pith. machine review for the scientific record. sign in

arxiv: 2511.07686 · v2 · pith:ID6GNMMFnew · submitted 2025-11-10 · ⚛️ physics.chem-ph · cs.LG

Kolmogorov-Arnold Chemical Reaction Neural Networks for learning pressure-dependent kinetic rate laws

Benjamin C. Koenig , Sili Deng This is my paper

Pith reviewed 2026-05-17 23:18 UTC · model grok-4.3

classification ⚛️ physics.chem-ph cs.LG
keywords chemical reaction networkspressure-dependent kineticsKolmogorov-Arnold networksmachine learning for kineticsArrhenius lawcombustion modelingdata-driven kineticsmass action kinetics
0
0 comments X

The pith

Kolmogorov-Arnold Chemical Reaction Neural Networks learn pressure-dependent kinetic rates directly from data while preserving Arrhenius and mass-action structure.

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

Standard chemical reaction neural networks follow Arrhenius and mass-action rules but cannot capture how rates change with pressure or the identity of other molecules present. The authors replace fixed kinetic parameters with functions of third-body concentrations that are learned through Kolmogorov-Arnold activations. This keeps the physical constraints and interpretability of the original framework yet removes the need for empirical falloff expressions or interpolation schemes. On two proof-of-concept reactions the resulting models reproduce observed pressure and mixture effects across wide ranges of temperature and bath gas and produce lower error than standard interpolative fits.

Core claim

KA-CRNNs generalize CRNNs by modeling each kinetic parameter as a learnable function of third-body concentrations through Kolmogorov-Arnold activations. The construction maintains strict adherence to Arrhenius temperature dependence and mass-action kinetics while permitting direct, assumption-free extraction of both global and collider-specific pressure dependence from data. Two proof-of-concept studies show that the networks recover accurate pressure-dependent kinetics over ranges of temperature, pressure, and bath-gas composition from sparse observations.

What carries the argument

Kolmogorov-Arnold activations that express each kinetic parameter as a function of third-body concentrations, keeping the overall rate law inside the Arrhenius and mass-action form.

If this is right

  • The networks reproduce pressure-dependent and collider-specific kinetics across wide ranges of temperature, pressure, and bath-gas mixtures.
  • Meaningful and generalizable models can be extracted from relatively sparse training data.
  • Prediction error is reduced by a factor of 2.88 relative to standard interpolative methods on the tested cases.
  • The approach supplies a foundation for data-driven discovery of extended kinetic behavior in complex reacting systems.

Where Pith is reading between the lines

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

  • The same structure could be inserted into existing combustion or atmospheric-chemistry mechanisms to replace manual choice of Troe or SRI parameters.
  • If the learned functions prove transferable, they might reveal systematic trends in collision efficiencies across families of reactions.
  • The method opens a route to joint inference of both rate coefficients and their pressure dependence inside a single physics-constrained model.

Load-bearing premise

Kolmogorov-Arnold activations can represent the full range of pressure and mixture dependencies without violating physical constraints or needing post-training corrections.

What would settle it

Measure the rate of one of the studied reactions at a new bath-gas composition or pressure outside the training set and check whether the KA-CRNN prediction matches the experimental value within the reported error.

Figures

Figures reproduced from arXiv: 2511.07686 by Benjamin C. Koenig, Sili Deng.

Figure 1
Figure 1. Figure 1: FIG. 1: KA-CRNN model with learnable activation [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: FIG. 3: KAN activations and individual basis functions [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 2
Figure 2. Figure 2: FIG. 2: KA-CRNN reconstructions of all 27 datasets, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. Figure 4: FIG. 4: KA-CRNN reaction rates. Synthetic data from [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
read the original abstract

