Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems

T0 review · 4 major / 4 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read An agentic workflow can turn a lightweight XML problem spec and a partially filled Python skeleton into a validated, differentiable, compiled parameter-estimation pipeline, recovering stiff ODE parameters to well under one percent and yield

desk verdict A plausible engineering tool with clean simulated-data demos, but the load-bearing LLM auto-correction claim rests on two anecdotes and the reproducibility link is missing. read the letter →

arxiv 2509.07283 v3 pith:LQ5746NI submitted 2025-09-08 cs.CE

Saakaar Bhatnagar This is my paper
classification cs.CE
keywords agenticAIparameterestimationordinarydifferentialequationsautomaticdifferentiationparticleswarmoptimizationneuralODEsstiffsystemsJAX
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper claims that a three-step agentic workflow can take a lightweight XML description of an ODE parameter-estimation problem plus a partially filled Python code skeleton and turn it into a ready-to-run, differentiable, just-in-time-compiled calibration pipeline. The built-in agent checks the user's code against the XML, flags or fixes inconsistent shapes, unused parameters, and other setup errors, then translates Python callables into pure functions for automatic differentiation and parallel execution. The pipeline first searches globally with particle swarm optimization and then refines locally with gradient-based optimization using reverse-mode derivatives through the ODE solve. On stiff Robertson kinetics, the recovered rate constants agree with true values to roughly 0.1–0.5 percent, and on a Van der Pol oscillator the recovered nonlinearity parameter matches to about 0.05 percent. On experimental data, the workflow produces plausible parameter sets for a piezoelectric hysteresis model and a battery thermal-runaway model.

What carries the argument

The key machinery is the agentic correction loop feeding a two-stage optimizer. The agent consumes the XML problem spec and user-filled Python skeleton, generates a validation report, separates critical errors from warnings, and iteratively rewrites the input files until no critical errors remain or the iteration budget is exhausted. The validated functions are then converted to pure, just-in-time-compilable callables with parallelization primitives; the pipeline runs particle swarm optimization—a zero-order global search—followed by a gradient-based refinement in which derivatives of the loss with respect to parameters are computed by reverse-mode automatic differentiation through the ODE i

What would settle it

Give the workflow a Robertson implementation with one deliberate semantic error—for example, swap the rate constants for the second and third reactions in the ODE—and observe whether the agent's validation report flags it before fitting. If the agent passes the corrupted model and the optimizer returns a low-loss fit to wrong parameters, the central claim about automatic code validation is not supported.

Watch

Extended reading notes

Core claim

The central claim is that the agentic AI workflow makes advanced ODE calibration accessible: a user with domain knowledge but little automatic-differentiation or optimization expertise can specify the problem in a human-readable XML template, fill in skeleton Python functions for the ODE, loss, and quantities of interest, and the system will validate the setup, auto-correct critical errors, generate differentiable just-in-time-compiled code with parallelization primitives, and run a two-stage optimization to estimate parameters. The paper demonstrates this by recovering known parameters from simulated stiff data—Robertson's kinetics and the Van der Pol oscillator—and by fitting approximate m

Load-bearing premise

The load-bearing premise is that the LLM agent, working with a bounded correction loop and no formal verification, can reliably find and fix errors in a user's ODE setup without silently replacing a valid model with a wrong one.

Editorial extensions

If this is right

  • Users with domain knowledge but no automatic-differentiation or optimizer expertise can go from an XML spec to a calibrated stiff ODE model in one run, so the bottleneck shifts from setup to modeling choices.
  • The auto-validation step could be reused as a general check for ill-posed optimization setups: shape mismatches, unused parameters, and infeasible bounds get caught before expensive solves.
  • For experimental datasets, the workflow yields baseline mechanistic models—such as a Bouc–Wen hysteresis model or an Arrhenius thermal-runaway model—that users can refine with domain-specific ranges or constraints.
  • Because the agent reports and applies corrections, the pipeline becomes a reproducible record: the final XML, corrected code, and optimization traces constitute an auditable history of the calibration.
  • The two-stage PSO-then-gradient recipe becomes a reusable template for similar inverse problems beyond the four systems demonstrated.

