CFDTwin: An open-source GUI and Python toolkit for POD-NN surrogate modeling of ANSYS Fluent simulations

Daniel Curl; Han Hu

arxiv: 2605.27725 · v1 · pith:B46WE7DJ · submitted 2026-05-26 · physics.flu-dyn · cs.LG

CFDTwin: An open-source GUI and Python toolkit for POD-NN surrogate modeling of ANSYS Fluent simulations

Daniel Curl , Han Hu This is my paper

Reviewed by Pith2026-06-29 15:06 UTCgrok-4.3pith:B46WE7DJ open to challenge →

classification physics.flu-dyn cs.LG

keywords CFD surrogate modelingPOD-NNANSYS Fluentopen-source toolkitdigital twinproper orthogonal decompositionworkflow automationreduced-order modeling

0 comments

The pith

CFDTwin packages the steps for building POD-NN surrogate models into a reusable open-source Python toolkit and GUI for ANSYS Fluent CFD simulations.

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

The paper introduces CFDTwin as a software platform that automates the full workflow of generating design samples, running Fluent batch simulations, extracting data, training proper orthogonal decomposition-neural network surrogates, validating them, and making predictions at new input points. This targets the repeated high cost of full CFD solves in design optimization and digital-twin applications by turning one-time simulation campaigns into fast reusable models for scalar values, surface fields, and cell-zone outputs. The toolkit exposes the same capabilities through both a scriptable API and a desktop interface, allowing users to resume interrupted runs and inspect accuracy metrics without writing custom code for each new case. By extending a prior electronics-cooling demonstration into a general platform, the work claims to make such surrogate techniques practical for a wider range of thermal-fluid problems.

Core claim

CFDTwin packages the steps of parameter sampling, Fluent automation, data extraction, reduced-order model construction, neural-network training, validation, and prediction into a reusable workflow that supports scalar, surface-field, and cell-zone outputs through both a scriptable API and a desktop GUI, extending the prior POD-NN approach from a case-specific implementation to a general research-software platform.

What carries the argument

CFDTwin, the open-source Python package and optional GUI that integrates design-of-experiments sampling, Fluent batch execution, POD-NN surrogate training for multiple output types, and model evaluation at new design points.

If this is right

Users can define inputs and outputs once, generate samples, and run or resume batch simulations without writing per-project scripts.
Trained surrogates can be evaluated at new design points for scalar, surface, and volume outputs without re-running full Fluent solves.
Validation metrics become directly inspectable inside the same interface used for training and prediction.
The same workflow supports both scripted reproducibility and interactive model checks through the GUI.
The platform turns repeated CFD campaigns into reusable models suitable for optimization loops and digital-twin use.

Where Pith is reading between the lines

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

If the workflow proves robust, it could shorten the time from problem setup to usable surrogate from weeks of custom coding to hours of configuration.
The separation of the Python API from the GUI suggests the same backend could later support web or cloud deployment for collaborative teams.
Because the toolkit preserves modal structure from the POD step, downstream users may still interpret which flow features drive the predictions even after the neural network is trained.
Extending the same structure to unsteady or multi-physics Fluent cases would be a direct next test of the packaging claim.

Load-bearing premise

That the POD-NN surrogate method shown on one electronics-cooling case can be turned into a general workflow that works across other Fluent simulations while keeping accuracy and physical interpretability without extra case-by-case tuning.

What would settle it

Running the CFDTwin workflow on a new Fluent case and finding that the resulting surrogate either requires substantial manual adjustments to match the accuracy of the original cold-plate example or produces errors large enough to make it unusable for design decisions.

Figures

Figures reproduced from arXiv: 2605.27725 by Daniel Curl, Han Hu.

**Figure 2.** Figure 2: Snapshot of the graphical user interface of CFDTwin showing the Analyze page. [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Design exploration using CFDTwin with geometric generation and optimization exten [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

