A forecast model for ionospheric TEC during solar flare events based on low-latitude GPS stations data Authors: R. Mukesh, Punyawi Jamjareegulgran, S. Riswana Fathima, J. Rubashri, P. Priyavarshini, J. Priyadharshini, Sarat C. Dass, S. Kiruthiga, A. R. S. Sahanaa, M. Sudhandra, S. Safana, S. Karthick, Prasert Kenpankho, Supawit Nambut, Pattawut Wongsak Published: 2026 ABSTRACT Solar flares can strongly disturb the ionosphere and significantly affect radio communication and satellite navigation systems. In this paper, the Ionospheric responses to four X-class solar flare events that occurred during the 25th solar cycle were investigated through the prediction of Total Electron Content (TEC). A Kolmogorov-Arnold Network (KAN) algorithm-based deep learning approach, combined both with the Adaptive Gradient Algorithm (Adagrad) optimiser and without an optimiser, was implemented to forecast TEC variations. A Deep Belief Network with Adagrad optimiser (DBN-Adagrad) is a deep learning model used to provide a clear comparison with the proposed KAN-based approach. The proposed models were analysed with three sets of dataset split strategies, and also the models were validated with Leave future out cross-validation. A geomagnetically quiet day on 28 April 2025, with Kp < 3, is also considered as a reference for evaluating model performance under undisturbed ionospheric conditions. These KAN-based TEC predictions were compared with TEC predictions based on the DBN-Adagrad, IRI-2020 model and with GNSS–derived TEC observations at the HYDE (Hyderabad, India) and CPN (Chumphon, Thailand) stations. The proposed KAN and DBN models used the important solar and geomagnetic parameters, such as Ap, Kp indices, SSN, and the 10.7 cm solar radio flux, as input parameters during quiet and disturbed conditions. Four major solar flares (SF) events on 26th October 2024 (X1.86 SF), 13th May 2025 (X1.21 SF), 14th May 2025 (X2.7 SF), and 25th May 2025 (X1.1 SF), were analysed orderly in this study. Moreover, model performance was evaluated using standard statistical measures, including Willmott’s Index of Agreement (WIA), Kling-Gupta Efficiency (KGE) and Spearman’s Correlation Coefficient (SCC). As for the X1.86 SF event (first SF) observed over HYDE and CPN stations, the KAN with Adagrad optimiser and DBN with Adagrad optimiser models achieved KGE values of 0.8586,0.7885 and 0.8573,0.8248, respectively, demonstrating their good performance as compared to other models (KAN without optimiser and IRI-2020 models). Overall, the results showed that the KAN model with Adagrad optimiser and DBN with Adagrad optimiser generally provided TEC predictions with improved accuracy at both stations as compared to other models, except on May 13th, 2025, over the HYDE station, where the KAN without optimiser performed better. As for the X1.1 SF event on 25th May 2025 (fourth SF) over HYDE station, two KAN models with Adagrad optimiser and without optimiser exhibited nearly equal performance. This study demonstrated that AI-driven deep learning models can effectively capture TEC variations during strong SF events and play a valuable role in improving the reliability of global navigation satellite systems (GNSS) under disturbed space weather conditions. Introduction Solar flares (SFs) are an immense explosion on the Sun that happens when energy stored in twisted magnetic fields due to plasma motion is suddenly released, causing immediate, abnormal ionisation in the lower layers of the ionosphere, thereby increasing electron density. Total Electron Content (TEC) is the integration of electron density along the whole path between a satellite and its receiver. Radio waves are affected by the presence of electrons. The more electrons in the path of the radio wave, the more the radio signal will be affected. Since communication and GNSS (Global Navigation Satellite System) systems use radio frequencies, navigation systems are also affected. TEC is a good parameter to monitor for potential space-weather impacts. During a SF event, large amounts of X-ray and extreme ultraviolet radiation are emitted and reach the Earth in approximately eight minutes1,2. This high-energy radiation is quickly absorbed by the dayside ionosphere, which triggers strong photoionisation and rapidly increases the free electron population. The release of additional electrons causes a rapid rise in the electron density. Consequently, the ionospheric TEC increases sharply during solar flare events3–9. Wulansari et al.10 analysed the Ionospheric response to a strong solar flare of X1.1 class that occurred on November 8th, 2013. GNSS observation data were collected from the Sumatran GPS Array (SuGAR) stations distributed across Indonesia. Ionospheric electron density disturbances were analysed by using the GNSS-TEC method, such as the detection of Sudden Ionospheric Disturbance (SID) using multiple GPS stations and the Detection of SID using all GPS satellites. The results suggested that the GNSS-TEC method was highly effective for understanding the Ionospheric disturbances caused by X-class flare events. Chakraborty et al.11 investigated TEC variations during X-class solar flares and geomagnetic storm conditions, highlighting the combined influence of solar and geomagnetic activity on ionospheric behaviour. Sanjay Kumar and Singh12, Singh et al.13, Domingo et al.14 investigated the solar flares at low latitude and showed that the magnitude of TEC enhancement depended on the class of the solar flares. Han