Reading between the lines

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

  • A systematic seeded-error benchmark across several ODE families would test the paper's auto-correction claim; the two worked examples do not establish a false-correction rate.
  • The workflow's separation of a human-readable spec from the AD implementation suggests the same orchestration could be retargeted to other automatic-differentiation backends or extended to PDE-constrained estimation, which the paper itself flags as future work.
  • For safety-relevant applications such as battery thermal-runaway models, users would likely want to review the agent's auto-corrections before trusting the fitted parameters, since no formal verification step is present.
  • The two-stage recipe could be stress-tested on higher-dimensional or weakly identifiable parameter spaces where global search cost grows.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a three-step agentic workflow for ODE parameter estimation: (I) the user supplies an XML problem specification and a Python code skeleton; (II) an LLM agent cross-checks the XML and code, reports critical errors and warnings, and auto-corrects the inputs in a bounded loop; (III) the corrected code is translated to JAX, compiled, parallelized, and calibrated by a two-stage global-then-local optimizer (PSO followed by gradient-based refinement). The workflow is demonstrated on synthetic stiff ODE problems (Robertson kinetics and Van der Pol oscillator), where known parameters are recovered accurately, and on two experimental datasets (piezoelectric hysteresis and battery thermal runaway), where qualitatively plausible fits are obtained. The main claimed contributions are reduced setup burden, AD-native and jit-compiled execution, automated validation/correction of user input, and an open-source release.

Significance. If the claims are substantiated, this is a useful contribution: it lowers the barrier for domain scientists to use differentiable programming, parallel compilation, and global-plus-local optimization for ODE calibration. The simulated-data results are the strongest part: the Robertson parameters are recovered to roughly 0.1–0.3% error and the Van der Pol μ to about 0.05% error, and the discussion correctly flags the sensitivity of reverse-mode derivatives to ODE solver tolerances. The experimental fits are plausible but are presented without uncertainty quantification or comparison to alternative methods. The distinctive novel component is the agentic auto-correction loop, but the evidence for its reliability is only anecdotal, which is the main gap between the paper's claims and what is demonstrated.

major comments (4)
  1. [§2.2, Appendix A.3, Discussion point 3] The workflow's central automation claim—that the agent reliably cross-checks and auto-corrects user code (Introduction feature 3)—is supported only by two anecdotal correction loops in Appendix A.3, both on the same problems that are later used in the results. No systematic evidence is given for the agent's detection rate, false-correction rate, failure modes, or behavior across varied user inputs. Because the correction loop is an LLM with a bounded number of iterations and no formal verification, a silent mis-correction could alter the intended ODE model and still pass the consistency re-check, yielding precisely fitted parameters for the wrong equations. The Discussion statement that the agentic workflow is 'highly effective' in correcting user errors therefore overstates the evidence. The paper should either add a systematic bug-injection evaluation (measuring detection success, fals
  2. [§3 and Appendix A.1, Eq. (14)] The numerical results cannot be fully reproduced or independently assessed because several key algorithmic choices are not reported. Specifically, the PSO constants c1, c2, w, swarm size, and number of iterations are not given; the ODE solver and tolerances used for the forward solve and for backpropagation are not specified; and the loss weights λ1 and λ2 in Eq. (14) are not reported. Since the abstract and feature 4 claim reproducibility, these settings should be listed in the text, a table, or a machine-readable configuration file.
  3. [§3.1 and §3.2] All results are based on single runs. No repeated runs with different random seeds, no error bars on the estimated parameters, and no statistical comparison against standard parameter-estimation baselines are provided. This matters because PSO is stochastic and the two-stage scheme may converge to different optima in different runs. At minimum, the paper should report repeated-run variability for the synthetic recoveries and, ideally, compare against a simpler baseline (e.g., scipy least_squares or a direct multi-start gradient method). The experimental fits are currently evaluated visually, which is not sufficient to support the claim that the workflow reliably produces 'good representations' of the data.
  4. [§3.2.1, Table 3] For the piezoelectric example, the fitted parameters β = 0 and γ ≈ 2.26e-9 make the Bouc–Wen state h essentially proportional to the voltage, so the identified model is a linear time-invariant filter rather than a hysteresis model. The 'hysteresis loop' in Fig. 7c then reflects only dynamic lag, not rate-independent hysteresis. The text acknowledges this, but as presented the example does not demonstrate hysteresis identification. The authors should either constrain β to be nonzero (and report the consequences) or relabel the demonstration as a linear-dynamics calibration benchmark.
minor comments (4)
  1. [§2.1] Typos and grammar issues: 'creats' should be 'creates', 'defintions' should be 'definitions'. Similar mechanical errors occur elsewhere and should be fixed in a copyedit pass.
  2. [Throughout] The formatting of the GitHub link appears broken in the text ('the/githubGitHub link'). Please ensure all URLs are rendered correctly.
  3. [Appendix A.1] The constants c1, c2, r1, r2, and w in Eq. (18) are not defined. Even a brief description of typical values or how they were chosen would help.
  4. [Figure 1] Figure 1 is labeled as '(a)' but appears to be a standalone figure. Please adjust the caption and labels for consistency with the other figures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: benchmark recoveries are genuine inverse problems and experimental fits use standard externally specified models.