read the original abstract

High-fidelity computational fluid dynamics (CFD) is widely used for thermal-fluid design, but repeated CFD solves remain expensive for design optimization, uncertainty analysis, and digital-twin workflows. Recently, our team has demonstrated that a proper orthogonal decomposition and neural-network (POD-NN) surrogate can predict two-dimensional thermal fields in an electronics-cooling cold plate with large inference speedups while preserving physically interpretable modal structure. Reproducing and extending such workflows, however, typically requires custom scripts for parameter sampling, Fluent automation, data extraction, reduced-order model construction, neural-network training, validation, and prediction. This paper introduces CFDTwin, an open-source Python package and optional desktop graphical user interface (GUI) that packages these steps into a reusable workflow for ANSYS Fluent simulations. CFDTwin allows users to define simulation inputs and output quantities, generate design-of-experiments samples, run and resume Fluent batch simulations, train POD-NN surrogate models for scalar, surface-field, and cell-zone outputs, inspect validation metrics, and evaluate trained models at new design points without re-running Fluent. The same workflow is exposed through a scriptable Python API and a GUI, supporting reproducible studies, user-facing model validation, and automated design exploration. CFDTwin extends the prior POD-NN modeling study from a case-specific research implementation to a reusable research-software platform for CFD surrogate modeling and digital-twin development.

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.

Desk Editor's Note private letter to a colleague

CFDTwin is a practical packaging of an existing POD-NN workflow into a documented toolkit with GUI, but its transfer to new Fluent cases stays untested.

read the letter

The main takeaway is that this paper releases CFDTwin as an open-source Python package plus optional GUI that bundles sampling, Fluent batch runs, POD-NN training, and prediction for scalar, surface, and cell-zone outputs. It turns the authors' earlier cold-plate script into something reusable rather than inventing new algorithms.

The work does a clear job laying out the end-to-end steps: DOE generation, resumable simulations, model training with validation metrics, and inference at new points. Exposing the same functions through both API and GUI is useful for people who want scripts for reproducibility but also a point-and-click option. The abstract is straightforward about the scope being limited to ANSYS Fluent.

The soft spot is exactly the one flagged in the stress test. The package claims to preserve accuracy and interpretability across cases, yet the only demonstrated example is the original electronics-cooling plate. No additional test problems, no checks on how POD rank or NN settings behave under different turbulence models or boundary conditions, and no ablation on defaults. That leaves the generality claim resting on the assumption that one workflow will transfer without extra tuning.

This is aimed at Fluent users already doing design optimization or digital-twin work who would rather not rebuild the surrogate pipeline each time. A reader who needs a ready-made tool for that specific stack could get value from the code if the documentation and examples hold up.

I would send it to peer review. The contribution is modest but the software angle is legitimate, and referees can check the actual implementation, test coverage, and whether the validation section is limited to the single prior case.

Referee Report

2 major / 2 minor

Summary. The manuscript presents CFDTwin, an open-source Python package and optional GUI that packages the workflow for POD-NN surrogate modeling of ANSYS Fluent CFD simulations. Users can define inputs and outputs, generate DoE samples, run and resume batch simulations, train POD-NN models for scalar, surface-field, and cell-zone outputs, inspect validation metrics, and evaluate predictions at new design points. It extends a prior case-specific POD-NN demonstration on an electronics-cooling cold plate to a reusable platform for surrogate modeling and digital-twin development.

Significance. If the implementation is robust, CFDTwin would provide a practical, reusable tool that lowers the barrier to applying POD-NN surrogates in Fluent-based design studies, supporting reproducibility through its dual API/GUI interface and open-source release. The explicit support for multiple output types and simulation resumption are engineering strengths that could accelerate uncertainty quantification and optimization workflows.

major comments (2)

[Abstract] Abstract: The claim that CFDTwin extends the prior POD-NN study 'to a reusable research-software platform for CFD surrogate modeling' across ANSYS Fluent simulations is load-bearing but unsupported; no multi-case validation results, ablation studies on POD rank or NN hyperparameters, or tests on problems with differing boundary conditions, turbulence models, or output topologies are presented to verify transfer without case-specific tuning.
[§4] §4 (Implementation and Workflow): The description of data extraction and POD-NN training routines does not include quantitative evidence or default settings that maintain physical consistency and accuracy on new cases, which directly undermines the reusability assertion central to the contribution.

minor comments (2)

[Figure 2] Figure 2 (GUI screenshot): Increase resolution or add annotations to clarify the workflow steps for readers unfamiliar with the interface.
Ensure the GitHub repository link includes a permanent archive (e.g., Zenodo DOI) and explicit licensing information for the code and any bundled Fluent journal files.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the detailed review and constructive comments. We address each major comment below, clarifying the scope of the contribution as a software platform and indicating revisions where the manuscript will be updated.

read point-by-point responses

Referee: [Abstract] Abstract: The claim that CFDTwin extends the prior POD-NN study 'to a reusable research-software platform for CFD surrogate modeling' across ANSYS Fluent simulations is load-bearing but unsupported; no multi-case validation results, ablation studies on POD rank or NN hyperparameters, or tests on problems with differing boundary conditions, turbulence models, or output topologies are presented to verify transfer without case-specific tuning.