et al.15 evaluated four ML algorithms (NN, LSTM, ANFIS, GBDT) for short-term TEC prediction at low-latitude IGS stations during high solar activity, finding that the GBDT model outperformed the others. Chen et al.16 investigated TEC prediction using deep learning approaches, such as spatio-temporal graph neural networks, for global ionospheric TEC prediction. Lobanova et al.17 examined TEC variations at the Mikhnevo Geophysical Observatory during X-class flares on 6th–7th September 2017, showing improved prediction accuracy during the recovery phase using a quiet-day background detrending approach. Zhao et al.18 investigated an Ionospheric TEC prediction model based on Variational Mode Decomposition (VMD) and heterogeneous ensemble learning. TEC data from six GNSS stations along with solar and geomagnetic activity parameters, were used for TEC prediction. Three heterogeneous models, namely BiLSTM, BiTCN and DBN, were used for final prediction. Modal performance was evaluated using R2, RMSE, MAPE and MAE metrics. The results showed that the proposed model achieved higher accuracy and stability during geomagnetic disturbances, demonstrating strong adaptability in low-latitude regions. Pietrow et al.19 analysed spectroscopic characteristics of X2.2 and X9.3 class solar flares on 6th September 2017, demonstrating that Hα-based indices reliably represent flare timing and strength. Zhou et al.20 analysed the influence of ionosphere disturbances on GNSS signal and positioning due to the X9.3 solar flare event on September 6th, 2017 and the M1.6 solar flare event on October 9th, 2021. Saharan et al.21 investigated the ionospheric response to 450 solar flares that occurred during 2014 (the peak of solar cycle 24) via GPS TEC signals from Bangalore (IGS station at low latitude). Previous research indicated that the ionospheric effects of solar flares depended on multiple parameters such as flare intensity, solar geometry, local time, and background space weather conditions. Observations from the GNSS-derived TEC, VLF signals and satellite measurements revealed strong regional and event-dependent variability in TEC enhancements, particularly over low-altitude and equatorial regions. Although several machine learning and signal processing methods have been developed for TEC prediction, many of these approaches struggle to capture the fast and nonlinear TEC changes associated with extreme flare activity. Their performances were often affected by seasonal variability and geomagnetic disturbances. To effectively capture the non-linear and dynamic nature of the ionospheric variations during solar flares, a robust alternative modelling strategy is necessary. The Kolmogorov-Arnold Network (KAN) model addresses this need by learning multivariate nonlinear relationships through flexible univariate functions. Hence, the main contribution of this research is the development of a KAN-based framework for TEC prediction during solar flares. To predict the TEC variabilities, the solar and geomagnetic parameters, as well as the Adaptive Gradient Algorithm (Adagrad) optimiser, are also applied along with the KAN algorithm (namely, KAN-Adagrad). Finally, it can be seen that the KAN-Adagrad model generally provides competitive accuracy for multivariate time-series forecasting under solar flare conditions as compared to the TEC values predicted by KAN without an optimiser (namely, the KAN model), Deep Belief Network with Adagrad (namely, DBN-Adagrad) and the IRI-2020 model (IRI-2020). While various deep learning models, such as Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), have been commonly applied in the ionospheric TEC prediction, they typically require a large amount of data for training, and are not able to capture fast nonlinear changes of TEC during extreme solar flare events. In this paper, we choose the Deep Belief Network (DBN) as a comparison model in this study, because the hierarchical pre-training strategy of the DBN model can learn complex feature representations from the limited training sample sets. In addition, the Adagrad optimiser is used to train the DBN model (called DBN-Adagrad) to provide a consistent comparison with the proposed KAN-Adagrad model. Note that IRI, International Reference Ionosphere, is an international project sponsored by the International Union of Radio Science (URSI) and the Committee on Space Research (COSPAR). It constructs an empirical standard model of the ionosphere to provide monthly averages of electron density, TEC, F2-layer critical frequency, etc., with the ionospheric altitude22. The latest version of the IRI model is the IRI-2020 model. This paper is organised as follows. Section 2 describes the data sources and the methodology used in this study. It includes the Kolmogorov-Arnold Network (KAN) algorithm (Sect. 2.1), the Adaptive Gradient (Adagrad) optimiser (Sect. 2.2), and the Deep Belief Network (DBN) (Sect. 2.3). Section 2.4 presents the statistical evaluation parameters used to assess the model performance, including Willmott’s Index of Agreement (WIA), Kling-Gupta Efficiency (KGE), and Spearman’s Correlation Coefficient (SCC). Section 2.5 presents the analysis of training data size and explains the dataset split strategy used in this study. Section 2.6 describes the model validation using PSIS-LFO-CV. Section 3 presents the TEC prediction results for the selected X-class solar flare events at the HYDE and CPN stations and discusses the performance of the KAN, KAN-Adagrad, DBN-Adagrad, and