full rationale

The paper's core contribution is a workflow, not a derived scientific law. For the Robertson and Van der Pol benchmarks, the optimizer minimizes a data-mismatch loss against simulated data generated from known parameter values; the reported parameter estimates are the outcome of that optimization and are compared with the true generative values as an independent check. The loss functions are not defined in terms of the fitted parameters, so there is no self-definitional reduction. The experimental sections fit standard Bouc-Wen and Arrhenius models to external data (Ha et al., Schöberl et al.), and the estimated parameters are outputs, not inputs. Self-citations [2], [10], [32] provide context, comparison, or standard model forms, but the equations are stated explicitly in the paper and the workflow's correctness is not justified by a self-citation chain. The only weakness is evidential: the claim in Discussion point 3 that the agent is 'highly effective in correcting user errors' is supported only by two anecdotal corrections in Appendix A.3. That is an unsupported empirical claim about reliability, not a circular derivation, and cannot be scored as circularity under the rules requiring an exhibited reduction by construction.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The free parameters are standard algorithm and loss-design choices that affect the reported fits but are not central scientific claims. The main unstated dependency is the reliability of the LLM agent in converting and correcting code, which is only anecdotally demonstrated. No new physical entities were introduced.

free parameters (5)
  • PSO hyperparameters (c1, c2, w, swarm size, iterations) = not stated in the manuscript
    The PSO update equations in Appendix A.1 depend on constants c1, c2, w and the number of particles and iterations. These control the global search behavior and are chosen by the author without a systematic sensitivity study.
  • Loss weights lambda_1 and lambda_2 for the battery thermal runaway model = not stated
    Equation 14 defines the thermal runaway loss as a weighted sum of log-scale temperature-rate error and temperature error. The weights are not given, and their choice directly affects the fitted parameters.
  • ODE solver tolerances and time-stepping settings = not stated; described as user-tunable
    Section 4 explicitly states that solver tolerances and time-stepping methods need to be tuned by the user for effective estimation, and that reverse-mode gradients are sensitive to these choices. These are influential free choices that affect the reported results.
  • Agentic loop parameters (maximum number of correction iterations) = 1 iteration in the shown examples
    Step II runs the correction loop until no critical errors or a maximum number of iterations. The examples run for 1 iteration; the chosen maximum affects the reliability of the auto-correction claim.
  • Bouc-Wen shape parameter n and mass m_p = n = 1, m_p = 0.1
    These are set to fixed values 'consistent with the used reference' rather than estimated. They are modeling choices that affect the piezoelectric fit.
assumptions (3)
  • domain assumption The LLM-based agent can correctly convert user Python skeletons into semantically equivalent JAX code and reliably detect semantic errors in ODE definitions and loss functions.
    This is the load-bearing premise of Step II and Step III. It is supported only by two anecdotal examples in Appendix A.3, not by a systematic evaluation.
  • domain assumption The two-stage PSO plus L-BFGS strategy converges to a good parameter estimate for stiff and multimodal ODE fitting problems.
    The paper states this is a commonly used approach, and the four examples support it empirically, but no convergence guarantees or failure-mode analysis are provided.
  • domain assumption Diffrax-based differentiable ODE integration with reverse-mode AD gives correct gradients for the problems considered.
    The paper cites Kidger and Rackauckas on the fragility of reverse-mode differentiation of ODE solvers, and Section 4 notes tolerances must be tightened. The correctness of the gradient stage depends on this assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems." pith.science (2026). https://pith.science/paper/LQ5746NI

