pith. sign in

arxiv: 2512.03476 · v2 · submitted 2025-12-03 · 💻 cs.LG · cs.AI· cs.MA· cs.NA· math.NA· physics.comp-ph

ATHENA: Agentic Team for Hierarchical Evolutionary Numerical Algorithms

Juan Diego Toscano , Daniel T. Chen , George Em Karniadakis This is my paper

Pith reviewed 2026-05-17 01:51 UTC · model grok-4.3

classification 💻 cs.LG cs.AIcs.MAcs.NAmath.NAphysics.comp-ph
keywords agentic frameworksevolutionary numerical algorithmsscientific computingscientific machine learningcontextual banditsnumerical solversautonomous discovery
0
0 comments X

The pith

An agentic framework treats numerical algorithm design as a contextual bandit problem to reach validation errors of 10 to the minus 14.

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

The paper presents ATHENA as a system that runs an end-to-end autonomous lab for developing computational methods in science and scientific machine learning. Its central mechanism is the HENA loop, which reviews past trials and chooses structural changes to code blueprints drawn from expert principles such as universal approximation or physics constraints. These choices convert into executable programs that deliver scientific rewards, allowing the system to locate exact analytical solutions through symmetries or to build stable solvers where other models break down. In more difficult cases the loop combines symbolic and numeric techniques to handle multiphysics problems. Human collaboration can close remaining stability gaps and raise performance by another factor of ten.

Core claim

ATHENA frames the iterative creation of numerical algorithms as a contextual bandit problem in which an online learner selects structural actions from combinatorial spaces guided by expert blueprints, translates each action into executable code, and measures the resulting scientific reward. This process autonomously identifies mathematical symmetries for exact solutions, derives stable solvers where foundation models fail, diagnoses ill-posed formulations, and couples hybrid workflows such as physics-informed networks with finite-element methods to resolve multiphysics tasks, ultimately producing validation errors as low as 10 to the minus 14 and further gains when a human intervenes on gaps

What carries the argument

The HENA loop, a knowledge-driven diagnostic process framed as a contextual bandit problem that selects structural actions from expert blueprints to produce executable high-reward code.

If this is right

  • The system can locate mathematical symmetries that deliver exact analytical solutions without numerical approximation.
  • It produces stable numerical solvers in settings where standard foundation models break down.
  • For ill-posed scientific machine learning problems it performs deep diagnosis and constructs hybrid symbolic-numeric workflows.
  • Coupling physics-informed networks with finite-element methods resolves complex multiphysics problems.
  • Overall validation accuracy exceeds typical human levels and improves by an order of magnitude with occasional human input.

Where Pith is reading between the lines

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

  • The bandit framing could transfer to other combinatorial design tasks such as choosing discretization schemes or mesh topologies.
  • A hybrid model in which the loop handles routine evolution while experts supply high-level blueprints may scale to larger problems.
  • If the same loop generalizes to inverse problems in fluid dynamics or materials science, the autonomous mode could shorten discovery cycles.
  • Systematic tests on a broader suite of ill-posed inverse problems would reveal the precise boundary between fully autonomous and human-assisted regimes.

Load-bearing premise

That framing the selection of numerical structures as a contextual bandit problem lets the system reliably choose actions that yield executable code with high scientific rewards without extra tuning.

What would settle it

Running the system on a fresh benchmark such as the incompressible Navier-Stokes equations and checking whether it reaches 10 to the minus 14 validation error without human fixes to the generated code.

Figures

Figures reproduced from arXiv: 2512.03476 by Daniel T. Chen, George Em Karniadakis, Juan Diego Toscano.

