Artificial Intelligence in HF Propagation: Feasible, Innovative, or Premature?

[Ricardo]

Dear Colleagues,

I would like to share with this esteemed scientific community a reflection that can be viewed as an initial proposal and an academic provocation.

Could an intelligent system be developed to analyze and predict ionospheric propagation in the HF (160 meters) band? This system would integrate AI, physical models, and real empirical data.

This experimental computer system would combine:

- Real-time ionospheric and solar data (simulated or obtained via APIs from NOAA, NASA, INPE, ESA, and UCAR);

- Modeling based on propagation physics (absorption, MUF, foF2, TEC, SNR, etc.);

- Artificial intelligence (AI) inference for analysis, explanation, and prediction based on geophysical and temporal conditions.

- Integration with the ARRL Logbook of the World (LoTW) database to cross-reference actual contact records (QSOs) with solar and ionospheric indicators.

It has an interactive visual interface with graphics, hourly forecasts, and data export.

Challenges to the community:

Can we rely on AI as a scientific assistant in propagation models?

What role does AI play in strengthening HF analysis without suppressing classical methods?

Can empirical data, such as QSOs confirmed in LoTW, validate and refine large-scale propagation forecasts?

Is it possible to predict extreme events, such as propagation blackouts or rare DX windows, with useful advance notice for operations or teaching?

We believe this debate could foster interdisciplinary collaborations between radio amateurs, engineers, space physicists, computer scientists, and science educators.

If you are interested in evaluating, validating, critiquing, or collaborating on this initiative, we would be honored to receive your feedback.

  Check out the simulation at the link below, along with the TSX code — a great starting point for those who want to improve or run their own tests.

https://claude.ai/public/artifacts/fb787e0f-8019-4d3b-8d43-8308a0292a9a  

Sincerely,

--

José Ricardo de Paula  Técnico em Infraestrutura e Suporte Diretoria de Tecnologia da Informação e Comunicação

[Steve Kaeppler]

Ricardo-

I have thought for a while that a reasonably good project for an undergraduate or graduate student would be to shove spots into a machine learning algorithm and see what you find in terms of trends or compare against some sort of independent set of parameters.  In other words, if you consider the spot number per grid point pair as a function of time, day of year, MLT, other parameters, and you put that into a machine learning algorithm, can you make something predictive?  Another non machine learning approach might be to develop some sort of empirical basis functions that you could then parameterize similarly.  

I haven't extensively looked into the literature, but I would assume what I am suggesting has already been done in some form.  And actually if someone (Phil Erickson, David Themens, etc) knows of the references, I would be very interested to read the papers.

73,

Steve

AD0AE

--
Please follow the HamSCI Community Participation Guidelines at http://hamsci.org/hamsci-community-participation-guidelines.
---
You received this message because you are subscribed to the Google Groups "HamSCI" group.
To unsubscribe from this group and stop receiving emails from it, send an email to hamsci+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/hamsci/CAB%3DSXAXK3G9TUVW9r_Ab7Ufv2p6_aHDZS9u2t%2Bzk-SRkGcxWGg%40mail.gmail.com.

--

--------------------------------------------------------------------
Steve Kaeppler
steve.k...@gmail.com
cell: 262-707-0932
Amateur Radio Call: AD0AE/6

[Ricardo]

Hi Steve,

I'm not a scientist specializing in ionospheric studies, just an Information Technology professional, but I recognize that this is a highly complex topic that would likely require careful modeling for proper exploration.

It would be interesting to involve undergraduate and graduate students, including those in scientific initiation programs, in developing and testing various machine learning and deep learning models, such as Random Forests, Gradient Boosted Machines, LSTMs, and CNNs, as well as hybrid models. Incorporating XAI (explainable AI) techniques could also help us identify the most influential variables in propagation behavior.

Regarding ready-to-use systems, I currently only have a simple simulation that uses AI to explore the concept.

I'd appreciate your thoughts on this and would be happy to stay in touch in case this evolves into a collaborative opportunity.

73,

Ricardo de Paula – PY2VOX

To view this discussion visit https://groups.google.com/d/msgid/hamsci/CANK6cPp5%3DVjEwURjzMSc159YvR8KhwwNUEiN1xBYmJw0mL4uSg%40mail.gmail.com.

[Lawrence Naif]

HI AlL,

This is an interesting topic that I am very passionate about.    I am very much into developing an ML algorithm to predict TIDs.  I have some papers written on this and am happy to collaborate with ionospheric and solar physicists. I bring with me an actuarial modeling background.

For example, the attached MP4 shows the behaviour of total electron count as a function of time over the solar eclipse period in April 2024.  

Physical Meaning of Extracted Features

• wt = TECnorm: Normalized Total Electron Content, representing ionospheric electron density.
• xt = S4norm: Normalized scintillation index, indicating signal amplitude fluctuations due to ionospheric
irregularities.
• yt = RFInorm: Normalized Radio Frequency Interference, reflecting external noise or interference.
• zt = latitudenorm: Normalized zenith angle (derived from elevation), representing the satellite’s position
relative to the vertical.

