pith. sign in

arxiv: 2605.04134 · v1 · submitted 2026-05-05 · 💻 cs.LG

Model synthesis and identifiability analysis of stiff chemical reaction systems with inVAErt networks

Sreejata Dey , Guoxiang Grayson Tong , Jonathan F. MacArt , Daniele E. Schiavazzi This is my paper

Pith reviewed 2026-05-08 17:52 UTC · model grok-4.3

classification 💻 cs.LG
keywords chemical kineticsstiff ODEsneural emulatorsinverse problemsparameter identifiabilityreaction rate inferencespecies concentrationsinVAErt networks
0
0 comments X

The pith

Neural emulators for stiff chemical reaction systems recover manifolds of non-identifiable reaction rates from species concentrations.

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

The paper trains conditional residual networks and LSTM architectures to create fast, accurate replicas of families of stiff ordinary differential equations that describe chemical reactions when reaction rates vary. It then applies inVAErt networks to the inverse problem of recovering the rates, integration time, and initial conditions that produce a given set of concentration trajectories. Demonstrations on reversible and irreversible kinetics cover systems from 2 to 20 equations and 3 to 25 parameters, with relative root-mean-square errors between 10^{-5} and 10^{-3}. The recovered sets of non-identifiable rates match exact analytical results in simple cases and local identifiability analysis in larger ones.

Core claim

Conditional residual networks and long-short term memory architectures serve as data-driven emulators for families of stiff reaction ODEs under varying rates. inVAErt networks then solve the ill-posed inverse problem of inferring reaction rates, integration time, and initial conditions from target species concentrations. On systems spanning 2 to 20 differential equations, 3 to 20 species, and 3 to 25 rate parameters, the emulators achieve relative root mean squared errors from 10^{-5} in low dimensions to 10^{-3} in an air-pollution model and a hydrogen-air system. Manifolds of non-identifiable rates recovered this way can be verified analytically for simple systems and remain consistent in

What carries the argument

inVAErt networks applied to the inverse mapping from observed species concentrations back to reaction rates, integration time, and initial conditions.

If this is right

  • Fast evaluation of entire families of stiff chemical models becomes possible without repeated numerical integration.
  • The inverse problem of recovering parameters from concentration data becomes tractable even when the mapping is many-to-one.
  • Non-identifiability of reaction rates can be quantified directly from data rather than through separate symbolic or local linear analysis.
  • The same workflow scales from two-equation toy systems to twenty-equation air-pollution and combustion models.

Where Pith is reading between the lines

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

  • The emulators could be embedded inside larger optimization or control loops that require repeated forward evaluations of chemical kinetics.
  • Extending the approach to experimental concentration time series would test whether the recovered manifolds remain consistent when measurement noise and model mismatch are present.
  • Hybrid schemes that combine the data-driven emulators with physics-based integrators for a subset of species could reduce error in the highest-dimensional cases.

Load-bearing premise

The trained neural emulators remain accurate across the full range of reaction rates and integration times without significant distribution shift or extrapolation error.

What would settle it

Analytic computation of the non-identifiable rate manifold for a simple reversible two-equation system followed by a direct numerical comparison with the manifold produced by the trained inVAErt network.

Figures

Figures reproduced from arXiv: 2605.04134 by Daniele E. Schiavazzi, Guoxiang Grayson Tong, Jonathan F. MacArt, Sreejata Dey.