Figure 1
Figure 1. Figure 1: Online Learning as a Model for Agentic Research. This diagram illustrates the research lifecycle modeled as a Contextual Bandit problem. The cycle begins with the Study phase (Policy π), where the Strategist synthesizes the Problem context and prior Rewards (Rn) to formulate a structural Plan or Action (An). The Implementation phase (Operator I) translates this abstract plan into executable Code or State (… view at source ↗
Figure 2
Figure 2. Figure 2: ATHENA (Agentic Team for Hierarchical Evolutionary Numerical Algorithms). The framework is organized into four logical groups, with specific icons indicating the heterogeneous allocation of Large Language Models (LLMs) to specialized roles (see legend). (A) The Conceptualization Group (Red): The user-facing triage system. The User interacts with a Coordinator to define a User Request, which the Gatekeeper … view at source ↗
Figure 3
Figure 3. Figure 3: Comparative analysis of the 2D Inviscid Burgers’ benchmark. The panels contrast the solution fields generated by state-of-the-art foundation models via direct prompting against the autonomous solution discovered by ATHENA. Baselines (Direct Prompting): Models such as GPT-5.1 and Claude Sonnet 4.5 incorrectly select Fourier Spectral methods and apply aggressive frequency filtering to handle the shock, resul… view at source ↗
Figure 4
Figure 4. Figure 4: Autonomous correction of the Kelvin-Helmholtz Instability (Euler Equations) The figure tracks the evolution of the solution fields—Density (ρ) and Velocity (v)—as ATHENA refines the solver configuration. Rows 1-2 (Iteration 1): The initial simulation completes with exit code 0 but exhibits a “silent physics failure.” The solution is dominated by numerical diffusion; the shear layer is smeared, and the vort… view at source ↗
Figure 5
Figure 5. Figure 5: Rayleigh-Taylor Instability (Compressible Navier-Stokes). Time evolution of density (ρ), velocity (u, v), and pressure (p). Unlike baselines that failed due to geometric distortions or instability, ATHENA stabilized the simulation by autonomously diagnosing two critical constraints: (1) It reconfigured the mesh topology into a 1 × 4 quadtree forest to match the domain’s 1:4 aspect ratio, preventing element… view at source ↗
Figure 6
Figure 6. Figure 6: Diagnostic dashboards for canonical benchmarks. These panels represent the exact visual state observed by the Advisor Agent at the end of the training stage. (a) Allen-Cahn: The system successfully captures the sharp phase transitions from the cosine initialization. (b) Viscous Burgers: The system resolves the shock formation with high fidelity (MSE ∼ 10−14). For both cases, the first row displays the loss… view at source ↗
Figure 7
Figure 7. Figure 7: Verification of advanced methodological capabilities. These dashboards illustrate how the Advisor Agent validates complex problem setups beyond standard data fitting. (a) Helmholtz: Demon￾strates the successful application of the Method of Manufactured Solutions (MMS). The agent verifies the steady-state solver by comparing it against a fabricated source term, ensuring the implementation correctly balances… view at source ↗
Figure 8
Figure 8. Figure 8: Diagnostic dashboard for the Inviscid Burgers equation with viscous continuation. This figure presents the comprehensive training data observed by the Advisor Agent. The “staircase” pattern prominent in the PDE Loss, Entropy Loss, and Relative L 2 Error histories (top row) and explicitly shown in the Viscosity Annealing Schedule (middle row, center) demonstrates the successful application of the continuati… view at source ↗
Figure 9
Figure 9. Figure 9: Human-Agent Collaboration: The Periodic Wavelet-KAN. By shifting from a Fourier basis (which struggled with the shock) to a Periodic Wavelet basis suggested by the human user, the system achieved a high-fidelity solution with minimal viscosity. The figure compares the final prediction against the exact solution, showing that the wavelet basis captures the discontinuity without the extensive ringing artifac… view at source ↗
Figure 10
Figure 10. Figure 10: Diagnostic Dashboard for the Final Refined Model (Run 13). This figure illustrates the performance of the optimal architecture—QR-DeepONet with KKAN Trunk and RBF Basis. Top Row: The history of Total Loss, Training Relative Error, and Testing Relative Error. The vertical red dashed line demarcates the two-stage training process inherent to the QR-DeepONet methodology. Note the stability of the Test Error … view at source ↗
Figure 11
Figure 11. Figure 11: Closed-Loop Artificial Intelligence Velocimetry Results. Top Row (u-velocity): Compar￾ison between the Ground Truth generated in Step 1 (Left), the PINN Reconstruction from Step 2 (Middle), and the Forward Numerical Verification from Step 3 (Right). The white circles in the first column indicate the sparse sensor locations where 5% Gaussian noise was injected for training. Middle Row (v-velocity): Reconst… view at source ↗
read the original abstract

Bridging the gap between theoretical conceptualization and computational implementation is a major bottleneck in Scientific Computing (SciC) and Scientific Machine Learning (SciML). We introduce ATHENA (Agentic Team for Hierarchical Evolutionary Numerical Algorithms), an agentic framework designed as an Autonomous Lab to manage the end-to-end computational research lifecycle. Its core is the HENA loop, a knowledge-driven diagnostic process framed as a Contextual Bandit problem. Acting as an online learner, the system analyzes prior trials to select structural `actions' ($A_n$) from combinatorial spaces guided by expert blueprints (e.g., Universal Approximation, Physics-Informed constraints). These actions are translated into executable code ($S_n$) to generate scientific rewards ($R_n$). ATHENA transcends standard automation: in SciC, it autonomously identifies mathematical symmetries for exact analytical solutions or derives stable numerical solvers where foundation models fail. In SciML, it performs deep diagnosis to tackle ill-posed formulations and combines hybrid symbolic-numeric workflows (e.g., coupling PINNs with FEM) to resolve multiphysics problems. The framework achieves super-human performance, reaching validation errors of $10^{-14}$. Furthermore, collaborative ``human-in-the-loop" intervention allows the system to bridge stability gaps, improving results by an order of magnitude. This paradigm shift focuses from implementation mechanics to methodological innovation, accelerating scientific discovery.