@misc{pith2026250907283,
  author       = {Pith},
  title        = {Pith review of: An Agentic AI Workflow to Simplify Parameter Estimation of Complex Differential Equation Systems},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQ5746NI}},
  note         = {Machine review of arXiv:2509.07283}
}
read the original abstract

Parameter identification for mechanistic Ordinary Differential Equation (ODE) models underpins prediction and control in several applications, yet remains a manual and labor-intensive process: datasets are noisy and partial, models can be stiff and complex, and differentiable implementations demand framework expertise. An agentic AI workflow is presented that converts a lightweight, human-readable specification into a compiled, parallel, and differentiable model calibration pipeline. Users supply an XML description of the problem and fill in a Python code skeleton; the agent automatically validates consistency between problem definition and code, and auto-corrects pathologies in the input deck. It transforms Python callables into pure JAX functions for efficient just-in-time compilation and parallelization. The system then orchestrates a two-stage search comprising global exploration of the parameter space followed by gradient-based refinement. The result is an AD-native, reproducible workflow that lowers the barrier to advanced calibration while preserving expert control. An open-source implementation with a documented API and examples is released, enabling rapid movement from problem statement to interpretable ODE models with minimal effort.

Figures

Figures reproduced from arXiv: 2509.07283 by the authors.

Figure 1
Figure 1. (a) The overall workflow proposed to automate ODE parameter estimation. Large parts of the workflow, which originally would have to be done by the user and would require expertise in programming, optimization and mathematics, can be handled by agentic AI. A three-step, agentic workflow is presented that turns a lightweight, human-readable specification into a compiled, parallel, and differentiable parameter-estimati… view at source ↗
Figure 2
Figure 2. (a) Step I of the workflow: Conversion of the input XML to a Python function skeleton. This skeleton contains clearly demarcated sections to be filled in by the user. (a) [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (a) Step II takes the completed (completions highlighted) Python functions and input XML, and generates a report containing critical errors and warnings. The agent automatically corrects the input files based on issues observed, and then reassesses the inputs. This is done in a loop until all critical errors are fixed, or a maximum number of iterations is reached. 2.2 Step II: Setup Error Corrections After the user … view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: (a) The corrected system is translated into pure JAX functions that are jit-compilable, fully differentiable and parallelized using primitives like pmap. The functions are then used to solve a two-stage optimization problem to do parameter estimation. Once critical iss…
Figure 5
Figure 5. Figure 5: (a) Fit loss showing the 2-stage loss evolution (b) Fitting result showing the accuracy of the obtained fit for solution variables of all time scales Robertson’s kinetics [26] provide a classical and demanding benchmark for ODE modeling pipelines. The system is markedl…
Figure 6
Figure 6. Figure 6: (a) Fit loss showing the 2-stage loss evolution (b) Fitting result showing the accuracy of the obtained fit The Van der Pol oscillator is a canonical relaxation oscillator that appears across electronics, biomechanics, and mechanical vibrations [29, 30]. Solutions of t…
Figure 7
Figure 7. Figure 7: (a)Loss function evolution. The gradient-based optimization only minimally improves the fit (b) Predicted v/s experimental displacement (c) Expected hysteresis loop from parameters. The framework is demonstrated on a piezoelectric actuator experiment (inspired from the…
Figure 8
Figure 8. Figure 8: (a) Loss function evolution during parameter estimation (b) Model match to experimental temperature data (c) Model match to experimental temperature rate user inputs, the agent automatically creates Python routine skeletons for users to fill, translates user routines i…
Figure 9
Figure 9. Figure 9: Workflow of correction of suboptimal inputs in the Robertson parameter estimation [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Workflow of correction of suboptimal inputs in the battery thermal runaway model [PITH_FULL_IMAGE:figures/full_fig_p019_10.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. ATHENA: Agentic Team for Hierarchical Evolutionary Numerical Algorithms