Authors: We agree that the manuscript does not provide new multi-case validation or ablation studies, as its focus is on the development and release of the CFDTwin software toolkit rather than on presenting additional empirical results. The reusability is achieved through the modular Python API and GUI that encapsulate the workflow from the prior study, allowing users to apply it to their own cases. To address this, we will revise the abstract to more precisely state that CFDTwin provides a reusable platform implementing the POD-NN approach for Fluent simulations, without claiming cross-case transfer validation in this work. revision: yes
Referee: [§4] §4 (Implementation and Workflow): The description of data extraction and POD-NN training routines does not include quantitative evidence or default settings that maintain physical consistency and accuracy on new cases, which directly undermines the reusability assertion central to the contribution.

Authors: Section 4 outlines the implementation details of the data extraction and training components. While the paper does not include new quantitative benchmarks for arbitrary cases, we will enhance this section by adding recommended default settings for key parameters such as POD energy threshold and neural network architecture, drawn from the prior validation study. Additionally, we will include more explicit discussion of how the built-in validation metrics help ensure physical consistency on user-specific problems. This will strengthen the guidance for reusability without requiring new simulations. revision: partial

Circularity Check

0 steps flagged

No circularity: contribution is software packaging with background citation only

full rationale

The manuscript introduces a Python/GUI toolkit that packages existing POD-NN steps (sampling, Fluent automation, model training, validation) into a reusable workflow. The single self-citation to the authors' prior cold-plate demonstration supplies only historical context and is not invoked to justify any derivation, uniqueness claim, or prediction inside this paper. No equations, fitted parameters renamed as predictions, or self-referential definitions appear; the central claim (reusable platform) is independent of the cited result and rests on standard software-engineering practices rather than any tautological reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

This is a software-description paper. No mathematical derivations, physical models, or fitted constants are introduced, so the ledger contains no entries.

pith-pipeline@v0.9.1-grok · 5787 in / 1180 out tokens · 48164 ms · 2026-06-29T15:06:11.988049+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

13 extracted references

[1]

Fei Tao, He Zhang, Ang Liu, and Andrew Y. C. Nee. Digital twin in industry: State-of-the-art. IEEE Transactions on Industrial Informatics, 15(4):2405–2415, 2019

2019
[2]

Stavroulakis, Bessy G

Georgios E. Stavroulakis, Bessy G. Charalambidi, and Petros Koutsianitis. Review of com- putational mechanics, optimization, and machine learning tools for digital twins applied to infrastructures.Applied Sciences, 12(23):11997, 2022

2022
[3]

Real-time optimization of microchannel cooling using a deep reinforcement learning-enabled digital twin

Tong Xue, Guoqing Yao, Haijun Xing, Yan Yu, Jun Zhao, and Dongsheng Wen. Real-time optimization of microchannel cooling using a deep reinforcement learning-enabled digital twin. Applied Thermal Engineering, 291:130164, 2026

2026
[4]

Darshil Panchigar, Kaushik Kar, Shashank Shukla, R. M. Mathew, Uday Chadha, and S. K. Selvaraj. Machine learning-based cfd simulations: a review, models, open threats, and future tactics.Neural Computing and Applications, 34(24):21677–21700, 2022

2022
[5]

P. J. Holmes, J. L. Lumley, G. Berkooz, J. C. Mattingly, and R. W. Wittenberg. Low- dimensional models of coherent structures in turbulence.Physics Reports, 287(4):337–384, 1997

1997
[6]

Hesthaven and Stefano Ubbiali

Jan S. Hesthaven and Stefano Ubbiali. Non-intrusive reduced order modeling of nonlinear problems using neural networks.Journal of Computational Physics, 363:55–78, 2018

2018
[7]

Michael MacRaild, Ali Sarrami-Foroushani, Toni Lassila, and Alejandro F. Frangi. Reduced order modelling of intracranial aneurysm flow using proper orthogonal decomposition and neural networks.International Journal for Numerical Methods in Biomedical Engineering, 40(10):e3848, 2024

2024
[8]

Ganti and P

H. Ganti and P. Khare. Data-driven surrogate modeling of multiphase flows using machine learning techniques.Computers & Fluids, 211:104626, 2020

