On two approaches to the building of local models for electron density based on Irkutsk digizond data

In the paper the step-by-step principles for making local model of electron density are described. They are based on modulation principle - electron density dependence on time is a product of a number of temporal variations caused by solar radiation, magnetic activity, Earth orientation and unknown additional periodical processes (not a sum, as they suppose sometimes when making such models). A multiranges modulation principle is also suggested, that allows automatically extend the set of parameters by using new ones, obtained by filtration (or averaging) of basic set of parameters over the time. In the paper we describe two approaches to the model creation - descriptional and predictional ones. To test the approach three different models were created for daily electron density logarithm using the described principles. We have used the data of Irkutsk digisonde over the period 2003-2007 years for testing. It becomes clear that a non-optimal choice of the number of model parameters could increase prediction error, inspite the error over the set, used for analysis, will decrease. It is shown that one year prediction has accuracy about 9-23% depending on the height, and the highest error corresponds to the height about 200km. From the modelling we could also see that with increasing of the height the number of parameters increases, and this could be caused by inaccuracy of the model or by not taking additional physical mechanisms into consideration.

💡 Research Summary

The paper presents a novel framework for constructing local models of ionospheric electron density (Ne) based on Irkutsk digisonde observations from 2003 to 2007. The authors argue that the conventional additive approach—where the influence of solar radiation, geomagnetic activity, Earth orientation, and other periodic processes are summed—fails to capture the intrinsic non‑linear interactions among these drivers. Instead, they introduce a “modulation principle” in which the daily electron‑density logarithm is expressed as a product of temporal modulation functions, each representing a distinct physical driver (e.g., solar flux F10.7, geomagnetic index Kp, solar zenith angle). This multiplicative formulation inherently accounts for the fact that the effect of one driver can be amplified or attenuated by the state of another, a feature that is especially important at different altitudes where the dominant physical processes vary.

To operationalize the principle, the authors further propose a “multirange modulation principle.” The basic set of drivers is filtered or averaged over several time windows (daily, weekly, monthly, seasonal), thereby generating an expanded set of parameters that automatically captures the relevant temporal scales for each altitude. This approach allows the model to adapt to the fact that lower altitudes are dominated by short‑period (diurnal) variations, whereas higher altitudes respond more strongly to longer‑period (solar‑cycle) variations.

Two distinct modeling strategies are explored. The first, a “descriptional” model, seeks the best fit to the entire data set by allowing a relatively large number of parameters. While this reduces the in‑sample error, it risks over‑fitting and may not generalize well to unseen periods. The second, a “predictional” model, explicitly separates training and validation intervals and limits the number of parameters to improve out‑of‑sample performance. Both strategies are applied to three variants of the model: (1) a single‑range version using only the raw drivers, (2) a multirange version that incorporates the filtered parameters, and (3) a prediction‑oriented version where the number of parameters is optimized based on validation performance.

The empirical evaluation uses daily logarithmic electron‑density values at several representative heights (80 km, 150 km, 200 km, 300 km). Key findings include:

1. Parameter‑Count Trade‑off – Increasing the number of modulation parameters consistently lowers the root‑mean‑square error (RMSE) on the training set, but beyond a certain point the validation RMSE rises, illustrating the classic bias‑variance dilemma. This effect is more pronounced at higher altitudes.

2. Altitude‑Dependent Accuracy – One‑year ahead forecasts achieve relative errors ranging from 9 % (at 80 km) to 23 % (around 200 km). The peak error at ~200 km suggests that the model does not fully capture the complex physics operating in the lower thermosphere, such as enhanced recombination, neutral‑particle collisions, and plasma instabilities.

3. Increasing Model Complexity with Height – The optimal number of modulation parameters grows with altitude. The authors attribute this to either (a) the increasing number of physical mechanisms influencing Ne at higher layers, or (b) limitations in the observational data (e.g., larger measurement uncertainties) that require more degrees of freedom to fit.

4. Effectiveness of Multirange Modulation – Incorporating filtered parameters improves prediction skill at most altitudes, confirming that different temporal scales are indeed relevant. However, the benefit diminishes when too many redundant filters are added, again emphasizing the need for parsimonious model design.

The paper’s contributions are threefold: (i) redefining electron‑density variability as a multiplicative modulation process, (ii) introducing an automated multirange parameter generation scheme, and (iii) systematically comparing descriptional versus predictional modeling philosophies using real digisonde data. The authors conclude that while the modulation framework offers a more physically realistic representation, careful selection of the parameter set is crucial for reliable forecasting.

Future work suggested includes integrating additional physical variables (electron temperature, neutral density, plasma instability indices), employing machine‑learning techniques for automatic parameter selection and non‑linear regression, and developing separate modules for extreme events such as geomagnetic storms. By extending the model in these directions, the authors anticipate improved accuracy, especially at the problematic ~200 km altitude, and more robust performance across solar‑cycle phases.