IRI-2020 models. It also includes the TEC prediction analysis during the geomagnetically quiet day on 28th April 2025 at HYDE station (Sect. 3.5) and at CPN station (Sect. 3.6), followed by the overall model performance comparison for TEC prediction (Sect. 3.7) and Sect. 3.8 presents the statistical comparison of the models using the Friedman test. Finally, Sect. 4 concludes the paper by summarising the main findings and highlighting the overall performance of the proposed approach. Data and methodology In this research, we focused on the impacts on ionospheric TEC variations created by four SF events, such as X1.86 SF, X1.21 SF, X2.7 SF and X1.1 SF that occurred on 26th October 2024, 13th May 2025, 14th May 2025 and 25th May 2025, respectively, on the ionosphere. A geomagnetically quiet day on 28 April 2025, with Kp < 3, is considered as a reference to evaluate model performance under undisturbed ionospheric conditions at both the HYDE and CPN stations. We considered four prediction models to study TEC variations in this work, for instance, the KAN model without optimiser (so-called KAN model), the KAN algorithm with Adagrad optimiser (so-called KAN-Adagrad), the Deep Belief Network with Adagrad (so-called DBN-Adagrad) and the IRI-2020 model (so-called IRI-2020). Moreover, Statistical parameters [e.g., Willmott’s Index of Agreement (WIA), Kling-Gupta Efficiency (KGE), and Spearman Correlation Coefficient (SCC)] are also used to evaluate the performance of the KAN model, KAN-Adagrad model, DBN-Adagrad model and IRI-2020 model. Firstly, one set of TEC data over Hyderabad station (HYDE), India, was obtained from the IONOLAB website. The TEC data obtained from the IONOLAB were extracted with 576 samples per day. To reduce the complexity and noise of the data, it was then down-sampled by taking the average of every 24 measurements, resulting in 24 data points per day, which were used as model inputs. Note that IONOLAB is a research group and laboratory that focuses on imaging and monitoring the ionosphere23. The latitude and longitude of the HYDE station are 17.42°N and 78.55°E, respectively. Secondly, another set of TEC values was processed from a GNSS receiver at Chumphon station (CPN), Thailand. The latitude and longitude of the CPN station are 10.72°N and 99.37°E, respectively. Both HYDE and CPN stations are located in low-latitude regions. Thirdly, the TEC values predicted by the IRI-2020 model (namely, IRI-2020 predicted TEC) were also collected and processed over HYDE and CPN stations. The input parameters given to the proposed KAN model included solar and geomagnetic indices such as sunspot number (SSN), 10.7 cm solar radio flux (F10.7), as well as Ap and Kp indices. These parameters were taken from NASA’s OMNIWeb. X-class solar flares are rare and extreme events, which naturally limit the availability of observational data. In this study, four events with intensities ranging from X1.1 to X2.7 were selected to capture different levels of ionospheric disturbances. A similar event-based approach has also been adopted in previous ionospheric TEC studies5. Note that OMNIWeb is a NASA website that provides a collection of near-Earth solar wind and magnetic field data24. As shown in Fig. 1, the forecasting framework integrates solar/geomagnetic indices and TEC observations of solar flares into a unified modelling pipeline. Four solar flare dates, a quiet day, and their corresponding training periods were selected to capture both disturbed and quiet ionospheric conditions, ensuring the models are exposed to diverse scenarios. TEC data was down-sampled to maintain consistency and reduce noise. The modelling stage comprised three models: KAN with Adagrad optimiser, KAN without optimiser and DBN with Adagrad optimiser. Each branch follows a structured training pipeline, beginning with parameter initialisation and proceeding through propagation, optimisation, or pretraining steps. Convergence was tested in each case. If convergence is achieved, it yields an optimised or converged model; if it is not, it yields partially trained models. The converged and optimised models proceed to the evaluation stage. Performance evaluation was conducted using three: Weighted Index of Agreement (WIA), Spearman’s Correlation Coefficient (SCC) and Kling-Gupta Efficiency (KGE). These metrics provided a balanced assessment of predictive accuracy, correlation strength and overall efficiency. Fig. 1Methodology of research work. Methodology of research work. KAN algorithm The Kolmogorov-Arnold Network (KAN) is a recently developed class of neural network architecture that is capable of effectively and transparently modelling complex non-linear relationships such as ionospheric TEC variations. Based on the Kolmogorov-Arnold representation theorem, the KAN algorithm is developed as an effective alternative for the traditional multilayer perceptron models, by enabling any multivariate continuous function to be expressed as a composition and summation of multiple univariate functions25. The KAN algorithm is mathematically expressed as follows:1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left({x}_{1},{x}_{2},\dots\:,{x}_{n}\right)={\sum\:}_{q=1}^{2n+1}{g}_{q}\left({\sum\:}_{p=1}^{n}{{\Phi\:}}_{pq}\left({x}_{p}\right)\right)$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left({x}_{1},{x}_{2},\dots\:,{x}_{n}\right)$$\end{document} - multivariate continuous function.