Chemical Reaction Neural Networks (CRNNs) have emerged as an interpretable machine learning framework for discovering reaction kinetics directly from data, while strictly adhering to the Arrhenius and mass action laws. However, standard CRNNs cannot represent pressure-dependent or mixture-based rate behavior, which is critical in many combustion and chemical systems and typically requires empirical falloff formulations such as Troe or SRI, or data-based interpolation or polynomial fits such as PLOG or Chebyshev Polynomials. Here, we develop Kolmogorov-Arnold Chemical Reaction Neural Networks (KA-CRNNs) that generalize CRNNs by modeling each kinetic parameter as a learnable function of third-body concentrations using Kolmogorov-Arnold activations. This structure maintains the Arrhenius and mass action interpretability and physical constraints of a vanilla CRNN while enabling assumption-free inference of global and collider-specific pressure effects directly from data. Two proof-of-concept reaction studies are presented to highlight the capability of KA-CRNNs to accurately reproduce pressure-dependent and collider-specific kinetics across a range of temperatures, pressures, and bath gas mixtures, extracting meaningful and generalizable models from sparse training data and significantly outperforming interpolative approaches (2.88x reduction in MSE). The framework establishes a foundation for data-driven discovery of extended kinetic behaviors in complex reacting systems, advancing interpretable and physics-constrained approaches for chemical model inference.

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

2 major / 2 minor

Summary. The manuscript introduces Kolmogorov-Arnold Chemical Reaction Neural Networks (KA-CRNNs) as an extension of Chemical Reaction Neural Networks (CRNNs) for modeling pressure-dependent kinetic rate laws. By representing each kinetic parameter as a learnable Kolmogorov-Arnold Network (KAN) function of third-body concentrations, the approach aims to infer global and collider-specific pressure effects from data while preserving the Arrhenius form and mass-action kinetics. The authors present two proof-of-concept studies demonstrating accurate reproduction of pressure-dependent kinetics and a 2.88x MSE reduction compared to interpolative approaches.

Significance. If the central claims hold after addressing constraint enforcement, this framework would represent a meaningful step toward assumption-free yet physics-constrained discovery of extended kinetic behaviors in combustion and reacting systems. It extends the interpretability of CRNNs to pressure and mixture dependencies without relying on empirical falloff forms, and the reported MSE improvement on sparse data suggests practical utility for model inference where traditional interpolation falls short.

major comments (2)
  1. [§3] §3 (KA-CRNN formulation): the claim that the architecture 'maintains ... physical constraints' and requires 'no post-training adjustments' is not supported by any explicit mechanism ensuring correct low- and high-pressure limiting behavior. Because the KAN activations are universal approximators without built-in asymptotic guarantees, nothing in the model automatically enforces that the effective rate constant approaches the expected Lindemann or Troe limits as total concentration tends to zero or infinity; this is load-bearing for the central 'assumption-free' and 'interpretable' assertions.
  2. [§5] §5 (numerical results): the reported 2.88x MSE reduction and 'accurate reproduction' are presented without details on data splits, cross-validation procedure, error bars on the learned parameters, or quantitative extrapolation tests outside the training pressure window. Without these, it is impossible to verify that the models remain physically valid or generalizable beyond the two proof-of-concept studies.
minor comments (2)
  1. [Abstract] The abstract and introduction would benefit from a brief explicit statement of the baseline interpolative methods (PLOG, Chebyshev, etc.) used for the 2.88x comparison.
  2. [§3] Notation for the KAN layer composition and how the third-body concentration vector is fed into the activations should be given a dedicated equation for clarity.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their insightful comments, which have helped us identify areas for improvement in our manuscript on KA-CRNNs. Below, we provide point-by-point responses to the major comments. We have revised the manuscript accordingly to address the concerns raised regarding constraint enforcement and the presentation of numerical results.

read point-by-point responses

  1. Referee: [§3] §3 (KA-CRNN formulation): the claim that the architecture 'maintains ... physical constraints' and requires 'no post-training adjustments' is not supported by any explicit mechanism ensuring correct low- and high-pressure limiting behavior. Because the KAN activations are universal approximators without built-in asymptotic guarantees, nothing in the model automatically enforces that the effective rate constant approaches the expected Lindemann or Troe limits as total concentration tends to zero or infinity; this is load-bearing for the central 'assumption-free' and 'interpretable' assertions.