I am mapping the features to an ionospheric grid with lon, lat, and then aggregating the variables per grid to detect abnormalities.

Happy to collaborate and discuss further.

Best Regards,

Lawrence Habahbeh

To view this discussion visit https://groups.google.com/d/msgid/hamsci/CANK6cPp5%3DVjEwURjzMSc159YvR8KhwwNUEiN1xBYmJw0mL4uSg%40mail.gmail.com.

[Andrew Rodland]

A bit of all of the above. "Just dump it into one big 'AI'" isn't really an approach I would espouse, but it's definitely time for some research into *targeted* models.

I've been doing some work on using latent diffusion to generate maps of ionospheric characteristics (foF2, hmF2, MUF(3000)), and it's recently started to show some promising results. My immediate goal is to plug it in as a new interpolation method for prop.kc2g.com, and do some GNSS-RO comparison to see how it compares quality-wise — but I also have some ideas on how it could be used to incorporate spot data and ionosonde data into a single forecast model.

Some other "AI" things that I think could be interesting:

* Create something that looks at the full spectrum of GOES SUVI images and see if they can be used to synthesize a better proxy for prop than F10.7. Besides the "it's UV, it *should* be better" factor, SUVI data is available near-real-time 24/7, compared to the three-times-daily cadence of Penticton F10.7.

* Try to predict the geomagnetic impact of an incoming CME in advance: major storm, or dud?

[Leandro Soares Indrusiak]

Dear all,

I am currently supervising a PhD student who is working on machine learning models to support scheduling under uncertainty in wireless comms, and we will use digimodes over HF as one of the case studies to evaluate the approach. In this particular case study, we are building datasets that can train models to learn correlations between the data received by a given station (e.g. FT8 messages decoded over a period of time) and the success of its subsequent transmission (e.g. PSK reporter or RBN hits after a transmission). Once the model is trained with data obtained over a large period of time, over different bands, and different levels of power, we hope to have models that can provide some guidance to a scheduler so that it can choose band, power levels and number of retransmissions so it can meet a target success rate. Our approach is different from some of those discussed in this thread because it does not use any ionospheric or solar data, only messages that a  simple station can decode on the air. The attached PDF tries to illustrate this case study. 

The PhD student has started only 4 months ago, so we are in the early stages of the project (the student receives a scholarship funded by ARDC - https://www.ardc.net). I'll update this group once we have results to share. But if anyone wants to know more, please feel free to contact me directly. 

Best,

Leandro G5LSI

University of Leeds, UK

[Ricardo]

Hello Lawrence,

Thank you for sharing your initiative. Seeing such concrete and well-structured applications of machine learning and ionospheric disturbances (TIDs) is inspiring, especially in relation to events such as solar eclipses. Your mapping and aggregation by ionospheric grid open up real possibilities for warning and forecasting systems, which would be extremely useful for the scientific community, radio operators, and HF communication systems.

Regarding our theme, "Artificial Intelligence in HF Propagation: Feasible, Innovative, or Premature?," I see your work as evidence that we are indeed at a feasible and innovative moment. Technology already enables us to combine geophysical time series, derived physical variables, and empirical records with ML/DL models for analysis and prediction.

I have been working on a project focusing on HF, especially the 160 m band. I am testing an experimental AI inference architecture that crosses solar and ionospheric data. I am also trying to connect APIs from the Logbook of the World (LoTW) to look for correlations between parameters such as signal-to-noise ratio (SNR), maximum usable frequency (MUF), total electron content (TEC), and effective propagation openings.

I believe approaches with solid physical foundations, spatial structuring, and indicators such as S4 and RFI are essential for validating and refining models, as yours demonstrates. In fact, I think that techniques such as long short-term memory (LSTM) networks, autoencoders, and explainable AI (XAI) can provide meaningful insights for interpreting the most influential variables.

I would be very happy to stay in touch and explore collaboration possibilities, although I am from the IT field and focus more on GTI than ionospheric modeling itself.

Congratulations on the initiative! Let's expand these connections between data science, space physics, and HF communication!

73 and a warm hug,
Ricardo de Paula – PY2VOX
🔗 linkedin.com/in/py2vox

To view this discussion visit https://groups.google.com/d/msgid/hamsci/CADHVvei4H-6Y0MfMrUn%2BaoY2TATsCsF_D9%3D4m3wDnA0P-WUymg%40mail.gmail.com.

[JORGE LOPEZ CANALES]

Claro que sera interesante el uso de IA en la predicción de eventos difíciles de propagacion .

To view this discussion visit https://groups.google.com/d/msgid/hamsci/CAB%3DSXAURp7BHHHEQi-H-Td9HkO25yT_wWZ%2B6%2Bwf_Q5-mgFvh1Q%40mail.gmail.com.