Society , volume=. 1959 , publisher=

    1959

  56. [56]

    , series =

    Olver, Peter J. , series =. Applications of. 1993 , publisher =

    1993

  57. [57]

    , series=

    Hall, Brian C. , series=. 2015 , publisher=

    2015

  58. [58]

    1983 , publisher =

    Semigroups of Linear Operators and Applications to Partial Differential Equations , author =. 1983 , publisher =

    1983

  59. [59]

    and Liggett, Thomas M

    Crandall, Michael G. and Liggett, Thomas M. , journal =. Generation of Semi-groups of Nonlinear Transformations on General

  60. [60]

    2024 , publisher=

    Dawson, Clint and Loveland, Mark and Pachev, Benjamin and Proft, Jennifer and Valseth, Eirik , journal=. 2024 , publisher=

    2024

  61. [61]

    and Dean, Joseph P

    Baratta, Igor A. and Dean, Joseph P. and Dokken, J. 2023 , doi=

    2023

  62. [62]

    Serrano et al., "Test-time Generalization

  63. [63]

    Zhang et al., "LegONet: Plug-and-Play Structure-

  64. [64]

    ^ Depends on whether the fixed schedule suits the target regime

    Koch et al., "Learning Neural Differential Algebraic related_works, proof plainnat table [t] Positioning. ^ Depends on whether the fixed schedule suits the target regime. tab:positioning 3pt tabular lccc & Classical splitting & Neural operators & HyCOP \\ Composition & Fixed schedule & None (monolithic) & Learned policy \\ Regime adaptivity & None & Learn...