    Authors: We appreciate the referee pointing out the distinction between preserving the Arrhenius and mass-action structure and enforcing specific asymptotic limits. The KA-CRNN is designed to maintain the overall physical form of the rate law while allowing the pressure dependence to be learned via KANs. However, we acknowledge that without explicit constraints, the universal approximation property of KANs does not guarantee correct limiting behavior a priori. In the proof-of-concept examples, the models were trained on data covering a wide pressure range including near-limiting regimes, and post-hoc analysis shows they approach the expected limits. To better support our claims, we will revise the manuscript in §3 to include a dedicated discussion on limiting behavior, provide plots of the learned rate constants vs. concentration in the low- and high-pressure limits, and clarify that the 'no post-training adjustments' refers to not needing to modify the model after training to fit empirical falloff forms, rather than automatic enforcement of all possible limits. revision: yes

  2. Referee: [§5] §5 (numerical results): the reported 2.88x MSE reduction and 'accurate reproduction' are presented without details on data splits, cross-validation procedure, error bars on the learned parameters, or quantitative extrapolation tests outside the training pressure window. Without these, it is impossible to verify that the models remain physically valid or generalizable beyond the two proof-of-concept studies.

    Authors: We agree that providing more details on the training and validation procedures is essential for assessing the reliability of our results. In the revised version of §5, we will add: a description of how the data was split for training and testing (e.g., by pressure ranges or random sampling with multiple seeds), the cross-validation approach employed to ensure robustness, error bars or standard deviations for the reported MSE and learned parameters based on multiple runs, and results from extrapolation tests where the model is evaluated at pressures outside the training window, comparing against reference data or known physical limits. These additions will substantiate the 2.88x MSE reduction and demonstrate generalizability. revision: yes

Circularity Check

0 steps flagged

KA-CRNN derivation is self-contained without circular reductions

full rationale

The paper extends standard CRNNs—which already enforce Arrhenius form and mass-action kinetics—by replacing fixed parameters with KAN-based functions of third-body concentrations. This is an architectural generalization for learning pressure dependence directly from data, not a re-derivation of prior results. No equation reduces a claimed prediction to a fitted input by construction, no uniqueness theorem is imported from self-citation, and the two proof-of-concept studies evaluate generalization on held-out conditions rather than tautological fits. The framework therefore remains independent of its own outputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

Abstract provides limited technical detail; main additions are the KAN-based functional dependence and the claim of maintained physical constraints.

free parameters (1)
  • KAN activation parameters
    Learnable parameters within each Kolmogorov-Arnold layer that define the pressure and mixture dependence for kinetic coefficients.
axioms (1)
  • domain assumption Reaction rates must obey Arrhenius temperature dependence and mass-action concentration dependence
    Framework is explicitly constructed to preserve these laws while adding pressure dependence.

pith-pipeline@v0.9.0 · 5540 in / 1240 out tokens · 56919 ms · 2026-05-17T23:18:49.178038+00:00 · methodology

discussion (0)

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

Forward citations

Cited by 1 Pith paper

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

  1. Learning continuous state of charge dependent thermal decomposition kinetics for Li-ion cathodes using Kolmogorov-Arnold Chemical Reaction Neural Networks (KA-CRNNs)

    physics.chem-ph 2025-12 unverdicted novelty 6.0

    KA-CRNN learns continuous SOC-dependent kinetic parameters for cathode-electrolyte decomposition directly from DSC data, reproducing heat-release features across all SOCs for NCA, NM, and NMA cathodes.

Reference graph

Works this paper leans on

34 extracted references · 34 canonical work pages · cited by 1 Pith paper · 6 internal anchors

  1. [1]

    of the KAN activation, ϕi (x) = NX j=1 wψ i,j ·ψ(||x−c j||) +w b i ·b(x),(5) ψ(r) = exp(− r2 2h2 ).(6) In this formulationNis the KAN basis function grid size, andw ψ i,j andw b j are the learnable network param- eters inθthat respectively encode thejgridded basis function scales and the single scale applied to the swish

    work page

  2. [2]

    ground truth

    residual base functionb(x). The grid itself is de- fined by its uniform gridpointsc j and gridpoint spacing or RBF spreading parameterh. The KA-CRNN redefines each kinetic parameter as its own univariate function of the pressure, as is visualized in Fig. 1. Instead of the scalar parameter inference of a standard CRNN, inference in a KA-CRNN is carried out...

    work page

  3. [3]

    Ji and S

    W. Ji and S. Deng, Autonomous Discovery of Unknown Reaction Pathways from Data by Chemical Reaction Neural Network, J. Phys. Chem. A.125, 1082 (2021)

    work page 2021

  4. [4]