Figure 1
Figure 1. Figure 1: Solution plots for the Robertson problem: parameters perturbed 100 times in range view at source ↗
Figure 2
Figure 2. Figure 2: Eigenvalues and stiffness ratio for the Robertson problem. view at source ↗
Figure 3
Figure 3. Figure 3: POLLU system solutions under 30 different parameter perturbations within a view at source ↗
Figure 4
Figure 4. Figure 4: Eigenvalues and stiffness ratio for the POLLU system. view at source ↗
Figure 5
Figure 5. Figure 5: Baseline solutions for the reversible and irreversible systems. view at source ↗
Figure 6
Figure 6. Figure 6: Solutions for 30 perturbed trajectories corresponding to each of three different initial conditions for the (a) view at source ↗
Figure 7
Figure 7. Figure 7: Solution plots for the hydrogen-air system, with initial conditions as in ( view at source ↗
Figure 8
Figure 8. Figure 8: Evolution of the temperature of the hydrogen-air system, with initial conditions as in ( view at source ↗
Figure 9
Figure 9. Figure 9: Solution plots for a parametric family of hydrogen-air systems. Rate parameters are perturbed 30 times view at source ↗
Figure 10
Figure 10. Figure 10: Time sampling strategies illustrated using the Robertson problem. view at source ↗
Figure 11
Figure 11. Figure 11: Training dataset generation for the emulator. The time instances in red are selected using one of the view at source ↗
Figure 12
Figure 12. Figure 12: Schematic of all components of an inVAErt network and their interactions. view at source ↗
Figure 13
Figure 13. Figure 13: Schematic of the ResNet emulator N N e. 4.2.2 Long-short term memory (LSTM) emulator The residuals in the emulator are dependent on ∆t, hence we are required to provide data with constant time-step during training. However, on account of the stiffness of our systems, we often have to choose a very small ∆t to capture the trends in the data. Consequently, errors may accumulate as we perform predictions aut… view at source ↗
Figure 14
Figure 14. Figure 14: Schematic of the LSTM emulator N N e. The quantities c and h represent cell and hidden states, respectively. obtained by concatenating the parameters k with the encoded hidden state, and encoding this latter into an updated hidden state through a feed-forward network. The conditional hidden state and cell state are then passed on to the LSTM decoder. The decoder takes in the final step data y(t − ∆t) and … view at source ↗
Figure 15
Figure 15. Figure 15: Rollout mechanism for ResNet architecture. Blue and red dots indicate input data and predictions, view at source ↗
Figure 16
Figure 16. Figure 16: Rollout mechanism for LSTM architecture. view at source ↗
Figure 17
Figure 17. Figure 17: Rollout errors for the Robertson problem. view at source ↗
Figure 18
Figure 18. Figure 18: Rollout errors for the POLLU system. 5.1.3 Systems with Reversible and Irreversible Kinetics The dataset for the emulator was trained using 30 time points in t ∈ [0, 10] per simulation, using log-sampling and ∆t = 10−3 . The rate parameters k have been generated by ±50% perturbations with respect to the nominal set k ∗ defined in (7) and (10). Additionally, the initial condition y(0) was also randomly sel… view at source ↗
Figure 19
Figure 19. Figure 19: Rollout predictions for the reversible and irreversible systems. The figure shows the system’s evolution for view at source ↗
Figure 20
Figure 20. Figure 20: LSTM Emulator rollout time traces for the hydrogen-air problem using three different combinations view at source ↗
Figure 21
Figure 21. Figure 21: Relative and absolute rollout errors in the hydrogen-air system for one of the ( view at source ↗
Figure 22
Figure 22. Figure 22: Trajectory reconstructions for the Robertson system at a randomly selected view at source ↗
Figure 23
Figure 23. Figure 23: Parallel chart for the Robertson problem parameters ( view at source ↗
Figure 24
Figure 24. Figure 24: Correlations between the Robertson problem parameters ( view at source ↗
Figure 25
Figure 25. Figure 25: POLLU system trajectories inverted from a randomly chosen view at source ↗
Figure 26
Figure 26. Figure 26: Parallel charts for the POLLU problem parameters ( view at source ↗
Figure 27
Figure 27. Figure 27: FIM eigenvalue decay for the POLLU system. Different colors are based on view at source ↗
Figure 28
Figure 28. Figure 28: Radar plots of FIM singular eigenvectors for different view at source ↗
Figure 29
Figure 29. Figure 29: Trajectory reconstructions for the reversible system from 300 latent space samples. We used a randomly view at source ↗
Figure 30
Figure 30. Figure 30: Parallel chart for the reversible system parameters ( view at source ↗
Figure 31
Figure 31. Figure 31: Correlations between the reversible system parameters for the view at source ↗
Figure 32
Figure 32. Figure 32: FIM eigenvalue decay for the reversible system, corresponding to 300 view at source ↗
Figure 33
Figure 33. Figure 33: Radar plots for the singular eigenvectors at different view at source ↗
Figure 34
Figure 34. Figure 34: DSS for the reversible system with larger priors ( view at source ↗
Figure 35
Figure 35. Figure 35: Irreversible system trajectories reconstructed using a randomly chosen view at source ↗
Figure 36
Figure 36. Figure 36: Parallel chart showing the irreversible system parameters ( view at source ↗
Figure 37
Figure 37. Figure 37: Correlations between parameters of the irreversible system determined by inversion from the same view at source ↗
Figure 38
Figure 38. Figure 38: Hydrogen-air system solutions corresponding to 200 latent space samples, and a randomly chosen view at source ↗
Figure 39
Figure 39. Figure 39: Parallel charts for the hydrogen-air kinetics problem parameters ( view at source ↗
Figure 40
Figure 40. Figure 40: FIM eigenvalue decay for the hydrogen-air system corresponding to a view at source ↗
Figure 41
Figure 41. Figure 41: Radar plots for the singular eigenvectors of the hydrogen-air system at view at source ↗
Figure 42
Figure 42. Figure 42: DSS of hydrogen-air system using target concentrations selected from pre- and post-ignition conditions. view at source ↗
Figure 43
Figure 43. Figure 43: Percentage of nearly-identifiable prior parameter combinations identified by DSS vs. view at source ↗
Figure 44
Figure 44. Figure 44: Trajectories of the POLLU system obtained by simulating 100 instances of the system corresponding to view at source ↗
read the original abstract