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

3 major / 1 minor

Summary. The manuscript introduces ATHENA, an agentic framework for end-to-end management of computational research in Scientific Computing and Scientific Machine Learning. Its core is the HENA loop, cast as a Contextual Bandit problem in which the system analyzes prior trials to select structural actions A_n from combinatorial spaces (guided by expert blueprints such as Universal Approximation or Physics-Informed constraints), translates them into executable code S_n, and obtains scientific rewards R_n. The paper claims that this framework achieves super-human performance, reaching validation errors of 10^{-14}, and that human-in-the-loop collaboration can further improve results by an order of magnitude.

Significance. If the performance claims were substantiated with reproducible benchmarks and clear methodological details, the work would be significant for demonstrating a knowledge-driven, agentic approach to automating numerical algorithm design and hybrid symbolic-numeric workflows where standard foundation models fail. The framing of hierarchical evolutionary search as an online bandit learner, combined with explicit integration of domain blueprints, offers a potentially useful paradigm shift from manual implementation to methodological innovation.

major comments (3)
  1. [Abstract] Abstract: The assertion of super-human performance with validation errors of 10^{-14} is presented without any benchmarks, baselines, error bars, experimental protocols, or verification procedures. This directly undermines the central performance claim.
  2. [Abstract] Abstract: The HENA loop is described as a Contextual Bandit problem, yet the manuscript supplies no definition of the reward function R_n, the state representation derived from prior trials, or evidence that the online learner converges in the stated combinatorial action spaces. Without these elements the reported precision cannot be assessed as autonomous.
  3. [Abstract] Abstract: The claim that human-in-the-loop intervention bridges stability gaps and improves results by an order of magnitude is stated without quantitative comparisons, specific case studies, or ablation data showing the magnitude of the improvement.
minor comments (1)
  1. [Abstract] The abstract refers to 'expert blueprints (e.g., Universal Approximation, Physics-Informed constraints)' without indicating how these are encoded as features or constraints within the bandit state or action space.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their thorough review and valuable feedback on our manuscript. We agree that the abstract requires strengthening to better support the central claims and will revise it along with relevant sections of the main text to include additional methodological details, references to experimental results, and quantitative evidence. Below we respond point-by-point to the major comments.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The assertion of super-human performance with validation errors of 10^{-14} is presented without any benchmarks, baselines, error bars, experimental protocols, or verification procedures. This directly undermines the central performance claim.

    Authors: We acknowledge that the abstract presents this performance claim without sufficient supporting context. The full manuscript reports these results from systematic experiments on SciC and SciML benchmarks, including direct comparisons against standard numerical solvers and foundation-model baselines. In the revision we will update the abstract to briefly note the experimental protocol (multiple independent runs on canonical test problems) and add a reference to the detailed results, tables, and error-bar plots in the Experiments section. revision: yes

  2. Referee: [Abstract] Abstract: The HENA loop is described as a Contextual Bandit problem, yet the manuscript supplies no definition of the reward function R_n, the state representation derived from prior trials, or evidence that the online learner converges in the stated combinatorial action spaces. Without these elements the reported precision cannot be assessed as autonomous.

    Authors: The Contextual Bandit formulation, including the explicit definition of R_n (a composite reward combining validation error, stability margin, and computational cost), the state vector (summary statistics of prior trial outcomes and constraint violations), and empirical convergence behavior across combinatorial action spaces, is provided in the Methodology section. We will revise the abstract to include a concise statement of these definitions and add a pointer to the formal description and convergence plots in the main text. revision: yes

  3. Referee: [Abstract] Abstract: The claim that human-in-the-loop intervention bridges stability gaps and improves results by an order of magnitude is stated without quantitative comparisons, specific case studies, or ablation data showing the magnitude of the improvement.

    Authors: We agree that the abstract would be strengthened by explicit quantitative support. The manuscript already contains case studies illustrating human-in-the-loop refinements; we will expand these with ablation tables that directly compare autonomous versus collaborative runs, reporting the observed order-of-magnitude gains in stability and accuracy. The abstract will be updated to reference these new quantitative comparisons. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected; framework claims are empirical rather than self-referential

full rationale

The paper describes an agentic system (ATHENA) whose core HENA loop is presented as a contextual bandit that selects actions A_n from blueprints, translates them to code S_n, and obtains rewards R_n via execution. Reported performance (validation errors of 10^{-14}) is framed as an outcome of running this loop on scientific problems, with optional human-in-the-loop for stability. No equations, self-citations, or derivation steps are supplied that reduce the claimed results to quantities defined in terms of the bandit parameters or prior trials by construction. The central claims rest on external execution and reward evaluation rather than internal redefinition or fitted-input renaming, making the presentation self-contained against the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 1 invented entities