    cs.LG 2025-12 unverdicted novelty 5.0 of 10

    ATHENA introduces an agentic team framework that autonomously manages the end-to-end computational research lifecycle via a knowledge-driven HENA loop to achieve validation errors of 10^{-14} in scientific computing a...

Reference graph

Works this paper leans on

39 extracted references · 31 canonical work pages · cited by 1 Pith paper

  1. [1]

    Com- parative study of equivalent circuit models performance in four common lithium-ion batteries: Lfp, nmc, lmo, nca.Batteries, 7(3):51, 2021

    Manh-Kien Tran, Andre DaCosta, Anosh Mevawalla, Satyam Panchal, and Michael Fowler. Com- parative study of equivalent circuit models performance in four common lithium-ion batteries: Lfp, nmc, lmo, nca.Batteries, 7(3):51, 2021

    work page 2021

  2. [2]

    A layered swarm optimization method for fitting battery thermal runaway models to accelerating rate calorimetry data.Journal of The Electrochemical Society, 2024

    Saakaar Bhatnagar, Andrew Comerford, Zelu Xu, Simone Reitano, Luigi Scrimieri, Luca Giuliano, and Araz Banaeizadeh. A layered swarm optimization method for fitting battery thermal runaway models to accelerating rate calorimetry data.Journal of The Electrochemical Society, 2024

    work page 2024

  3. [3]

    Kinetic modeling of gasoline surrogate components and mixtures under engine conditions.Proceedings of the Com- bustion Institute, 33(1):193–200, 2011

    Marco Mehl, William J Pitz, Charles K Westbrook, and Henry J Curran. Kinetic modeling of gasoline surrogate components and mixtures under engine conditions.Proceedings of the Com- bustion Institute, 33(1):193–200, 2011

    work page 2011

  4. [4]

    Gauss–seidel iteration for stiff odes from chemical kinetics.SIAM Journal on Scientific Computing, 15(5):1243–1250, 1994

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

    work page 1994

  5. [5]

    Stefano Giampiccolo, Federico Reali, Anna Fochesato, Giovanni Iacca, and Luca Marchetti. Ro- bust parameter estimation and identifiability analysis with hybrid neural ordinary differential equations in computational biology.NPJ Systems Biology and Applications, 10(1):139, 2024

    work page 2024

  6. [6]

    Fem-simulation of laminar flame propagation

    B Michaelis and B Rogg. Fem-simulation of laminar flame propagation. i: Two-dimensional flames. Journal of computational physics, 196(2):417–447, 2004

    work page 2004

  7. [7]

    Paul T Coman, Eric C Darcy, Christian T Veje, and Ralph E White. Modelling li-ion cell thermal runaway triggered by an internal short circuit device using an efficiency factor and arrhenius formulations.Journal of The Electrochemical Society, 164(4):A587, 2017

    work page 2017

  8. [8]