We consider the problem of learning data-driven replicas for stiff systems of ordinary differential equations arising in chemical kinetics that can be evaluated with high computational efficiency. We first focus on training emulators for families of reaction equations under varying reaction rates, using conditional residual networks or long-short term memory architectures. We then apply a recently proposed data-driven framework known as ``inVAErt networks'' to address the ill-posed inverse problem of inferring reaction rates, integration time, and possibly initial conditions from a target set of species concentrations - a problem that has received relatively little attention in the literature. The proposed approach is demonstrated on chemical systems with reversible and irreversible kinetics, spanning 2 to 20 differential equations, 3 to 20 chemical species, and 3 to 25 reaction rate parameters. Relative root mean squared errors produced by the proposed emulators range from $10^{-5}$ for lower-dimensional systems to $10^{-4}$ and $10^{-3}$ for an air pollution model and a hydrogen-air reaction system, respectively. Manifolds of non-identifiable reaction rates recovered by the proposed approach can be analytically verified for simple systems and are consistent with local identifiability analysis in higher dimensions.

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 / 1 minor

Summary. The paper proposes training conditional residual networks or LSTMs as efficient emulators for families of stiff chemical ODEs under varying reaction rates, then applies inVAErt networks to solve the inverse problem of recovering reaction rates, integration times, and initial conditions from observed species concentrations. Demonstrations span systems with 2-20 equations and 3-25 parameters, with reported relative RMSEs from 10^{-5} (low-dimensional) to 10^{-3} (hydrogen-air system). The central claim is that manifolds of non-identifiable rates recovered via this approach can be analytically verified for simple systems and are consistent with local identifiability analysis in higher dimensions.

Significance. If the emulators remain faithful across the full parameter and time ranges and the inversion accurately isolates true non-identifiability, the work would offer a scalable data-driven route to model synthesis and identifiability analysis for stiff kinetics, useful in combustion and atmospheric chemistry where direct integration is expensive. The analytic verification on simple cases and consistency checks with local analysis are concrete strengths that could support broader adoption if the generalization claims hold.

major comments (2)
  1. [Abstract and Results] The manuscript reports relative RMSEs of 10^{-5} to 10^{-3} for the emulators but provides no quantitative tests of emulator fidelity for reaction rates or integration times outside the training support. In stiff systems, even modest extrapolation can produce qualitatively different trajectories; without such tests the claim that recovered manifolds reflect true non-identifiability (rather than artifacts of the learned map) remains unsubstantiated.
  2. [Identifiability Analysis] The consistency statement with local identifiability analysis in higher dimensions is load-bearing for the paper's identifiability contribution, yet the text does not detail the comparison procedure, metrics, or figures that establish agreement between the inVAErt posterior and the local analysis.
minor comments (1)
  1. [Methods] Provide explicit training protocols, data-exclusion criteria, and hyperparameter choices for the conditional residual nets and LSTMs to allow independent verification of the reported error levels.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We appreciate the referee's constructive feedback on our manuscript. The comments highlight important aspects regarding the robustness of our emulator and the transparency of our identifiability analysis. We will revise the paper to include the requested tests and details, as detailed in the point-by-point responses below.

read point-by-point responses

  1. Referee: [Abstract and Results] The manuscript reports relative RMSEs of 10^{-5} to 10^{-3} for the emulators but provides no quantitative tests of emulator fidelity for reaction rates or integration times outside the training support. In stiff systems, even modest extrapolation can produce qualitatively different trajectories; without such tests the claim that recovered manifolds reflect true non-identifiability (rather than artifacts of the learned map) remains unsubstantiated.