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}_{pq}\left({x}_{p}\right)$$\end{document} - univariate inner functions that transform each input variable.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}(.)$$\end{document} - outer nonlinear functions that combine the inner transformations.n - input dimension.q - summation terms. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left({x}_{1},{x}_{2},\dots\:,{x}_{n}\right)$$\end{document} - multivariate continuous function. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:f\left({x}_{1},{x}_{2},\dots\:,{x}_{n}\right)$$\end{document} - multivariate continuous function. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}_{pq}\left({x}_{p}\right)$$\end{document} - univariate inner functions that transform each input variable. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{{\Phi\:}}_{pq}\left({x}_{p}\right)$$\end{document} - univariate inner functions that transform each input variable. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}(.)$$\end{document} - outer nonlinear functions that combine the inner transformations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{q}(.)$$\end{document} - outer nonlinear functions that combine the inner transformations. n - input dimension. n - input dimension. q - summation terms. q - summation terms. Adagrad optimiser The Adaptive Gradient (Adagrad) optimiser adapts the learning rate individually for each parameter based on the cumulative sum of its squared gradients. This allows frequently updated parameters to receive smaller updates and rarely updated ones to receive larger updates26. This is a useful feature when handling heterogeneous or sparse data. The mathematical expression of the Adagrad optimiser is shown below:2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{t}={G}_{t-1}+{g}_{t}\odot\:{g}_{t}$$\end{document}3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\theta\:}_{t+1}={\theta\:}_{t}-\frac{\eta\:}{\sqrt{{G}_{t}+\epsilon}}\odot {g}_{t}$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{t}={\nabla\:}_{{\theta\:}_{t}}{f}_{t}\left({\theta\:}_{t}\right)$$\end{document} be the gradient at step t.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{t}\:$$\end{document}- diagonal matrix containing the sum of squared gradients up to time t.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} - initial learning rate.⊙ - element-wise multiplication.ε - small constant to prevent division by zero. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{t}={\nabla\:}_{{\theta\:}_{t}}{f}_{t}\left({\theta\:}_{t}\right)$$\end{document} be the gradient at step t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{g}_{t}={\nabla\:}_{{\theta\:}_{t}}{f}_{t}\left({\theta\:}_{t}\right)$$\end{document} be the gradient at step t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{t}\:$$\end{document}- diagonal matrix containing the sum of squared gradients up to time t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{G}_{t}\:$$\end{document}- diagonal matrix containing the sum of squared gradients up to time t. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} - initial learning rate. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\eta\:$$\end{document} - initial learning rate. ⊙ - element-wise multiplication. ⊙ - element-wise multiplication. ε - small constant to prevent division by zero. ε - small constant to prevent division by zero. Deep belief network A Deep Belief Network (DBN) is a probabilistic generative model composed of multiple stacked layers of Restricted Boltzmann Machines (RBMs). DBN learns hierarchical feature representations from input data through a two-phase training process: unsupervised greedy layer-wise pre-training followed by supervised fine-tuning27. The mathematical expression of the DBN energy function is shown below:4\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$E(v,h) = - \sum_i b_i v_i - \sum_j c_j h_j - \sum_{i,j} v_i w_{ij} h_j$$\end{document}5\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$P(v, h) = e^{-E(v,h)} /Z$$\end{document} where: where: E(v, h) - energy function of the RBM.vi - visible unit (input).hj - hidden unit (feature layer).wij - weight matrix between visible and hidden units.bi - bias of visible unit.cj - bias of hidden unit.Z - partition function (normalization constant).P(v, h) - joint probability of visible and hidden units. E(v, h) - energy function of the RBM. E(v, h) - energy function of the RBM. vi - visible unit (input). vi - visible unit (input). hj - hidden unit (feature layer). hj - hidden unit (feature layer). wij - weight matrix between visible and hidden units. wij - weight matrix between visible and hidden units. bi - bias of visible unit. bi - bias of visible unit. cj - bias of hidden unit. cj - bias of hidden unit. Z - partition function (normalization constant). Z - partition function (normalization constant). P(v, h) - joint probability of visible and hidden units. P(v, h) - joint probability of visible and hidden units. Table 1Parameters and values for training the KAN algorithm without adagrad optimiser.ParametersValuesModel TypeKAN (Kolmogorov – Arnold Network)Optimization MethodNo OptimiserEpochs500Hidden Units64Learning Rate0.01Train/Test Split80% - training dataset and 20% - testing datasetNormalizationZ-score Table 1Parameters and values for training the KAN algorithm without adagrad optimiser.ParametersValuesModel TypeKAN (Kolmogorov – Arnold Network)Optimization MethodNo OptimiserEpochs500Hidden Units64Learning Rate0.01Train/Test Split80% - training dataset and 20% - testing