Review is based on abstract only. The framework relies on domain assumptions such as the validity of universal approximation and physics-informed constraints as guiding blueprints. No explicit free parameters or invented entities with independent evidence are stated in the abstract.

axioms (1)
  • domain assumption Universal Approximation theorem and Physics-Informed constraints provide reliable expert blueprints for guiding structural actions
    Explicitly listed as examples of expert blueprints that steer the selection of actions A_n in the HENA loop.
invented entities (1)
  • HENA loop no independent evidence
    purpose: Knowledge-driven diagnostic process that selects structural actions as an online learner
    Core novel component introduced to frame algorithm design as a contextual bandit problem.

pith-pipeline@v0.9.0 · 5556 in / 1522 out tokens · 52091 ms · 2026-05-17T01:51:52.224672+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Forward citations

Cited by 2 Pith papers

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

  1. GRAFT-ATHENA: Self-Improving Agentic Teams for Autonomous Discovery and Evolutionary Numerical Algorithms

    cs.LG 2026-05 unverdicted novelty 6.0

    GRAFT-ATHENA projects combinatorial method choices into factored trees that embed as fingerprints in a metric space, enabling an agentic system to accumulate experience across domains and autonomously discover new num...

  2. ALL-FEM: Agentic Large Language models Fine-tuned for Finite Element Methods

    cs.CE 2026-01 unverdicted novelty 6.0

    ALL-FEM fine-tunes LLMs on a corpus of verified FEniCS scripts and uses multi-agent workflows to automate finite element code generation, achieving 71.79% success on 39 benchmarks across elasticity, flow, and coupled ...

Reference graph

Works this paper leans on

83 extracted references · 83 canonical work pages · cited by 2 Pith papers · 5 internal anchors

  1. [1]

    W. F. Wiggins, A. S. Tejani, On the opportunities and risks of foundation models for natural language processing in radiology, Radiology: Artificial Intelligence 4 (4) (2022) e220119

    work page 2022

  2. [2]

    Schneider, C

    J. Schneider, C. Meske, P. Kuss, Foundation models: A new paradigm for artificial intelligence, Business & Information Systems Engineering 66 (2) (2024) 221–231

    work page 2024

  3. [3]

    C. Lu, C. Lu, R. T. Lange, J. Foerster, J. Clune, D. Ha, The ai scientist: Towards fully automated open-ended scientific discovery, arXiv preprint arXiv:2408.06292 (2024). 29

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  4. [4]

    Y. Choi, S. W. Cheung, Y. Kim, P.-H. Tsai, A. N. Diaz, I. Zanardi, S. W. Chung, D. M. Copeland, C. Kendrick, W. Anderson, et al., Defining foundation models for computational science: A call for clarity and rigor, arXiv preprint arXiv:2505.22904 (2025)

    work page arXiv 2025

  5. [5]

    Subramanian, P

    S. Subramanian, P. Harrington, K. Keutzer, W. Bhimji, D. Morozov, M. W. Mahoney, A. Gholami, Towards foundation models for scientific machine learning: Characterizing scaling and transfer behavior, Advances in Neural Information Processing Systems 36 (2023) 71242–71262

    work page 2023

  6. [6]

    Herde, B

    M. Herde, B. Raonic, T. Rohner, R. K¨ appeli, R. Molinaro, E. de B´ ezenac, S. Mishra, Poseidon: Efficient foundation models for pdes, Advances in Neural Information Pro- cessing Systems 37 (2024) 72525–72624

    work page 2024

  7. [7]

    Karniadakis, S

    G. Karniadakis, S. J. Sherwin, Spectral/hp element methods for computational fluid dynamics, Oxford University Press, USA, 2005

    work page 2005

  8. [8]

    Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations

    M. Raissi, P. Perdikaris, G. E. Karniadakis, Physics Informed Deep Learning (Part I): Data-driven Solutions of Nonlinear Partial Differential Equations, arXiv preprint arXiv:1711.10561 (2017)

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  9. [9]

    Jiang, Z

    Q. Jiang, Z. Gao, G. E. Karniadakis, Deepseek vs. chatgpt vs. claude: A comparative study for scientific computing and scientific machine learning tasks, Theoretical and Applied Mechanics Letters 15 (3) (2025) 100583

    work page 2025

  10. [10]

    N. F. Liu, K. Lin, J. Hewitt, A. Paranjape, M. Bevilacqua, F. Petroni, P. Liang, Lost in the middle: How language models use long contexts, Transactions of the Association for Computational Linguistics 12 (2024) 157–173

    work page 2024

  11. [11]