    C. W. Gao, J. W. Allen, W. H. Green, and R. H. West, Reaction mechanism generator: Automatic construction of chemical kinetic mechanisms, Comput. Phys. Com- mun.203, 212 (2016)

    work page 2016

  5. [5]

    S. Deng, L. Wang, S. Kim, and B. C. Koenig, Scientific machine learning in combustion for discovery, simulation, and control, Proc. Combust. Inst.41, 105796 (2025)

    work page 2025

  6. [6]

    Kumar and T

    A. Kumar and T. Echekki, Combustion chemistry accel- eration with deeponets, Fuel365, 131212 (2024)

    work page 2024

  7. [7]

    B. C. Koenig and S. Deng, Multi-target active subspaces generated using a neural network for computationally ef- ficient turbulent combustion kinetic uncertainty quan- tification in the flamelet regime, Combust. Flame258, 113015 (2023)

    work page 2023

  8. [8]

    W. Ji, F. Richter, M. J. Gollner, and S. Deng, Au- tonomous kinetic modeling of biomass pyrolysis using chemical reaction neural networks, Combust. Flame240, 111992 (2022)

    work page 2022

  9. [9]

    Q. Li, H. Chen, B. C. Koenig, and S. Deng, Bayesian chemical reaction neural network for autonomous kinetic uncertainty quantification, Phys. Chem. Chem. Phys.25, 3707 (2023)

    work page 2023

  10. [10]

    B. C. Koenig, H. Chen, Q. Li, P. Zhao, and S. Deng, Un- certain lithium-ion cathode kinetic decomposition model- ing via Bayesian chemical reaction neural networks, Proc. Combust. Inst.40, 105243 (2024)

    work page 2024

  11. [11]

    F. A. D¨ oppel and M. Votsmeier, Robust mechanism dis- covery with atom conserving chemical reaction neural networks, Proc. Combust. Inst.40, 105507 (2024)

    work page 2024

  12. [12]

    B. C. Koenig, P. Zhao, and S. Deng, Accommodating physical reaction schemes in DSC cathode thermal sta- bility analysis using chemical reaction neural networks, J. Power Sources581, 233443 (2023)

    work page 2023

  13. [13]

    B. C. Koenig, P. Zhao, and S. Deng, Comprehensive thermal-kinetic uncertainty quantification of lithium-ion battery thermal runaway via bayesian chemical reaction neural networks, Chem. Eng. J.507, 160402 (2025)

    work page 2025

  14. [14]

    H. Wang, Y. Xu, M. Wen, W. Wang, Q. Chu, S. Yan, S. Xu, and D. Chen, Kinetic modeling of cl-20 decom- position by a chemical reaction neural network, J. Anal. Appl. Pyrolysis169, 105860 (2023)

    work page 2023

  15. [15]

    W. Sun, Y. Xu, X. Chen, Q. Chu, and D. Chen, Kinetic models of hmx decomposition via chemical reaction neu- ral network, J. Anal. Appl. Pyrolysis179, 106519 (2024)

    work page 2024

  16. [16]

    X. Chen, Y. Xu, M. Wen, Y. Wang, K. Pang, S. Wang, Q. Chu, and D. Chen, EM-HyChem: Bridging molec- ular simulations and chemical reaction neural network- enabled approach to modelling energetic material chem- istry, Combust. Flame275, 114065 (2025)

    work page 2025

  17. [17]

    Bhatnagar, A

    S. Bhatnagar, A. Comerford, Z. Xu, D. B. Polato, A. Ba- naeizadeh, and A. Ferraris, Chemical Reaction Neural Networks for fitting Accelerating Rate Calorimetry data, J. Power Sources628, 235834 (2025)

    work page 2025

  18. [18]

    Zhong, W

    Y. Zhong, W. Gao, C. Li, and Y. Ding, Pyrolysis mecha- nism study on xylose by combining experiments, chemical reaction neural networks and density functional theory, Bioresour. Technol. , 132530 (2025)

    work page 2025

  19. [19]

    Stagge and R

    H. Stagge and R. G¨ uttel, The findability of microkinetic parameters by heterogeneous chemical reaction neural networks (hCRNNs), Chem. Eng. J. , 161460 (2025)

    work page 2025

  20. [20]