    Authors: We agree with the referee that explicit tests of emulator performance outside the training support are important to substantiate the claims, particularly given the sensitivity of stiff systems to parameter variations. Although the training distributions were chosen to encompass the expected ranges for the chemical systems considered, the manuscript does not report out-of-sample evaluations. In the revised version, we will add quantitative results for emulator fidelity on extrapolated parameter values and integration times for at least the low-dimensional systems, and discuss how these affect the recovered manifolds in the inverse problem. revision: yes

  2. Referee: [Identifiability Analysis] The consistency statement with local identifiability analysis in higher dimensions is load-bearing for the paper's identifiability contribution, yet the text does not detail the comparison procedure, metrics, or figures that establish agreement between the inVAErt posterior and the local analysis.

    Authors: The referee is correct that the current text lacks sufficient detail on the comparison between the inVAErt-derived non-identifiable manifolds and the local identifiability analysis. The manuscript states consistency but omits the specific methodology (e.g., computation of the sensitivity matrix or null space of the Fisher information matrix), the quantitative metrics (such as subspace angles or sample overlap), and supporting figures. We will revise the identifiability analysis section to include a full description of the local analysis procedure, the metrics employed, and additional figures illustrating the agreement for the higher-dimensional cases (e.g., the 20-equation system). revision: yes

Circularity Check

0 steps flagged

Minor self-citation of inVAErt framework; central claims remain independently verifiable

full rationale

The derivation chain trains conditional residual or LSTM emulators on families of stiff ODE trajectories and then applies the cited inVAErt inversion to recover rate manifolds. For low-dimensional cases the recovered manifolds are checked against direct analytic non-identifiability conditions, and higher-dimensional results are compared to standard local identifiability analysis; neither step reduces the claimed manifolds to quantities defined by the network weights themselves. The inVAErt reference is external and not load-bearing for the analytic verification step, producing only a minor self-citation score.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The work rests on the assumption that neural networks can faithfully approximate the solution operator of stiff ODE families and that inVAErt can meaningfully characterize identifiability; no new physical entities are postulated.

axioms (1)
  • domain assumption Neural networks trained on sampled trajectories generalize to unseen rate combinations within the training distribution.
    Required for the emulator to serve as a reliable surrogate across the stated ranges of 3-25 rate parameters.

pith-pipeline@v0.9.0 · 5530 in / 1225 out tokens · 45451 ms · 2026-05-08T17:52:46.748413+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.

Reference graph

Works this paper leans on

44 extracted references · 17 canonical work pages · 1 internal anchor

  1. [1]

    ChemNODE: A neural ordinary differential equations framework for efficient chemical kinetic solvers.Energy and AI, 7:100118, 2022

    Opeoluwa Owoyele and Pinaki Pal. ChemNODE: A neural ordinary differential equations framework for efficient chemical kinetic solvers.Energy and AI, 7:100118, 2022. doi: 10.1016/j.egyai.2021.100118

    work page doi:10.1016/j.egyai.2021.100118 2022

  2. [2]

    Journal of Mathematical Physics64(9), 091902 (2023) https://doi.org/10.1063/5

    Suyong Kim, Weiqi Ji, Sili Deng, Yingbo Ma, and Christopher Rackauckas. Stiff neural ordinary differential equations.Chaos: An Interdisciplinary Journal of Nonlinear Science, 31(9):093122, 2021. doi: 10.1063/5. 0060697

    work page doi:10.1063/5 2021

  3. [3]

    Tadbhagya Kumar, Anuj Kumar, and Pinaki Pal. A posteriori evaluation of a physics-constrained neural ordinary differential equations approach coupled with CFD solver for modeling stiff chemical kinetics.arXiv preprint arXiv:2312.00038, 2024

    work page arXiv 2024

  4. [4]

    Extension of the trotter- ized unitary coupled cluster to triple excitations

    Weiqi Ji, Weilun Qiu, Zhiyu Shi, Shaowu Pan, and Sili Deng. Stiff-PINN: Physics-informed neural network for stiff chemical kinetics.The Journal of Physical Chemistry A, 125(36):8098–8106, 2021. doi: 10.1021/acs.jpca. 1c05102

    work page doi:10.1021/acs.jpca 2021

  5. [5]