    J. Gama, I. ˇZliobait˙ e, A. Bifet, M. Pechenizkiy, A. Bouchachia, A survey on concept drift adaptation, ACM computing surveys (CSUR) 46 (4) (2014) 1–37

    work page 2014

  12. [12]

    S. Ross, G. Gordon, D. Bagnell, A reduction of imitation learning and structured pre- diction to no-regret online learning, in: Proceedings of the fourteenth international conference on artificial intelligence and statistics, JMLR Workshop and Conference Pro- ceedings, 2011, pp. 627–635

    work page 2011

  13. [13]

    C. Wu, A. J. Varghese, V. Oommen, G. E. Karniadakis, Gpt vs human for scientific reviews: A dual source review on applications of chatgpt in science, arXiv preprint arXiv:2312.03769 (2023)

    work page arXiv 2023

  14. [14]

    Ghafarollahi, M

    A. Ghafarollahi, M. J. Buehler, SciAgents: automating scientific discovery through bioinspired multi-agent intelligent graph reasoning, Advanced Materials 37 (22) (2025) 2413523

    work page 2025

  15. [15]

    M. E. Samadi, A. Schuppert, Flexible swarm learning may outpace foundation models in essential tasks, arXiv preprint arXiv:2510.06349 (2025). 30

    work page arXiv 2025

  16. [16]

    Q. Wuwu, C. Gao, T. Chen, Y. Huang, Y. Zhang, J. Wang, J. Li, H. Zhou, S. Zhang, PINNsAgent: Automated PDE surrogation with large language models, arXiv preprint arXiv:2501.12053 (2025)

    work page arXiv 2025

  17. [17]

    X. He, L. You, H. Tian, B. Han, I. Tsang, Y.-S. Ong, Lang-PINN: From language to physics-informed neural networks via a multi-agent framework, arXiv preprint arXiv:2510.05158 (2025)

    work page arXiv 2025

  18. [18]

    Bhatnagar, An agentic AI workflow to simplify parameter estimation of complex differential equation systems, arXiv preprint arXiv:2509.07283 (2025)

    S. Bhatnagar, An agentic AI workflow to simplify parameter estimation of complex differential equation systems, arXiv preprint arXiv:2509.07283 (2025)

    work page arXiv 2025

  19. [19]

    Lupoiu, Y

    R. Lupoiu, Y. Shao, T. Dai, C. Mao, K. Ed´ ee, J. A. Fan, A multi-agentic framework for real-time, autonomous freeform metasurface design, Science Advances 11 (44) (2025) eadx8006

    work page 2025

  20. [20]

    Mathematical exploration and discovery at scale

    B. Georgiev, J. G´ omez-Serrano, T. Tao, A. Z. Wagner, Mathematical exploration and discovery at scale, arXiv preprint arXiv:2511.02864 (2025)

    work page Pith review arXiv 2025

  21. [21]

    An AI system to help scientists write expert-level empirical software

    E. Ayg¨ un, A. Belyaeva, G. Comanici, M. Coram, H. Cui, J. Garrison, R. J. A. Kast, C. Y. McLean, P. Norgaard, Z. Shamsi, et al., An AI system to help scientists write expert-level empirical software, arXiv preprint arXiv:2509.06503 (2025)

    work page Pith review arXiv 2025

  22. [22]

    Jiang, G

    Q. Jiang, G. Karniadakis, AgenticSciML: Collaborative multi-agent systems for emer- gent discovery in scientific machine learning, arXiv preprint arXiv:2511.07262 (2025)

    work page arXiv 2025

  23. [23]

    Kumar, G

    V. Kumar, G. E. Karniadakis, Toward autonomous engineering design: A knowledge- guided multi-agent framework, arXiv preprint arXiv:2511.03179 (2025)

    work page arXiv 2025

  24. [24]

    Lattimore, C

    T. Lattimore, C. Szepesv´ ari, Bandit algorithms, Cambridge University Press, 2020

    work page 2020

  25. [25]

    Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals and Systems 2 (4) (1989) 303–314

    G. Cybenko, Approximation by superpositions of a sigmoidal function, Mathematics of Control, Signals and Systems 2 (4) (1989) 303–314

    work page 1989

  26. [26]

    Hornik, M

    K. Hornik, M. Stinchcombe, H. White, Multilayer feedforward networks are universal approximators, Neural Networks 2 (5) (1989) 359–366

    work page 1989

  27. [27]

    Kolmogorov, On the representation of continuous functions of several variables as superpositions of continuous functions of one variable and additionEnglish translation: Amer

