Volcanic eruptions generate atmospheric gravity waves that propagate into the ionosphere and can cause significant ionospheric disturbances, including Equatorial Plasma Bubbles (EPBs) and large variations in Total Electron Content (TEC). These disturbances affect radio and satellite communications, degrading GPS accuracy and other services. This study evaluates a Hybrid ML-DL Model—a weighted combination of Light Gradient Boosting Machine (LightGBM) and Long Short-Term Memory (LSTM) networks—for forecasting TEC during eruption-driven perturbations. The goal was to assess whether the hybrid approach improves TEC prediction accuracy compared with individual LightGBM and LSTM models and with conventional Autoregressive (AR) and Multiple Linear Regression (MLR) models.
Volcanic eruptions launch intense atmospheric gravity waves through vertical air displacements. When these waves reach the ionosphere they perturb ion and electron densities, producing irregularities such as EPBs—localized low-density pockets that scatter radio waves and disrupt signal propagation. Such ionospheric perturbations can lead to GPS positioning errors and degraded satellite communications.
Contemporary TEC forecasting methods include empirical, physics-based and machine learning approaches. The authors note these methods often perform well during quiet conditions but lose accuracy during extreme or rapidly evolving space-weather events. Models based on global indices or spherical harmonics can capture regular diurnal and seasonal patterns but degrade during storms. Deep learning architectures can learn complex dynamics but tend to be computationally intensive, less interpretable, and sensitive to data gaps. The study aims to address these limitations by testing a hybrid ML–DL architecture on eruption-related TEC anomalies across multiple events and latitudes.
TEC data used in the analysis were obtained primarily from the BAKO GPS station in Indonesia, supplemented by observations at other latitudes to examine model generalizability. Six volcanic eruptions were selected as case events: Mt. Kelud (2014), Mt. Sinabung (2016), Mt. Semeru (2021), Mt. Ruang (2024), Mt. Etna (2013), and Mt. La Soufriere (2021). The dataset and supporting information are reported as available with the publication.
The proposed Hybrid ML-DL Model combines predictions from a LightGBM model and an LSTM network using a weighted average. LightGBM is an ensemble gradient-boosting tree method well-suited to tabular and feature-driven forecasting, while LSTM is a recurrent neural network architecture designed for sequence modeling and temporal dependencies. The hybrid approach leverages strength from both: LightGBM’s structured-feature learning and LSTM’s temporal pattern extraction. The exact model weights and hyperparameter details were reported in the full methods of the source article.
Model comparisons included standalone LightGBM and LSTM models and two conventional statistical approaches: Autoregressive (AR) and Multiple Linear Regression (MLR). These baselines assess whether the hybridization yields consistent improvement over both modern ML/DL and classical time-series regression techniques during perturbed ionospheric conditions.
Four evaluation measures were used to quantify model performance; the abstract highlights Root Mean Square Error (RMSE) and reports other metrics such as Normalized RMSE (NRMSE), Mean Bias Deviation (MBD), and Relative Log Error (RLE) in examples. Models were trained and tested on the eruption-related TEC time series to evaluate forecasting accuracy during disturbed periods.
Across all six eruption cases the Hybrid ML-DL Model consistently outperformed the independent LightGBM and LSTM models as well as the AR and MLR approaches. The hybrid model produced lower errors on the reported evaluation metrics, demonstrating better fit to the disturbed TEC time series and more accurate short- to medium-term forecasting in the studied events.
The Mt. Ruang eruption (April 16–May 7, 2024) produced the largest disturbance analyzed, with a maximum perturbation of approximately 100–110 TECU. For this event the Hybrid ML-DL Model achieved an RMSE of 2.841 TECU. By comparison, RMSEs for the other models were reported as LightGBM 3.484 TECU, LSTM 5.084 TECU, MLR 5.345 TECU and AR 6.285 TECU. The Hybrid model also recorded the lowest values for NRMSE (0.030), MBD (0.546 TECU) and RLE (0.069) among the compared models in this case.
The findings indicate that combining tree-based gradient boosting and sequence-learning networks can improve TEC forecasting during ionospheric disturbances driven by volcanic eruptions. Improved TEC prediction has practical implications for mitigation planning in satellite communications, GPS-dependent navigation, and other services sensitive to ionospheric variability. The study reinforces that models trained primarily on quiet periods will underperform during extreme events and that hybrid architectures can enhance robustness across disturbances and geographic coverage.
The Hybrid ML-DL Model (LightGBM + LSTM) demonstrated consistent improvement over individual ML/DL models and traditional statistical models for TEC prediction during six volcanic eruptions. The source article states that the data are available as Supplementary Information. Funding for the work was provided by the Anusandhan National Research Foundation (project ID 2375), and the authors declared no competing interests.