    Shukla, X

    J. Shukla, X. Qu, Z. Darbari, M. Iloska, J. A. Boscoboinik, and Q. Wu, Discovering CO adsorption and desorption pathways from chemical reaction neural net- work modeling of transient kinetics spectroscopy, J. Phys. Chem. Lett.16, 3562 (2024)

    work page 2024

  21. [21]

    the radiation theory of chemical action

    F. A. Lindemann, S. Arrhenius, I. Langmuir, N. Dhar, J. Perrin, and W. M. Lewis, Discussion on “the radiation theory of chemical action”, Trans. Faraday Soc.17, 598 (1922)

    work page 1922

  22. [22]

    Gilbert, K

    R. Gilbert, K. Luther, and J. Troe, Theory of thermal unimolecular reactions in the fall-off range. ii. weak col- lision rate constants, Berichte der Bunsengesellschaft f¨ ur physikalische Chemie87, 169 (1983)

    work page 1983

  23. [23]

    B. C. Koenig, S. Kim, and S. Deng, KAN-ODEs: Kol- mogorov–Arnold network ordinary differential equations for learning dynamical systems and hidden physics, Com- put. Methods Appl. Mech. Eng.432, 117397 (2024)

    work page 2024

  24. [24]

    Z. Liu, Y. Wang, S. Vaidya, F. Ruehle, J. Halver- son, M. Soljaˇ ci´ c, T. Y. Hou, and M. Tegmark, KAN: Kolmogorov-Arnold Networks, arXiv preprint arXiv:2404.19756 10.48550/arXiv.2404.19756 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2404.19756 2024

  25. [25]

    B. C. Koenig, S. Kim, and S. Deng, LeanKAN: a parameter-lean Kolmogorov-Arnold network layer with improved memory efficiency and convergence behavior, Neural Netw.192, 107883 (2025)

    work page 2025

  26. [26]

    R. T. Q. Chen, Y. Rubanova, J. Bettencourt, and D. Du- venaud, Neural Ordinary Differential Equations (2019), arxiv preprint arXiv:1806.07366

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  27. [27]

    B. C. Koenig, S. Kim, and S. Deng, ChemKANs for combustion chemistry modeling and acceleration, Phys. Chem. Chem. Phys.27, 17313 (2025)

    work page 2025

  28. [28]

    Kolmogorov-Arnold networks are radial basis function networks,

    Z. Li, Kolmogorov-Arnold Networks are Radial Ba- sis Function Networks, arxiv preprint arXiv:2405.06721 10.48550/arXiv.2405.06721 (2024)

    work page doi:10.48550/arxiv.2405.06721 2024

  29. [29]

    Searching for Activation Functions

    P. Ramachandran, B. Zoph, and Q. V. Le, Search- ing for Activation Functions (2017), arxiv preprint arXiv:1710.05941

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  30. [30]

    Rackauckas and Q

    C. Rackauckas and Q. Nie, DifferentialEquations.jl – A Performant and Feature-Rich Ecosystem for Solving Dif- ferential Equations in Julia, J. Open Res. Softw.5, 15 (2017)

    work page 2017

  31. [31]

    Forward-Mode Automatic Differentiation in Julia

    J. Revels, M. Lubin, and T. Papamarkou, Forward-Mode Automatic Differentiation in Julia (2016), arxiv preprint arXiv:1607.07892

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  32. [32]

    D. P. Kingma and J. Ba, Adam: A Method for Stochastic Optimization (2017), arxiv preprint arXiv:1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  33. [33]

    A. F. Wagner and D. M. Wardlaw, Study of the Recom- bination Reaction CH 3 + CH 3 →C 2H6. 2. Theory, J. Phys. Chem92(1988). [32]ANSYS Chemkin Theory Manual, Release 17.0, ANSYS, Inc. (2016)

    work page 1988

  34. [34]

    Interpretable Machine Learning for Science with PySR and SymbolicRegression.jl

    M. Cranmer, Interpretable Machine Learning for Sci- ence with PySR and SymbolicRegression.jl (2023), arxiv preprint arXiv:2305.01582

    work page internal anchor Pith review Pith/arXiv arXiv 2023