Release v8.0 Preview - Atmospheric model

[sylvain-guillet]

What's Changed

• Feature/nrlmsis by @sylvain-guillet in Feature/nrlmsis #255
• Feature/upgrade nrlmsise by @sylvain-guillet in Feature/upgrade nrlmsise #257
• Feature/nrlmsise optimisation by @sylvain-guillet in Feature/nrlmsise optimisation #258

Full Changelog: v7.1.0...v8.0.0-preview

---

This discussion was created from the release Release v8.0 Preview - Atmospheric model.

[ChristopherRabotin]

Hi Sylvain,

I've been following your work for a few months. Adding the NRLMSIS model is a impressive, and a lot of work. It seems like you used Claude as support. I don't see any Fortran code. Out of curiosity, did you feed it the "official" FORTRAN code and ask for a rewrite in C#? Or did you have a C# baseline already?

Thank you

[sylvain-guillet]

Hi Chris,

Thank you for your kind words.

To answer your question directly: no, I did not ask an LLM to rewrite the official FORTRAN NRLMSIS code into C#. I deliberately avoided a direct FORTRAN to C# rewrite because the risk of subtle numerical regressions is high, especially due to differences in floating-point behavior, array semantics, and implicit assumptions in the original code.

Instead, I relied on validated C implementation as the computational reference and integrated it into the .NET ecosystem while preserving the original algorithms and numerical behavior. To validate correctness, I reused the official test datasets and reference outputs from the FORTRAN version to systematically verify the results produced on the C# side. This was essential to ensure numerical equivalence and avoid silent scientific drift. AI tools were used only as auxiliary support (documentation, data model, sanity checks,...), not for source-to-source translation.

On a side note, your work on nyxspace in Rust is genuinely impressive. I’m particularly curious about the performance you’re reaching: do you have published benchmarks or comparative figures (e.g. versus SPICE or other toolchains)?

Best regards,
Sylvain

[ChristopherRabotin]

Hi Sylvain,

Great work: that sounds like the correct approach. Is that C implementation public by any chance? All of the repositories I've found on Github expose bindings to the original code. Nyx itself does not support a high fidelity drag model. I tried to have an LLM convert the original FORTRAN code for NRLMSIS several months ago, and while I have yet to check the details, I would be very surprised if it worked correctly, but I have yet to check in detail.

Thanks for your kind words on my Nyx Space work. It's consumed a lot of my life outside of work for many years now. The benchmarking of ANISE versus SPICE shows that ANISE is about ten times faster to query the planetary kernels than SPICE (mean of 176.17 µs for ANISE and 1.5219 ms for SPICE) , but almost two times slower to query the spacecraft kernels with the Hermite interpolation (mean of 38.261 ms for ANISE and 17.779 ms for SPICE). I am not certain, but I think the run-time frame checks in ANISE add that overhead. We use ANISE at Rocket Lab on several projects, and it's also used at Firefly Space, K2 Space, and Astranis. At Rocket Lab, we've found that the analysis engine in ANISE is about 500x faster than the one in STK, while generating identical results: we recently rewrote a service that would take 5-6 minutes to build the data we needed in ANISE, and it returns in less than 0.5 seconds. While Nyx does not have a good drag model, we're considering using it for the orbit determination of Blue Ghost Mission 2 because we don't need to account for drag given our launch profile.

I don't know anything about C#, if you're interested in using it in your code, I can try to learn how to build the bindings. I tried to learn how to use bazel to build a C++ interface to ANISE, but I royally failed, so I can't guarantee anything.

Best regards,
Chris

[sylvain-guillet]

Hi Chris, Thank you for your kind words. To answer your question directly: no, I did not ask an LLM to rewrite the official FORTRAN NRLMSIS code into C#. I deliberately avoided a direct FORTRAN  to C# rewrite because the risk of subtle numerical regressions is high, especially due to differences in floating-point behavior, array semantics, and implicit assumptions in the original code. Instead, I relied on validated  C implementation as the computational reference and integrated it into the .NET ecosystem while preserving the original algorithms and numerical behavior. To validate correctness, I reused the official test datasets and reference outputs from the FORTRAN version to systematically verify the results produced on the C# side. This was essential to ensure numerical equivalence and avoid silent scientific drift. AI tools were used only as auxiliary support (documentation, data model, sanity checks,...), not for source-to-source translation. On a side note, your work on nyxspace in Rust is genuinely impressive. I’m particularly curious about the performance you’re reaching: do you have published benchmarks or comparative figures (e.g. versus SPICE or other toolchains)? Best regards, Sylvain

…

-- Secured with Tuta Mail: https://tuta.com/free-email Dec 30, 2025, 19:25 by ***@***.***:

Hi Sylvain, I've been following your work for a few months. Adding the NRLMSIS model is a impressive, and a lot of work. It seems like you used Claude as support. I don't see any Fortran code. Out of curiosity, did you feed it the "official" FORTRAN code and ask for a rewrite in C#? Or did you have a C# baseline already? Thank you — Reply to this email directly, > view it on GitHub <#259 (comment)>> , or > unsubscribe <https://github.com/notifications/unsubscribe-auth/AD6PDQ5QOY5UYQFDEK62KLL4EK7SBAVCNFSM6AAAAACQKG6QPWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMZXGY4DQNI>> . You are receiving this because you authored the thread.> Message ID: > <IO-Aerospace-software-engineering/Astrodynamics/repo-discussions/259/comments/15376885> @> github> .> com>