datasetNormalizationZ-score Parameters and values for training the KAN algorithm without adagrad optimiser. Table 2Parameters and values for training KAN and DBN algorithms with adagrad optimiser.ParametersValuesValuesModel TypeKAN (Kolmogorov – Arnold Network)DBN (Deep Belief Network)Optimization MethodAdagrad OptimiserAdagrad OptimiserEpochs500500Hidden Units6464Learning Rate0.010.01Train/Test Split80% - training dataset and 20% - testing dataset80% - training dataset and 20% - testing datasetNormalizationZ-scoreZ-score Table 2Parameters and values for training KAN and DBN algorithms with adagrad optimiser.ParametersValuesValuesModel TypeKAN (Kolmogorov – Arnold Network)DBN (Deep Belief Network)Optimization MethodAdagrad OptimiserAdagrad OptimiserEpochs500500Hidden Units6464Learning Rate0.010.01Train/Test Split80% - training dataset and 20% - testing dataset80% - training dataset and 20% - testing datasetNormalizationZ-scoreZ-score Parameters and values for training KAN and DBN algorithms with adagrad optimiser. From Tables 1 and 2, the KAN model is trained without an optimiser and with Adagrad optimiser and DBN is trained only with Adagrad optimiser. The KAN without optimiser relies on basic gradient descent, and the KAN and DBN with Adagrad optimiser adaptively adjust learning rates for each parameter. Epochs represent the number of complete passes through the training dataset. 500 epochs were chosen because they provide sufficient iterations for convergence while avoiding excessive computation. Hidden units denote the number of neurons in the hidden layer. For the KAN model, 64 units were selected and for the DBN model, a configuration of 64 units in the first layer and 32 units in the second layer was selected for generalisation and balance complexity, offering enough power to capture temporal patterns without overfitting. The learning rate defines the step size for parameter updates, and a 0.01 was chosen because it ensures stable convergence without overshooting minima. The train/test split was explained in Sect. 2.5. Normalisation means scaling of input data; Z-score standardisation was applied to centre and scale TEC values, thus improving convergence and stabilising training. Statistical parameters Willmott’s index of agreement Willmott’s Index of Agreement (WIA) is a standardised metric used to quantify the degree of prediction error in models28. It ranges from 0 to 1, where 1 indicates perfect agreement between the model’s predictions and True TEC, and 0 indicates no agreement. The WIA is derived from a ratio of the mean squared error to the potential error as shown in Eq. (6), providing a holistic evaluation of model performance by considering both systematic and random discrepancies between the observed and modelled data.6\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:d=1-\frac{{\sum\:}_{i=1}^{n}{({P}_{i}-{O}_{i})}^{2}}{{\sum\:}_{i=1}^{n}{({|P}_{i}-{\stackrel{-}{O}|+|O}_{i}-\stackrel{-}{O}|)}^{2}}$$\end{document} where: where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{i}$$\end{document} - Predicted value.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{i}$$\end{document} - True TEC value.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{O}$$\end{document} - Mean of True TEC values. n - number of samples. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{i}$$\end{document} - Predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{P}_{i}$$\end{document} - Predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{i}$$\end{document} - True TEC value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{O}_{i}$$\end{document} - True TEC value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{O}$$\end{document} - Mean of True TEC values. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\stackrel{-}{O}$$\end{document} - Mean of True TEC values. n - number of samples. n - number of samples. Kling-gupta efficiency Kling-Gupta Efficiency (KGE) is a goodness-of-fit indicator as expressed in Eq. (7) that tells how well a model’s predicted data matches the observed data, combining correlation, bias and variability into one score. It ranges from -∞ to 1, where 1 indicates a perfect model fit29.7\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:KGE=1-\sqrt{{(r-1)}^{2}+{(\beta\:-1)}^{2}+{(\alpha\:-1)}^{2}}$$\end{document} where: r - Pearson correlation coefficient.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} - variability of prediction errors computed by Eq. (8)\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} - bias term calculated by Eq. (9) 8\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:=\frac{{\sigma\:}_{s}}{{\sigma\:}_{o}}$$\end{document} 9\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:=\frac{{\mu\:}_{s}}{{\mu\:}_{o}}$$\end{document} r - Pearson correlation coefficient. r - Pearson correlation coefficient. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} - variability of prediction errors computed by Eq. (8) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\alpha\:$$\end{document} - variability of prediction errors computed by Eq. (8) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} - bias term calculated by Eq. (9) \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\beta\:$$\end{document} - bias term calculated by Eq. (9) where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{s}$$\end{document} - mean of the predicted value.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{o}$$\end{document} - mean of the observed value.