    Thermal kinetics comparison of delithiated li [nixcoymn1-xy] o2 cathodes.Journal of Power Sources, 514:230582, 2021

    Yu Wang, Dongsheng Ren, Xuning Feng, Li Wang, and Minggao Ouyang. Thermal kinetics comparison of delithiated li [nixcoymn1-xy] o2 cathodes.Journal of Power Sources, 514:230582, 2021

    work page 2021

Show all 39 references
  1. [9]

    Model-based thermal runaway prediction of lithium-ion batteries from kinetics analysis of cell components.Applied energy, 228:633–644, 2018

    Dongsheng Ren, Xiang Liu, Xuning Feng, Languang Lu, Minggao Ouyang, Jianqiu Li, and Xiang- ming He. Model-based thermal runaway prediction of lithium-ion batteries from kinetics analysis of cell components.Applied energy, 228:633–644, 2018. 12

    2018

  2. [10]

    Chemical reaction neural networks for fitting accelerating rate calorimetry data.Journal of Power Sources, 628:235834, 2025

    Saakaar Bhatnagar, Andrew Comerford, Zelu Xu, Davide Berti Polato, Araz Banaeizadeh, and Alessandro Ferraris. Chemical reaction neural networks for fitting accelerating rate calorimetry data.Journal of Power Sources, 628:235834, 2025

    2025

  3. [11]

    Identification of parameters for equivalent circuit model of li-ion battery cell with population based optimization algorithms.Ain Shams Engineering Journal, 15(3):102481, 2024

    Yu-Shan Cheng. Identification of parameters for equivalent circuit model of li-ion battery cell with population based optimization algorithms.Ain Shams Engineering Journal, 15(3):102481, 2024

    2024

  4. [12]

    Electrochemical model parameter iden- tification of a lithium-ion battery using particle swarm optimization method.Journal of Power Sources, 307:86–97, 2016

    Md Ashiqur Rahman, Sohel Anwar, and Afshin Izadian. Electrochemical model parameter iden- tification of a lithium-ion battery using particle swarm optimization method.Journal of Power Sources, 307:86–97, 2016

    2016

  5. [13]

    Hysteretic tuned mass damper with bumpers for seismic protection: Modeling, identification, and shaking table tests.Journal of Sound and Vibration, 597:118816, 2025

    Vinay Yadav Janga, Pranath Kumar Gourishetty, Biagio Carboni, Giuseppe Quaranta, and Walter Lacarbonara. Hysteretic tuned mass damper with bumpers for seismic protection: Modeling, identification, and shaking table tests.Journal of Sound and Vibration, 597:118816, 2025

    2025

  6. [14]

    Jih-Lian Ha, Ying-Shieh Kung, Rong-Fong Fung, and Shao-Chien Hsien. A comparison of fitness functions for the identification of a piezoelectric hysteretic actuator based on the real-coded genetic algorithm.Sensors and Actuators A: Physical, 132(2):643–650, 2006

    2006

  7. [15]

    A physics-informed neural network approach to parameter estimation of lithium-ion battery electro- chemical model.Journal of Power Sources, 621:235271, 2024

    Jingrong Wang, Qiao Peng, Jinhao Meng, Tianqi Liu, Jichang Peng, and Remus Teodorescu. A physics-informed neural network approach to parameter estimation of lithium-ion battery electro- chemical model.Journal of Power Sources, 621:235271, 2024

    2024

  8. [16]

    A physics-informed neural networks framework for model parameter identification of beam-like structures.Mechanical Systems and Signal Processing, 224:112189, 2025

    Rafael de O Teloli, Roberta Tittarelli, Ma¨ el Bigot, Lucas Coelho, Emmanuel Ramasso, Patrice Le Moal, and Morvan Ouisse. A physics-informed neural networks framework for model parameter identification of beam-like structures.Mechanical Systems and Signal Processing, 224:112189, 2025

    2025

  9. [17]