    Multiscale physics-informed neural networks for stiff chemical kinetics.The Journal of Physical Chemistry A, 126(45):8534–8543, 2022

    Yuting Weng and Dezhi Zhou. Multiscale physics-informed neural networks for stiff chemical kinetics.The Journal of Physical Chemistry A, 126(45):8534–8543, 2022. doi: 10.1021/acs.jpca.2c06513

    work page doi:10.1021/acs.jpca.2c06513 2022

  6. [6]

    Christo, Assaad R

    Farid C. Christo, Assaad R. Masri, Eduardo Mario Nebot, and Stephen B. Pope. An integrated PDF/neural network approach for simulating turbulent reacting systems.Symposium (International) on Combustion, 26(1): 43–48, 1996. doi: 10.1016/S0082-0784(96)80198-6. 31

    work page doi:10.1016/s0082-0784(96)80198-6 1996

  7. [7]

    Blasco, N

    J.A. Blasco, N. Fueyo, C. Dopazo, and J. Ballester. Modelling the temporal evolution of a reduced combustion chemical system with an artificial neural network.Combustion and Flame, 113(1-2):38–52, 1998. doi: 10.1016/ S0010-2180(97)00211-3

    1998

  8. [8]

    Enforcing physical conservation in neural network surrogate models for complex chemical kinetics.Combustion and Flame, 275: 114105, 2025

    Tinghao Wang, Yuxiao Yi, Junjie Yao, Zhi-Qin John Xu, Tianhan Zhang, and Zheng Chen. Enforcing physical conservation in neural network surrogate models for complex chemical kinetics.Combustion and Flame, 275: 114105, 2025. doi: 10.1016/j.combustflame.2025.114105

    work page doi:10.1016/j.combustflame.2025.114105 2025

  9. [9]

    A multi- scale sampling method for accurate and robust deep neural network to predict combustion chemical kinetics

    Tianhan Zhang, Yuxiao Shu, Xi Jiang, Pengfei Yi, Chengxi Xu, Ziming Wang, and Huangwei Wang. A multi- scale sampling method for accurate and robust deep neural network to predict combustion chemical kinetics. Combustion and Flame, 245:112319, 2022. doi: 10.1016/j.combustflame.2022.112319

    work page doi:10.1016/j.combustflame.2022.112319 2022

  10. [10]

    Efficient machine learning based surrogate models for surface kinetics by approximating the rates of the rate-determining steps.Chemical Engineering Science, 262:117964, 2022

    Felix A D¨ oppel and Martin Votsmeier. Efficient machine learning based surrogate models for surface kinetics by approximating the rates of the rate-determining steps.Chemical Engineering Science, 262:117964, 2022. doi: 10.1016/j.ces.2022.117964

    work page doi:10.1016/j.ces.2022.117964 2022

  11. [11]

    ChemKANs for combustion chemistry modeling and acceler- ation.Physical Chemistry Chemical Physics, 27(33):17313–17330, 2025

    Benjamin C Koenig, Suyong Kim, and Sili Deng. ChemKANs for combustion chemistry modeling and acceler- ation.Physical Chemistry Chemical Physics, 27(33):17313–17330, 2025

    2025

  12. [12]

    Learning kinetics from non-ideal reactors by implicitly solved finite volumes and global reaction neural networks.Chemical Engineering Journal, 522:166749, 2025

    Tim Kircher and Martin Votsmeier. Learning kinetics from non-ideal reactors by implicitly solved finite volumes and global reaction neural networks.Chemical Engineering Journal, 522:166749, 2025. doi: 10.1016/j.cej.2025. 166749

    work page doi:10.1016/j.cej.2025 2025

  13. [13]

    A data-driven reduced-order model for stiff chemical kinetics using dynamics-informed training.Energy and AI, 15:100325, 2024

    Vijayamanikandan Vijayarangan, Harshavardhana A Uranakara, Shivam Barwey, Riccardo Malpica Galassi, Mohammad Rafi Malik, Mauro Valorani, Venkat Raman, and Hong G Im. A data-driven reduced-order model for stiff chemical kinetics using dynamics-informed training.Energy and AI, 15:100325, 2024

    2024

  14. [14]