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{s}$$\end{document} - standard deviation of the predicted value.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{o}$$\end{document} - standard deviation of the observed value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{s}$$\end{document} - mean of the predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{s}$$\end{document} - mean of the predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{o}$$\end{document} - mean of the observed value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\mu\:}_{o}$$\end{document} - mean of the observed value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{s}$$\end{document} - standard deviation of the predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{s}$$\end{document} - standard deviation of the predicted value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{o}$$\end{document} - standard deviation of the observed value. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{\sigma\:}_{o}$$\end{document} - standard deviation of the observed value. Spearman’s correlation coefficient Spearman’s Correlation Coefficient (SCC) is a non-parametric statistical measure, as stated in Eq. (10), that quantifies the strength and direction of a monotonic relationship between two variables, using their ranks instead of their actual values. It ranges from − 1 to + 1, in which 0 indicates no correlation30.10\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{s}=\rho\:\left[R\left[X\right],R\left[Y\right]\right]=\frac{cov[R\left[X\right],R\left[Y\right]]}{\left(\sigma\:R\left[X\right]\right)\left(\sigma\:R\left[Y\right]\right)}$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} - Pearson correlation coefficient operator.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cov[R\left[X\right],R\left[Y\right]]$$\end{document} - covariance of the rank variables.\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:R\left[X\right],\:\sigma\:R\left[Y\right]$$\end{document} - standard deviations of the rank variables. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} - Pearson correlation coefficient operator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\rho\:$$\end{document} - Pearson correlation coefficient operator. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cov[R\left[X\right],R\left[Y\right]]$$\end{document} - covariance of the rank variables. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:cov[R\left[X\right],R\left[Y\right]]$$\end{document} - covariance of the rank variables. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:R\left[X\right],\:\sigma\:R\left[Y\right]$$\end{document} - standard deviations of the rank variables. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\sigma\:R\left[X\right],\:\sigma\:R\left[Y\right]$$\end{document} - standard deviations of the rank variables. When all n ranks are distinct integers (no ties), it can be computed using Eq. (11).11\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{r}_{s}=1-\frac{6\sum\:{d}_{i}^{2}}{n({n}^{2}-1)}$$\end{document} where: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}\equiv\:R\left[{X}_{i}\right]-R\left[{Y}_{i}\right]$$\end{document} - the difference between the two ranks of each observation. n - the number of observations. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}\equiv\:R\left[{X}_{i}\right]-R\left[{Y}_{i}\right]$$\end{document} - the difference between the two ranks of each observation. \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{d}_{i}\equiv\:R\left[{X}_{i}\right]-R\left[{Y}_{i}\right]$$\end{document} - the difference between the two ranks of each observation. n - the number of observations. n - the number of observations. Analysis of training data size and dataset split strategy To determine the optimum training data size for TEC prediction, the KAN-Adagrad and DBN-Adagrad models were trained using three different training data sizes: 30 days, 60 days, and 100 days. The predicted TEC values were compared with the actual TEC values from the test data, and the Mean Absolute Error (MAE) was calculated. The size of the training data is an important factor in time-series prediction, as it affects both the accuracy and generalisation ability of the model31. Table 3 presents the MAE values for both models under the three training data sizes. From Table 3, it is observed that the KAN-Adagrad model gives the best performance with 100 days of training data, achieving the lowest MAE of 3.5215 TECU. Since this provides more stable and accurate results, the 100-day dataset was selected for the main experiments in this study. After selecting the training data size, the 100-day dataset was divided using an 80/20 split. The first 80% of the data was used for training, and the remaining 20% was used for testing. This split was done in a chronological order to preserve the natural sequence of TEC variations and to ensure that future data was not used during training. This helps to avoid information leakage and provides a more reliable evaluation of the model performance31. Table 3Mean absolute error (MAE) of KAN-Adagrad and DBN-Adagrad models for different training data sets.Training Data (Days)KAN-Adagrad MAE (TECU)DBN-Adagrad MAE (TECU)3010.210010.8037609.884310.0219 100 3.5215 4.0216 Table 3Mean absolute error (MAE) of KAN-Adagrad and DBN-Adagrad models for different training data sets.Training Data (Days)KAN-Adagrad MAE (TECU)DBN-Adagrad MAE (TECU)3010.210010.8037609.884310.0219 100 3.5215 