Researchers Deploy Deep Learning to Predict F10.7 and F30

A research team led by Zhenduo Wang published a new model, `SINet`, for daily forecasting of the solar radio flux indices F10.7 and F30, targeting medium-term horizons of 1-60 days. The authors trained SINet on historical measurements from NOAA and ground-based radio facilities at Toyokawa and Nobeyama, and report that SINet outperforms five benchmark statistical and deep-learning methods on the F10.7 task. This submission appears on arXiv and is associated with an AGU journal record (article number e2025JA034868).

The study frames the task as daily time-series forecasting for two solar radio flux indices with different sensitivity profiles: F10.7 (10.7 cm) is a long-standing proxy for solar UV output, and F30 (30 cm) is described as more sensitive for thermospheric response. The authors emphasize medium-term prediction windows up to 60 days. Key data inputs include:

• •Daily observational flux values from NOAA radio flux archives
• •Independent ground-based records from the Toyokawa and Nobeyama observatories
• •A benchmarking suite consisting of five statistical and deep-learning baselines (unnamed in the abstract)

`SINet` is presented as a task-specific deep-learning architecture for solar-index forecasting. The paper reports comparative experiments showing SINet yields superior predictive performance versus the selected baselines for F10.7, and it is the first deep-learning application the authors are aware of for daily F30 prediction. The public metadata and repository links on arXiv and institutional pages indicate open access to the preprint; the full paper will include architecture diagrams, training procedures, and evaluation protocols necessary for reproducibility.

Improved daily forecasts of solar radio flux have direct operational value for space-weather modeling. F10.7 is widely used to model upper-atmosphere heating and ionization, which affects satellite drag calculations and HF radio propagation. F30 is increasingly valuable because its spectral response can better capture variations that drive thermospheric density changes. Applying deep learning to these indices aligns with a broader trend of domain-adapted time-series models replacing or augmenting empirical and linear statistical predictors in geoscience. This work is notable for two reasons: it reports measurable gains on a long-standing operational proxy (F10.7), and it extends deep-learning methods into the less-explored F30 index. For practitioners, a validated ML model that produces more accurate medium-term forecasts can be integrated into propagation, orbit prediction, and collision-avoidance pipelines.

Confirmatory details in the full paper are important: the exact model architecture, temporal input windows, loss functions, training-validation splits across solar cycle phases, and absolute error metrics versus operational thresholds. Follow-up work should test generalization across solar maximum and minimum and evaluate end-to-end impact on thermosphere density forecasts and downstream operational systems.