    A. Kolmogorov, On the representation of continuous functions of several variables as superpositions of continuous functions of one variable and additionEnglish translation: Amer. Math. Soc. Transl., 28: Sixteen Papers on Analysis (1963) (1957)

    work page 1963

  28. [28]

    L. Song, J. D. Toscano, L.-L. Wang, Explicit construction of approximate kolmogorov- arnold superpositions with c2-smoothness, arXiv preprint arXiv:2508.04392 (2025)

    work page internal anchor Pith review arXiv 2025

  29. [29]

    J. D. Toscano, D. T. Chen, V. Oommen, J. Darbon, G. E. Karniadakis, A variational framework for residual-based adaptivity in neural pde solvers and operator learning, arXiv preprint arXiv:2509.14198 (2025)

    work page arXiv 2025

  30. [30]

    URLhttps://openai.com/index/gpt-5-1/ 31

    OpenAI, Gpt-5.1: A smarter, more conversational chatgpt, OpenAI BlogReleased November 12, 2025 (November 2025). URLhttps://openai.com/index/gpt-5-1/ 31

    work page 2025

  31. [31]

    DeepMind, Gemini 3 technical report, Tech

    G. DeepMind, Gemini 3 technical report, Tech. rep., Google, released November 18, 2025 (November 2025). URLhttps://blog.google/technology/ai/

    work page 2025

  32. [32]

    rep., Anthropic, released November 24, 2025 (November 2025)

    Anthropic, Claude sonnet 4.5 system card, Tech. rep., Anthropic, released November 24, 2025 (November 2025). URLhttps://www.anthropic.com/claude-sonnet-4-5-system-card

    work page 2025

  33. [33]

    rep., xAI, released November 17, 2025 (November 2025)

    xAI, Grok 4.1 model card, Tech. rep., xAI, released November 17, 2025 (November 2025). URLhttps://data.x.ai/2025-11-17-grok-4-1-model-card.pdf

    work page 2025

  34. [34]

    J. Shen, T. Tang, L.-L. Wang, Spectral methods: algorithms, analysis and applications, Vol. 41, Springer Science & Business Media, 2011

    work page 2011

  35. [35]

    Cockburn, C.-W

    B. Cockburn, C.-W. Shu, The runge–kutta discontinuous galerkin method for conserva- tion laws v: multidimensional systems, Journal of computational physics 141 (2) (1998) 199–224

    work page 1998

  36. [36]

    Cockburn, G

    B. Cockburn, G. E. Karniadakis, C.-W. Shu, Discontinuous Galerkin methods: theory, computation and applications, Vol. 11, Springer Science & Business Media, 2012

    work page 2012

  37. [37]

    M. J. Berger, P. Colella, Local adaptive mesh refinement for shock hydrodynamics, Journal of computational Physics 82 (1) (1989) 64–84

    work page 1989

  38. [38]

    Persson, J

    P.-O. Persson, J. Peraire, Sub-cell shock capturing for discontinuous galerkin methods, in: 44th AIAA aerospace sciences meeting and exhibit, 2006, p. 112

    work page 2006

  39. [39]

    Hennemann, A

    S. Hennemann, A. M. Rueda-Ram´ ırez, F. J. Hindenlang, G. J. Gassner, A provably entropy stable subcell shock capturing approach for high order split form dg for the compressible euler equations, Journal of Computational Physics 426 (2021) 109935

    work page 2021

  40. [40]

    R. J. LeVeque, Finite volume methods for hyperbolic problems, Vol. 31, Cambridge university press, 2002

    work page 2002

  41. [41]

    Ranocha, M

    H. Ranocha, M. Schlottke-Lakemper, A. R. Winters, E. Faulhaber, J. Chan, G. J. Gassner, Adaptive numerical simulations with trixi. jl: A case study of julia for scientific computing, arXiv preprint arXiv:2108.06476 (2021)

    work page arXiv 2021

  42. [42]

    Ranocha, Generalised summation-by-parts operators and entropy stability of numer- ical methods for hyperbolic balance laws, Cuvillier Verlag, 2018

    H. Ranocha, Generalised summation-by-parts operators and entropy stability of numer- ical methods for hyperbolic balance laws, Cuvillier Verlag, 2018

    work page 2018

  43. [43]

    D. A. Kopriva, Implementing spectral methods for partial differential equations: Algo- rithms for scientists and engineers, Springer Science & Business Media, 2009

    work page 2009

  44. [44]

    J. D. Toscano, V. Oommen, A. J. Varghese, Z. Zou, N. Ahmadi Daryakenari, C. Wu, G. E. Karniadakis, From pinns to pikans: Recent advances in physics-informed machine learning, Machine Learning for Computational Science and Engineering 1 (1) (2025) 1–43. 32

    work page 2025

  45. [45]