4.0216 Mean absolute error (MAE) of KAN-Adagrad and DBN-Adagrad models for different training data sets. As the dataset is divided chronologically, the data of the observed solar flare event will not overlap across the training and testing 80/20 boundary, but the data of other solar flares that may have occurred on the training dates may create an impact on the prediction. In this study, the prediction horizon was explicitly defined as 24 h ahead, corresponding to one full diurnal cycle of ionospheric variation. This rolling window approach provides predictions aligned with the practical monitoring needs, as daily forecasts are most useful for space weather monitoring, providing insights for solar flare-driven ionospheric variability. The forecasting setup was designed to operate in a forward-prediction mode, not in a retrospective fitting. 72 h were used as a training sequence to generate forecasts for the subsequent 24 h. This makes sure that the models were evaluated under realistic temporal constraints, and the chronological 80/20 train-test split preserved event order, simulating operation deployment conditions. The models were assessed only on unseen data to reflect true forecasting conditions. Model validation using PSIS-LFO-CV The predictive performance of the models was assessed using Leave-Future-out Cross Validation (LFO-CV) methodology, as shown in previous studies32, and to tackle possible over-fitting and time-series data splitting concerns. This approach is based on the assumption that only data from the past can be used to forecast future data. This is ideal for use with ionospheric TEC data, for which the time order needs to be maintained. The PSIS-LFO-CV (Pareto Smoothed Importance Sampling) validation was performed for two events: X1.21 SF (CPN station) and X1.86 SF (HYDE station), X.1.21 SF (CPN) as shown in Fig. 2 and the values for both events are shown in Tables 4 and 5, since the TEC prediction results and model behaviour for the remaining events and stations were similar. As in the X1.21 SF event at CPN station, the Pareto-k diagnostic plot shows that the majority of the prediction points are within the acceptable range (k < 0.5) with the occasional refit points at h = 19 and h = 22, demonstrating good predictive behaviour for both models. The performance metrics of PSIS-LFO-CV, such as Total Expected Log Predictive Density (ELPD), Mean Pointwise ELPD and number of refits for all the solar flare events and stations studied are summarised in Table 6. These results suggest that both KAN-Adagrad and DBN-Adagrad models perform well under solar flare conditions without overfitting33. Fig. 2PSIS-LFO-CV results for the X1.21 solar flare event at the CPN station: (a) comparison of observed and predicted TEC values, (b) pointwise ELPD contributions, and (c) Pareto-k diagnostics. PSIS-LFO-CV results for the X1.21 solar flare event at the CPN station: (a) comparison of observed and predicted TEC values, (b) pointwise ELPD contributions, and (c) Pareto-k diagnostics. Table 4PSIS-LFO-CV Performance Summary for X1.21 Solar Flare Event at CPN Station.MetricKAN-AdagradDBN-AdagradTotal ELPD-79.267-61.018Mean Pointwise ELPD-9.908-7.627No. of Refits22Training Hours(L)1616 Table 4PSIS-LFO-CV Performance Summary for X1.21 Solar Flare Event at CPN Station.MetricKAN-AdagradDBN-AdagradTotal ELPD-79.267-61.018Mean Pointwise ELPD-9.908-7.627No. of Refits22Training Hours(L)1616 PSIS-LFO-CV Performance Summary for X1.21 Solar Flare Event at CPN Station. Table 5PSIS-LFO-CV Performance Summary for X1.86 Solar Flare Event at HYDE Station.MetricKAN-AdagradDBN-AdagradTotal ELPD-38.661-44.9312Mean Pointwise ELPD-4.833-8.560No. of Refits22Training Hours(L)1616 Table 5PSIS-LFO-CV Performance Summary for X1.86 Solar Flare Event at HYDE Station.MetricKAN-AdagradDBN-AdagradTotal ELPD-38.661-44.9312Mean Pointwise ELPD-4.833-8.560No. of Refits22Training Hours(L)1616 PSIS-LFO-CV Performance Summary for X1.86 Solar Flare Event at HYDE Station. Table 6Comparison of PSIS-LFO-CV validation results for KAN-Adagrad and DBN-adagrad models during solar flare and quiet-day events.EventStationTotal ELPD (KAN-Adagrad)Total ELPD (DBN-Adagrad)Mean ELPD (KAN-Adagrad)Mean ELPD (DBN-Adagrad)No. of Refits (KAN-Adagrad)No. of Refits (DBN-Adagrad)Training hours (L)X1.86(26-Oct-2024)HYDE-38.661-44.9312-4.833-8.5602216X1.86(26-Oct-2024)CPN-40.639-38.987-5.080-4.8732216X1.21(13-May-2025)HYDE-75.922-69.514-9.490-8.6892216X1.21(13-May-2025)CPN-79.267-61.018-9.908-7.6272216X2.7(14-May-2025)HYDE-37.300-35.181-4.663-4.3982216X2.7(14-May-2025)CPN-42.284-41.506-5.286-5.1882216X1.1(25-May-2025)HYDE-60.291-62.664-7.536-7.8332216X1.1(25-May-2025)CPN-126.262-124.724-15.783-15.5912216Quiet Day (28-Apr-2025)HYDE-47.948-47.729-5.993-5.9662216Quiet Day (28-Apr-2025)CPN-60.144-61.517-7.518-7.6902216 Table 6Comparison of PSIS-LFO-CV validation results for KAN-Adagrad and DBN-adagrad models during solar flare and quiet-day events.EventStationTotal ELPD (KAN-Adagrad)Total ELPD (DBN-Adagrad)Mean ELPD (KAN-Adagrad)Mean ELPD (DBN-Adagrad)No. of Refits (KAN-Adagrad)No. of Refits (DBN-Adagrad)Training hours (L)X1.86(26-Oct-2024)HYDE-38.661-44.9312-4.833-8.5602216X1.86(26-Oct-2024)CPN-40.639-38.987-5.080-4.8732216X1.21(13-May-2025)HYDE-75.922-69.514-9.490-8.6892216X1.21(13-May-2025)CPN-79.267-61.018-9.908-7.6272216X2.7(14-May-2025)HYDE-37.300-35.181-4.663-4.3982216X2.7(14-May-2025)CPN-42.284-41.506-5.286-5.1882216X1.1(25-May-2025)HYDE-60.291-62.664-7.536-7.8332216X1.1(25-May-2025)CPN-126.262-124.724-15.783-15.5912216Quiet