    Jagtap, Hessam Babaee, Bryan T

    Somdatta Goswami, Ameya D. Jagtap, Hessam Babaee, Bryan T. Susi, and George Em Karniadakis. Learning stiff chemical kinetics using extended deep neural operators.Computer Methods in Applied Mechanics and Engineering, 419:116674, 2024. doi: 10.1016/j.cma.2023.116674

    work page doi:10.1016/j.cma.2023.116674 2024

  15. [15]

    Combustion chemistry acceleration with DeepONets

    Tarek Echekki, Majid Farjoo, and Opeoluwa Owoyele. Combustion chemistry acceleration with DeepONets. Fuel, 365:131212, 2024. doi: 10.1016/j.fuel.2024.131212

    work page doi:10.1016/j.fuel.2024.131212 2024

  16. [16]

    Air traffic control complexity as workload driver

    Yuting Weng, Han Li, Hao Zhang, Zhi X Chen, and Dezhi Zhou. Extended Fourier neural operators to learn stiff chemical kinetics under unseen conditions.Combustion and Flame, 272:113847, 2025. doi: 10.1016/j. combustflame.2024.113847

    work page doi:10.1016/j 2025

  17. [17]

    Susi, Hessam Babaee, and George Em Karniadakis

    Ameya D Jagtap, Zhiping Mao, and George Em Karniadakis. AMORE: Adaptive multi-output operator network for stiff chemical kinetics.arXiv preprint arXiv:2510.12999, 2024

    work page internal anchor Pith review arXiv 2024

  18. [18]

    and Bekas, Costas and Lee, Alpha A

    Philippe Schwaller, Teodoro Laino, Th´ eophile Gaudin, Peter Bolgar, Christopher A. Hunter, Costas Bekas, and Alpha A. Lee. Molecular transformer: A model for uncertainty-calibrated chemical reaction prediction.ACS Central Science, 5(9):1572–1583, 2019. doi: 10.1021/acscentsci.9b00576

    work page doi:10.1021/acscentsci.9b00576 2019

  19. [19]

    Machine learning of reaction properties via learned representations of the condensed graph of reaction.Journal of Chemical Information and Modeling, 62(9):2101–2110, 2022

    Esther Heid and William H Green. Machine learning of reaction properties via learned representations of the condensed graph of reaction.Journal of Chemical Information and Modeling, 62(9):2101–2110, 2022. doi: 10.1021/acs.jcim.1c00975

    work page doi:10.1021/acs.jcim.1c00975 2022

  20. [20]

    Parameter estimation for models of chemical reaction networks from experimental data of reaction rates.International Journal of Control, 96(2):392–407, 2023

    Manvel Gasparyan, Arnout Van Messem, and Shodhan Rao. Parameter estimation for models of chemical reaction networks from experimental data of reaction rates.International Journal of Control, 96(2):392–407, 2023

    2023

  21. [21]

    Laser diagnostics and minor species detection in combustion using resonant four-wave mixing.Progress in Energy and Combustion Science, 37(5):525–564, 2011

    Johannes Kiefer and Paul Ewart. Laser diagnostics and minor species detection in combustion using resonant four-wave mixing.Progress in Energy and Combustion Science, 37(5):525–564, 2011

    2011

  22. [22]

    ´Erdi, J

    P. ´Erdi, J. Toth, and J. T´ oth.Mathematical Models of Chemical Reactions: Theory and Applications of Deter- ministic and Stochastic Models. Nonlinear science : theory and applications. Princeton University Press, 1989. ISBN 9780691085326

    1989

  23. [23]

    C.K. Law. Combustion physics.Cambridge Univ. Press, N. J, 2006. 32

    2006

  24. [24]

    Robertson

    H.H. Robertson. The solution of a set of reaction rate equations.Numerical Analysis: An Introduction, 178182, 1966

    1966

  25. [25]

    Springer Berlin Heidelberg New York, 1996

    Gerhard Wanner and Ernst Hairer.Solving Ordinary Differential Equations II, volume 375. Springer Berlin Heidelberg New York, 1996

    1996

  26. [26]

    Jan G. Verwer. Gauss–Seidel iteration for stiff ODEs from chemical kinetics.SIAM Journal on Scientific Computing, 15(5):1243–1250, 1994

    1994

  27. [27]

    Powers.Combustion Thermodynamics and Dynamics

    Joseph M. Powers.Combustion Thermodynamics and Dynamics. Cambridge University Press, 2016

    2016

  28. [28]