    The findability of microkinetic parameters by heterogeneous chemical reaction neural networks (hcrnns).Chemical Engineering Journal, 510:161460, 2025

    Hannes Stagge and Robert G¨ uttel. The findability of microkinetic parameters by heterogeneous chemical reaction neural networks (hcrnns).Chemical Engineering Journal, 510:161460, 2025

    2025

  10. [18]

    Accommodating physical reaction schemes in dsc cathode thermal stability analysis using chemical reaction neural networks.Journal of Power Sources, 581:233443, 2023

    Benjamin C Koenig, Peng Zhao, and Sili Deng. Accommodating physical reaction schemes in dsc cathode thermal stability analysis using chemical reaction neural networks.Journal of Power Sources, 581:233443, 2023

    2023

  11. [19]

    On neural differential equations.arXiv preprint arXiv:2202.02435, 2022

    Patrick Kidger. On neural differential equations.arXiv preprint arXiv:2202.02435, 2022

    2022 arXiv

  12. [20]

    Chatgpt for programming numerical methods.Journal of Machine Learning for Modeling and Computing, 4(2), 2023

    Ali Kashefi and Tapan Mukerji. Chatgpt for programming numerical methods.Journal of Machine Learning for Modeling and Computing, 4(2), 2023

    2023

  13. [21]

    Fine-tuning a large language model for automating com- putational fluid dynamics simulations.Theoretical and Applied Mechanics Letters, page 100594, 2025

    Zhehao Dong, Zhen Lu, and Yue Yang. Fine-tuning a large language model for automating com- putational fluid dynamics simulations.Theoretical and Applied Mechanics Letters, page 100594, 2025

    2025

  14. [22]

    Openfoamgpt: A retrieval-augmented large language model (llm) agent for openfoam-based computational fluid dynamics.Physics of Fluids, 37(3), 2025

    Sandeep Pandey, Ran Xu, Wenkang Wang, and Xu Chu. Openfoamgpt: A retrieval-augmented large language model (llm) agent for openfoam-based computational fluid dynamics.Physics of Fluids, 37(3), 2025

    2025

  15. [23]

    Ai agents in engineering design: a multi-agent framework for aesthetic and aerodynamic car design.arXiv preprint arXiv:2503.23315, 2025

    Mohamed Elrefaie, Janet Qian, Raina Wu, Qian Chen, Angela Dai, and Faez Ahmed. Ai agents in engineering design: a multi-agent framework for aesthetic and aerodynamic car design.arXiv preprint arXiv:2503.23315, 2025

    2025

  16. [24]

    From concept to manufacturing: Evaluating vision-language models for engineering design.Artificial Intelligence Review, 58(9):288, 2025

    Cyril Picard, Kristen M Edwards, Anna C Doris, Brandon Man, Giorgio Giannone, Md Ferdous Alam, and Faez Ahmed. From concept to manufacturing: Evaluating vision-language models for engineering design.Artificial Intelligence Review, 58(9):288, 2025

    2025

  17. [25]

    How an ai-enabled software product development life cycle will fuel innovation.https://www

    Chandra Gnanasambandam, Martin Harrysson, Rikki Singh, and Aditi Chawla. How an ai-enabled software product development life cycle will fuel innovation.https://www. mckinsey.com/industries/technology-media-and-telecommunications/our-insights/ how-an-ai-enabled-software-product...