    Cai, Z.-Q

    W. Cai, Z.-Q. J. Xu, Multi-scale deep neural networks for solving high dimensional PDEs, arXiv preprint arXiv:1910.11710 (2019)

    work page arXiv 1910

  46. [46]

    S. Wang, H. Wang, P. Perdikaris, On the eigenvector bias of Fourier feature networks: From regression to solving multi-scale PDEs with physics-informed neural networks, arXiv preprint arXiv:2012.10047 (2020)

    work page arXiv 2012

  47. [47]

    S. Wang, B. Li, Y. Chen, P. Perdikaris, PirateNets: Physics-informed Deep Learning with Residual Adaptive Networks, arXiv preprint arXiv:2402.00326 (2024)

    work page arXiv 2024

  48. [48]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halverson, M. Soljaˇ ci´ c, T. Y. Hou, M. Tegmark, KAN: Kolmogorov-Arnold Networks, arXiv preprint arXiv:2404.19756 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  49. [49]

    Shukla, J

    K. Shukla, J. D. Toscano, Z. Wang, Z. Zou, G. E. Karniadakis, A comprehensive and FAIR comparison between MLP and KAN representations for differential equations and operator networks, Computer Methods in Applied Mechanics and Engineering 431 (2024) 117290

    work page 2024

  50. [50]

    J. D. Toscano, L.-L. Wang, G. E. Karniadakis, KKANs: Kurkova-Kolmogorov-Arnold networks and their learning dynamics, Neural Networks (2025) 107831

    work page 2025

  51. [51]

    J. D. Toscano, Y. Guo, Z. Wang, Y. Mori, M. Vaezi, G. E. Karniadakis, K. A. Boster, D. H. Kelley, MR-AIV reveals in-vivo brain-wide fluid flow with physics-informed ai, bioRxiv (2025) 2025–07

    work page 2025

  52. [52]

    S. Wang, Y. Teng, P. Perdikaris, Understanding and mitigating gradient flow pathologies in physics-informed neural networks, SIAM Journal on Scientific Computing 43 (5) (2021) A3055–A3081

    work page 2021

  53. [53]

    J. D. Toscano, T. K¨ aufer, Z. Wang, M. Maxey, C. Cierpka, G. E. Karniadakis, AIVT: Inference of turbulent thermal convection from measured 3d velocity data by physics- informed kolmogorov-arnold networks, Science advances 11 (19) (2025) eads5236

    work page 2025

  54. [54]

    J. D. Toscano, C. Wu, A. Ladr´ on-de Guevara, T. Du, M. Nedergaard, D. H. Kelley, G. E. Karniadakis, K. A. Boster, Inferring in vivo murine cerebrospinal fluid flow using ar- tificial intelligence velocimetry with moving boundaries and uncertainty quantification, Interface Focus 14 (6) (2024) 20240030

    work page 2024

  55. [55]

    L. D. McClenny, U. M. Braga-Neto, Self-adaptive physics-informed neural networks, Journal of Computational Physics 474 (2023) 111722

    work page 2023

  56. [56]

    S. J. Anagnostopoulos, J. D. Toscano, N. Stergiopulos, G. E. Karniadakis, Residual- based attention in physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 421 (2024) 116805

    work page 2024

  57. [57]

    S. J. Anagnostopoulos, J. D. Toscano, N. Stergiopulos, G. E. Karniadakis, Learning in pinns: Phase transition, diffusion equilibrium, and generalization, Neural Networks (2025) 107983. 33

    work page 2025

  58. [58]

    W. Chen, A. A. Howard, P. Stinis, Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks, Journal of Computational Physics (2025) 114226

    work page 2025

  59. [59]

    C. Wu, J. D. Toscano, K. Shukla, Y. Chen, A. Shahmohammadi, E. Raymond, T. Toupy, N. Nazemifard, C. Papageorgiou, G. E. Karniadakis, Fmenets: Flow, material, and energy networks for non-ideal plug flow reactor design, arXiv preprint arXiv:2505.20300 (2025)

    work page arXiv 2025

  60. [60]

    L. Lu, X. Meng, Z. Mao, G. E. Karniadakis, DeepXDE: A deep learning library for solving differential equations, SIAM Review 63 (1) (2021) 208–228

    work page 2021

  61. [61]

    C. Wu, M. Zhu, Q. Tan, Y. Kartha, L. Lu, A comprehensive study of non-adaptive and residual-based adaptive sampling for physics-informed neural networks, Computer Methods in Applied Mechanics and Engineering 403 (2023) 115671

    work page 2023

  62. [62]

    W. Chen, A. Howard, P. Stinis, Self-adaptive weighting and sampling for physics- informed neural networks, arXiv preprint arXiv:2511.05452 (2025)

    work page arXiv 2025

  63. [63]

