A New Way to Work with PFLOTRAN — Free AI-Powered Simulation Workflows

[keni...@gmail.com]

Dear PFLOTRAN Community,

I would like to invite you to try a new PFLOTRAN service:

https://getlectra.com

The goal of this project is to explore a fundamentally different way of performing numerical simulations. Instead of requiring users to master every detail of a complex modeling process, users can describe their simulation task, objectives, constraints, and available information. The TMG Agent then helps translate those requirements into a working PFLOTRAN model.

The website currently provides:

• A fully automated simulation workflow that can prepare or review input files, run PFLOTRAN, diagnose problems, revise the model, repeat simulations, post-process results, and prepare a technical report.

• Simulation Diagnostics for identifying input errors, convergence problems, inappropriate parameters, numerical failures, and potentially unreasonable results.

• Expert-level PFLOTRAN Q/A, designed to provide detailed technical answers with equations, code, tables, and references when available.

• PFLOTRAN source modification and compilation support for users working with customized source code.

The service is  free to use.

This project was created as a personal, after-hours effort and currently has no external funding. Therefore, only very limited computing resources are available. Model size, uploaded files, simulation runtime, and the number of concurrent tasks must be restricted. These limits reflect available resources rather than the intended capabilities of the workflow.

This is still an evolving system, and community feedback is extremely valuable. If you find an incorrect answer, a failed workflow, a confusing interface, or any other problem, please let me know. Every problem discovered is an opportunity to improve the TMG Agent and make it more capable for future PFLOTRAN users.

I would greatly appreciate your comments, test cases, and suggestions.

Best regards,

Kenny Zhang
https://getlectra.com

[Mills, Richard Tran]

Hi Kenny,

The TMG Agent that The TOUGH group has developed looks interesting. There is certainly a big space of possibilities for facilitating simulation setup that has recently opened up with powerful agentic LLM-based tools, in general. I've been amazed at what I've been able to pull together by just pointing Claude Code at things like the PETSc and PFLOTRAN repos and describing simulations I want it to build and run, so it's interesting to see what can be done as people start to make custom-tailored agents for such work.

I realize that you have mostly just copied over the overall structure for your TOUGH agent to build this PFLOTRAN one, so I understand why the info may not be there yet, but I suggest that you say a little bit on your PFLOTRAN agent landing page about what PFLOTRAN is, who has contributed to its development, and how its licensed, where to get it, etc. I realize that PFLOTRAN's LGPL license doesn't necessarily require this, but whenever I make serious use of an open-source code, I consider it good form to advertise this. (I did ask your Q&A agent to provide such details, and after a web search or pulling for its RAG data base or whatever it does, it did give me a summary that appeared to be correct and without hallucinations, so that's good, at least.)

I'd also like to know what LLM(s) you are using to drive this, if you can disclose this information. In my case, it's just curiosity, but for other people it's important to know. At many of the DOE Labs staff keep getting stern warnings to "MAKE SURE TO NEVER USE CHINESE LLMs" (even though I think that a bunch of these warnings don't even make sense, in the case of locally run open-weight models); there are some legit reasons to know what LLMs are being employed and where the inference is being run. At Argonne we've been banned from using many fairly innocuous LLM tools due to concerns about where the data ends up, or even just due to "telemetry" that gets sent.

Cheers,

Richard

---

From: pflotra...@googlegroups.com <pflotra...@googlegroups.com> on behalf of keni...@gmail.com <keni...@gmail.com>
Sent: Sunday, August 30, 2026 5:53 PM
To: pflotran-users <pflotra...@googlegroups.com>
Subject: [pflotran-users: 8840] A New Way to Work with PFLOTRAN — Free AI-Powered Simulation Workflows

 

You don't often get email from keni...@gmail.com. Learn why this is important

Dear PFLOTRAN Community, I would like to invite you to try a new PFLOTRAN service: https: //getlectra. com The goal of this project is to explore a fundamentally different way of performing numerical simulations. Instead of requiring users to

ZjQcmQRYFpfptBannerStart

This Message Is From an External Sender

This message came from outside your organization.

 

ZjQcmQRYFpfptBannerEnd

--
You received this message because you are subscribed to the Google Groups "pflotran-users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pflotran-user...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/pflotran-users/5a225cbc-bcc9-4910-93f0-75fdde4c3a78n%40googlegroups.com.

[Kenny]

Subject: Re: TMG Agent and PFLOTRAN

Hi Richard,

Thank you for your thoughtful comments and suggestions.

First, I should clarify that, due to funding constraints, I am no longer working at LBL. I developed TMG Agent after leaving LBL, so the project is entirely independent of LBL and was developed 100% by me personally.

TMG Agent is a general-purpose simulation-agent framework rather than an agent developed specifically for one simulator. I have currently applied it to TOUGH3, PFLOTRAN, and my own simulator. The framework is designed to support different large language models that can be connected through the Codex interface. I have successfully tested it with ChatGPT, Claude, and several Chinese LLMs. The version currently running on the website uses ChatGPT 5.6 SOL.

For the PFLOTRAN implementation, I instructed the agent to study the PFLOTRAN source code, manuals, mailing-list archives, and relevant materials from other repositories. These materials have been incorporated into its knowledge base. It should therefore already have a fairly comprehensive understanding of PFLOTRAN, although I will continue testing and improving its technical accuracy.

Thank you also for your suggestion regarding the landing page. I agree that it should provide a clear introduction to PFLOTRAN, acknowledge its developers and contributors, explain its open-source license, and provide links to the official repository and other relevant resources. I will add this information to the landing page.

The website already includes a Privacy Notice and Terms of Use describing how account information, uploaded files, project data, public-document retrieval, account and project isolation, data retention and deletion, access controls, and security limitations are handled. In particular, the public retrieval layer sends only a limited set of general PFLOTRAN technical search terms and does not send the user’s full question, uploaded files, project description, email address, or account details to public search services.

I agree, however, that disclosure of the particular LLM in use and the location and handling of model inference should be made more explicit and easier to find. I will add clearer information about the active model and inference arrangement to the website and add an option for users to use their own LLM API. 

Thank you again for taking the time to review the site and provide such useful feedback.

Best regards,

Kenny

To view this discussion visit https://groups.google.com/d/msgid/pflotran-users/CO6PR09MB7352DDFD4312D90723074FDFB8A92%40CO6PR09MB7352.namprd09.prod.outlook.com.