    2025

  18. [26]

    Robertson’s example for stiff differential equations.Arizona State Univer- sity, Technical report, 1996

    Matthias K Gobbert. Robertson’s example for stiff differential equations.Arizona State Univer- sity, Technical report, 1996

    1996

  19. [27]

    Stiff-pinn: Physics-informed neural network for stiff chemical kinetics.The Journal of Physical Chemistry A, 125(36):8098–8106, 2021

    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

    2021

  20. [28]

    Accelerating simulation of stiff nonlinear systems using continuous-time echo state networks.arXiv preprint arXiv:2010.04004, 2020

    Ranjan Anantharaman, Yingbo Ma, Shashi Gowda, Chris Laughman, Viral Shah, Alan Edelman, and Chris Rackauckas. Accelerating simulation of stiff nonlinear systems using continuous-time echo state networks.arXiv preprint arXiv:2010.04004, 2020

    2010 arXiv

  21. [29]

    Analysis of an experimental technique for determining van der pol parameters of a transistor oscillator.IEEE transactions on microwave theory and techniques, 46(7):914–922, 1998

    Kuang Yi Chen, Paul D Biernacki, A Lahrichi, and Alan Mickelson. Analysis of an experimental technique for determining van der pol parameters of a transistor oscillator.IEEE transactions on microwave theory and techniques, 46(7):914–922, 1998

    1998

  22. [30]

    Yanli Wang, Xianghong Li, and Yongjun Shen. Study on mechanical vibration control of limit cycle oscillations in the van der pol oscillator by means of nonlinear energy sink.Journal of Vibration Engineering & Technologies, 12(1):811–819, 2024

    2024

  23. [31]

    Thermal runaway characterization of cylindrical lithium-ion and sodium-ion batteries with various sizes and energy contents.Journal of Power Sources, 648:237240, 2025

    Jan Sch¨ oberl, Sebastian Ohneseit, Stefan Schaeffler, Dominic F¨ orstermann, Linus Grahl, Andreas Jossen, Carlos Ziebert, and Markus Lienkamp. Thermal runaway characterization of cylindrical lithium-ion and sodium-ion batteries with various sizes and energy contents.Journal o...

    2025

  24. [32]

    Experimental and simulation-based characterization of thermal runaway in lithium-ion batteries using altair simlab®

    Luca Giuliano, Luigi Scrimieri, Simone Reitano, Davide Berti Polato, Alessandro Ferraris, An- drew Comerford, and Saakaar Bhatnagar. Experimental and simulation-based characterization of thermal runaway in lithium-ion batteries using altair simlab®. Technical report, SAE Techn...

    2025

  25. [33]

    Chris Rackauckas. The numerical analysis of differentiable simulation: Auto- matic differentiation can be incorrect.https://www.stochasticlifestyle.com/ the-numerical-analysis-of-differentiable-simulation-automatic-differentiation-can-be-incorrect/, April 2025. Stochastic Life...

    2025

  26. [34]

    Maziar Raissi, Paris Perdikaris, and George E Karniadakis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.Journal of Computational physics, 378:686–707, 2019

    2019

  27. [35]

    Φ flow (PhiFlow): Differentiable simulations for pytorch, tensorflow and jax

    Philipp Holl and Nils Thuerey. Φ flow (PhiFlow): Differentiable simulations for pytorch, tensorflow and jax. InInternational Conference on Machine Learning. PMLR, 2024

    2024

  28. [36]

    Particle swarm optimization

    James Kennedy and Russell Eberhart. Particle swarm optimization. InProceedings of ICNN’95- international conference on neural networks, volume 4, pages 1942–1948. ieee, 1995

    1942

  29. [37]

    Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations.Advances in neural information processing systems, 31, 2018

    2018

  30. [38]

    Autonomous kinetic modeling of biomass pyrolysis using chemical reaction neural networks.Combustion and Flame, 240:111992, 2022

    Weiqi Ji, Franz Richter, Michael J Gollner, and Sili Deng. Autonomous kinetic modeling of biomass pyrolysis using chemical reaction neural networks.Combustion and Flame, 240:111992, 2022

    2022

  31. [39]

    Jax md: a framework for differentiable physics

    Samuel Schoenholz and Ekin Dogus Cubuk. Jax md: a framework for differentiable physics. Advances in Neural Information Processing Systems, 33:11428–11441, 2020. A Appendix A.1 Particle Swarm Optimization (PSO) Particle Swarm Optimization [36] is a zero-order optimization metho...

    2020

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.