    Z. Gao, L. Yan, T. Zhou, Failure-informed adaptive sampling for pinns, SIAM Journal on Scientific Computing 45 (4) (2023) A1971–A1994

    work page 2023

  64. [64]

    Z. Gao, T. Tang, L. Yan, T. Zhou, Failure-informed adaptive sampling for PINNs, part ii: combining with re-sampling and subset simulation, Communications on Applied Mathematics and Computation (2023) 1–22

    work page 2023

  65. [65]

    J. F. Urb´ an, P. Stefanou, J. A. Pons, Unveiling the optimization process of Physics In- formed Neural Networks: How accurate and competitive can PINNs be?, arXiv preprint arXiv:2405.04230 (2024)

    work page arXiv 2024

  66. [66]

    Kiyani, K

    E. Kiyani, K. Shukla, J. F. Urb´ an, J. Darbon, G. E. Karniadakis, Optimizing the opti- mizer for physics-informed neural networks and kolmogorov-arnold networks, Computer Methods in Applied Mechanics and Engineering 446 (2025) 118308

    work page 2025

  67. [67]

    N. A. Daryakenari, K. Shukla, G. E. Karniadakis, Representation meets optimization: Training PINNs and PIKANs for gray-box discovery in systems pharmacology, arXiv preprint arXiv:2504.07379 (2025)

    work page arXiv 2025

  68. [68]

    Y. Wang, Z. Liu, Z. Li, A. Anandkumar, T. Y. Hou, High precision pinns in un- bounded domains: application to singularity formulation in pdes, arXiv preprint arXiv:2506.19243 (2025)

    work page arXiv 2025

  69. [69]

    S. Wang, A. K. Bhartari, B. Li, P. Perdikaris, Gradient alignment in physics- informed neural networks: A second-order optimization perspective, arXiv preprint arXiv:2502.00604 (2025)

    work page arXiv 2025

  70. [70]

    W. Chen, A. A. Howard, P. Stinis, Self-adaptive weights based on balanced residual decay rate for physics-informed neural networks and deep operator networks, Journal of Computational Physics (2025) 114226. 34

    work page 2025

  71. [71]

    Takamoto, T

    M. Takamoto, T. Praditia, R. Leiteritz, D. MacKinlay, F. Alesiani, D. Pfl¨ uger, M. Niepert, Pdebench: An extensive benchmark for scientific machine learning, Ad- vances in Neural Information Processing Systems 35 (2022) 1596–1611

    work page 2022

  72. [72]

    J. Kim, K. Lee, D. Lee, S. Y. Jhin, N. Park, DPM: A novel training method for physics- informed neural networks in extrapolation, in: Proceedings of the AAAI conference on artificial intelligence, Vol. 35, 2021, pp. 8146–8154

    work page 2021

  73. [73]

    U. Ghia, K. N. Ghia, C. Shin, High-re solutions for incompressible flow using the navier- stokes equations and a multigrid method, Journal of computational physics 48 (3) (1982) 387–411

    work page 1982

  74. [74]

    E. L. Allgower, K. Georg, Introduction to numerical continuation methods, SIAM, 2003

    work page 2003

  75. [75]

    Martin, M

    B. Martin, M. Buhmann, J. Ablowitz, Radial basis functions: theory and implementa- tions, Cambridge University (ISBN: 0-521-63338-9.) (2003) 5

    work page 2003

  76. [76]

    L. Lu, P. Jin, G. E. Karniadakis, DeepOnet: Learning nonlinear operators for identifying differential equations based on the universal approximation theorem of operators, arXiv preprint arXiv:1910.03193 (2019)

    work page internal anchor Pith review Pith/arXiv arXiv 1910

  77. [77]

    S. Lee, Y. Shin, On the training and generalization of deep operator networks, SIAM Journal on Scientific Computing 46 (4) (2024) C273–C296

    work page 2024

  78. [78]

    Bozorgasl and H

    Z. Bozorgasl, H. Chen, Wav-kan: Wavelet kolmogorov-arnold networks, 2024, arXiv preprint arXiv:2405.12832

    work page arXiv 2024

  79. [79]

    Peyvan, V

    A. Peyvan, V. Oommen, A. D. Jagtap, G. E. Karniadakis, Riemannonets: Interpretable neural operators for riemann problems, Computer Methods in Applied Mechanics and Engineering 426 (2024) 116996

    work page 2024

  80. [80]

    S. Cai, H. Li, F. Zheng, F. Kong, M. Dao, G. E. Karniadakis, S. Suresh, Artificial intelligence velocimetry and microaneurysm-on-a-chip for three-dimensional analysis of blood flow in physiology and disease, Proceedings of the National Academy of Sciences 118 (13) (2021)

    work page 2021

Showing first 80 references.