2020
[9]

Daniel Curl and Han Hu. Physically interpretable surrogate modeling of thermal fields in electronics cooling using combined proper orthogonal decomposition and neural networks.AI Thermal Fluids, 6:100040, 2026

2026
[10]

CFDTwin: Digital twin toolkit for cfd simulations using ansys fluent

Daniel Curl and Han Hu. CFDTwin: Digital twin toolkit for cfd simulations using ansys fluent. GitHub repository, 2026

2026
[11]

CFDTwin v0.2.0, 2026

Daniel Curl and Han Hu. CFDTwin v0.2.0, 2026

2026
[12]

CFDTwin documentation.https://uark-ned3.github.io/ CFDTwin/, 2026

Daniel Curl and Han Hu. CFDTwin documentation.https://uark-ned3.github.io/ CFDTwin/, 2026. 10

2026
[13]

M. D. McKay, R. J. Beckman, and W. J. Conover. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code.Technometrics, 21(2):239–245, 1979. 11

1979

[1] [1]

Fei Tao, He Zhang, Ang Liu, and Andrew Y. C. Nee. Digital twin in industry: State-of-the-art. IEEE Transactions on Industrial Informatics, 15(4):2405–2415, 2019

2019

[2] [2]

Stavroulakis, Bessy G

Georgios E. Stavroulakis, Bessy G. Charalambidi, and Petros Koutsianitis. Review of com- putational mechanics, optimization, and machine learning tools for digital twins applied to infrastructures.Applied Sciences, 12(23):11997, 2022

2022

[3] [3]

Real-time optimization of microchannel cooling using a deep reinforcement learning-enabled digital twin

Tong Xue, Guoqing Yao, Haijun Xing, Yan Yu, Jun Zhao, and Dongsheng Wen. Real-time optimization of microchannel cooling using a deep reinforcement learning-enabled digital twin. Applied Thermal Engineering, 291:130164, 2026

2026

[4] [4]

Darshil Panchigar, Kaushik Kar, Shashank Shukla, R. M. Mathew, Uday Chadha, and S. K. Selvaraj. Machine learning-based cfd simulations: a review, models, open threats, and future tactics.Neural Computing and Applications, 34(24):21677–21700, 2022

2022

[5] [5]

P. J. Holmes, J. L. Lumley, G. Berkooz, J. C. Mattingly, and R. W. Wittenberg. Low- dimensional models of coherent structures in turbulence.Physics Reports, 287(4):337–384, 1997

1997

[6] [6]

Hesthaven and Stefano Ubbiali

Jan S. Hesthaven and Stefano Ubbiali. Non-intrusive reduced order modeling of nonlinear problems using neural networks.Journal of Computational Physics, 363:55–78, 2018

2018

[7] [7]

Michael MacRaild, Ali Sarrami-Foroushani, Toni Lassila, and Alejandro F. Frangi. Reduced order modelling of intracranial aneurysm flow using proper orthogonal decomposition and neural networks.International Journal for Numerical Methods in Biomedical Engineering, 40(10):e3848, 2024

2024

[8] [8]

Ganti and P

H. Ganti and P. Khare. Data-driven surrogate modeling of multiphase flows using machine learning techniques.Computers & Fluids, 211:104626, 2020

2020

[9] [9]

Daniel Curl and Han Hu. Physically interpretable surrogate modeling of thermal fields in electronics cooling using combined proper orthogonal decomposition and neural networks.AI Thermal Fluids, 6:100040, 2026

2026

[10] [10]

CFDTwin: Digital twin toolkit for cfd simulations using ansys fluent

Daniel Curl and Han Hu. CFDTwin: Digital twin toolkit for cfd simulations using ansys fluent. GitHub repository, 2026

2026

[11] [11]

CFDTwin v0.2.0, 2026

Daniel Curl and Han Hu. CFDTwin v0.2.0, 2026

2026

[12] [12]

CFDTwin documentation.https://uark-ned3.github.io/ CFDTwin/, 2026

Daniel Curl and Han Hu. CFDTwin documentation.https://uark-ned3.github.io/ CFDTwin/, 2026. 10

2026

[13] [13]

M. D. McKay, R. J. Beckman, and W. J. Conover. A comparison of three methods for selecting values of input variables in the analysis of output from a computer code.Technometrics, 21(2):239–245, 1979. 11

1979