    Al-Khateeb, Joseph M

    Ashraf N. Al-Khateeb, Joseph M. Powers, and Samuel Paolucci. Analysis of the spatio-temporal scales of laminar premixed flames near equilibrium.Combustion Theory and Modelling, 17(1):76–108, 2013

    2013

  29. [29]

    Kee, Fran M

    Robert J. Kee, Fran M. Rupley, and James A. Miller. Chemkin-II: A fortran chemical kinetics package for the analysis of gas-phase chemical kinetics. Technical report, Sandia National Laboratories (SNL-CA), Livermore, CA (United States), 1989

    1989

  30. [30]

    Nonuniform random variate generation.Handbooks in Operations Research and Management Science, 13:83–121, 2006

    Luc Devroye. Nonuniform random variate generation.Handbooks in Operations Research and Management Science, 13:83–121, 2006

    2006

  31. [31]

    Schiavazzi, and George Em Karniadakis

    Mario De Florio, Zongren Zou, Daniele E. Schiavazzi, and George Em Karniadakis. Quantification of total uncertainty in the physics-informed reconstruction of CVSim-6 physiology.Philosophical Transactions A, 383 (2292):20240221, 2025

    2025

  32. [32]

    Sing Long, and Daniele E

    Guoxiang Grayson Tong, Carlos A. Sing Long, and Daniele E. Schiavazzi. InVAErt networks: A data-driven framework for model synthesis and identifiability analysis.Computer Methods in Applied Mechanics and Engi- neering, 423:116846, 2024

    2024

  33. [33]

    Density estimation using Real NVP

    Laurent Dinh, Jascha Sohl-Dickstein, and Samy Bengio. Density estimation using Real NVP. InInternational Conference on Learning Representations, 2017. URLhttps://openreview.net/forum?id=HkpbnH9lx

    2017

  34. [34]

    Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    George Papamakarios, Eric Nalisnick, Danilo Jimenez Rezende, Shakir Mohamed, and Balaji Lakshminarayanan. Normalizing flows for probabilistic modeling and inference.Journal of Machine Learning Research, 22(57):1–64, 2021

    2021

  35. [35]

    Prince, and Marcus A

    Ivan Kobyzev, Simon J.D. Prince, and Marcus A. Brubaker. Normalizing flows: An introduction and review of current methods.IEEE transactions on pattern analysis and machine intelligence, 43(11):3964–3979, 2020

    2020

  36. [36]

    Kingma and Max Welling

    Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In2nd International Conference on Learning Representations, ICLR 2014, Banff, AB, Canada, April 14-16, 2014, Conference Track Proceedings,

    2014

  37. [37]

    URLhttp://arxiv.org/abs/1312.6114

    work page Pith review arXiv

  38. [38]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, pages 770–778, 2016

    2016

  39. [39]

    Data driven governing equations approximation using deep neural networks.Journal of Computational Physics, 395:620–635, 2019

    Tong Qin, Kailiang Wu, and Dongbin Xiu. Data driven governing equations approximation using deep neural networks.Journal of Computational Physics, 395:620–635, 2019

    2019

  40. [40]

    LSTM can solve hard long time lag problems.Advances in Neural Information Processing Systems, 9, 1996

    Sepp Hochreiter and J¨ urgen Schmidhuber. LSTM can solve hard long time lag problems.Advances in Neural Information Processing Systems, 9, 1996

    1996

  41. [41]

    Lstm encoder decoder

    Lucas Kulowski. Lstm encoder decoder. GitHub repository:https://github.com/lkulowski/LSTM_encoder_ decoder, 2020

    2020

  42. [42]

    Schiavazzi

    Guoxiang Grayson Tong and Daniele E. Schiavazzi. Data-driven synchronization-avoiding algorithms in the explicit distributed structural analysis of soft tissue.Computational Mechanics, 71(3):453–479, 2023

    2023

  43. [43]

    Simone Lauria and Mohammed F. Saleh. Conditional recurrent neural networks for broad applications in nonlinear optics.Optics Express, 32(4):5582–5591, 2024

    2024

  44. [44]

    important

    Lukas Baumgartner and Peter Szmolyan. A multiparameter singular perturbation analysis of the Robertson model.Studies in Applied Mathematics, 154(2):e70020, 2025. 33 A Network architecture and hyperparameters Hyperparameter choices for the Robertson, POLLU, reversible, irreversible and hydrogen-air systems are reported in Table 9, 10, 11, 12 and 13, respec...

    2025