Day (28-Apr-2025)HYDE-47.948-47.729-5.993-5.9662216Quiet Day (28-Apr-2025)CPN-60.144-61.517-7.518-7.6902216 Comparison of PSIS-LFO-CV validation results for KAN-Adagrad and DBN-adagrad models during solar flare and quiet-day events. X1.86 (26-Oct-2024) X1.86 (26-Oct-2024) X1.21 (13-May-2025) X1.21 (13-May-2025) X2.7 (14-May-2025) X2.7 (14-May-2025) X1.1 (25-May-2025) X1.1 (25-May-2025) Results and discussion TEC prediction during X1.86 SF on October 26th, 2024 The X1.86 SF that happened on October 26, 2024, originated from the Sun’s active region AR 3873, which was located on the southeastern limb of the Sun’s visible disk. The event began at approximately 05:57 UTC, reached its peak at 07:17 UTC, and ended around 07:56 UTC. Due to its X-class magnitude and Earth-facing side of the sun, the flare caused a substantial increase in ionization level and noticeable anomalies in ionospheric TEC. TEC, solar and geomagnetic variations during X1.86 SF Figure 3 shows the TEC values over HYDE and CPN stations along with the solar and geomagnetic indices, including F10.7 solar radio flux, Ap index, SSN, and Kp index for the period from 28th September 2024 to 28th October 2024. The TEC values at both HYDE and CPN stations show regular diurnal variations throughout the observation period. A significant geomagnetic disturbance was observed around 10th–11th October 2024, where the Ap index reached approximately 294 nT and the Kp index exceeded 8, indicating a severe geomagnetic storm event prior to the X1.86 SF on 26th October 2024. The F10.7 solar radio flux values ranged from 184 to 322 sfu, confirming high solar activity during Solar Cycle 25. The SSN values ranged from 135 to 225, further indicating the active solar conditions. On 26th October 2024, the X1.86 SF event caused noticeable variations in the TEC values at both HYDE and CPN stations, associated with the sudden enhancement in solar radiation during the solar flare. Fig. 3Solar and Geomagnetic Indices for studying the X1.86 SF event from 28th September 2024 to 28th October 2024. Solar and Geomagnetic Indices for studying the X1.86 SF event from 28th September 2024 to 28th October 2024. TEC prediction comparison during X1.86 SF at HYDE station Figure 4 compares the True TEC values (shortly, True TEC in all figures) and the TEC values predicted by the KAN algorithm with Adagrad optimiser or KAN-Adagrad model (shortly, KAN-Adagrad in all figures), the KAN algorithm without optimiser or KAN model (shortly, KAN in all figures), DBN with Adagrad (shortly, DBN-Adagrad in all figures) and the IRI-2020 model (shortly, IRI-2020 in all figures) at Hyderabad (HYDE) station, India. The X-axis represents the dates during 23rd -28th October 2024 (3 days before and 2 days after the solar flare event) along with the universal time in Hrs [UTC]. The Y-axis represents the TEC values in units: TECU. Figure 4 consists of five lines such as (a) a blue solid line representing the True TEC values observed by GNSS receiver at HYDE station, (b) a red dotted line denoting the TEC value predicted by KAN model, (c) a yellow dashed line representing the TEC value predicted by KAN-Adagrad model, and (d) a black dashed line denoting the TEC values predicted by the DBN-Adagrad model, and (e) a purple dashed line denoting the TEC predicted by IRI-2020 model. The vertical dashed lines in Fig. 4 indicate the day on which the solar flare occurred. The descriptions of the variables, lines and legends used in Fig. 4 were also employed in Figs. 6, 9, 11, 14, 16, 19, 21, 23 and 25 in this study work, except for different stations, SF intensities, and SF periods. Fig. 4Comparison of the True TEC and the TEC values predicted by KAN, KAN-Adagrad and IRI-2020 models during 23rd -28th October 2024 at HYDE station. Comparison of the True TEC and the TEC values predicted by KAN, KAN-Adagrad and IRI-2020 models during 23rd -28th October 2024 at HYDE station. Fig. 5Comparison between TEC values during X1.86 SF from 23rd to 28th October 2024 as contour plots at HYDE station. Comparison between TEC values during X1.86 SF from 23rd to 28th October 2024 as contour plots at HYDE station. Fig. 6Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during 23rd-28th October 2024 at the CPN station. Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during 23rd-28th October 2024 at the CPN station. Fig. 7Comparison between the TEC values during X1.86 SF from 23rd to 28th October 2024 through contour plots at the CPN station. Comparison between the TEC values during X1.86 SF from 23rd to 28th October 2024 through contour plots at the CPN station. Fig. 8Solar and geomagnetic indices during X1.21 SF (the studied second SF event) from 15th April 2025 to 16th May 2025. Solar and geomagnetic indices during X1.21 SF (the studied second SF event) from 15th April 2025 to 16th May 2025. Fig. 9Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, and IRI-2020 models during 10th -15th May 2025 at HYDE station. Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, and IRI-2020 models during 10th -15th May 2025 at HYDE station. Fig. 10Comparison between TEC values during X1.21 SF from 10th -15th May 2025 as contour plots at HYDE station. Comparison between TEC values during X1.21 SF from 10th -15th May 2025 as contour plots at HYDE station. Fig. 11Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during 10th -15th May 2025 at CPN station. Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during 10th -15th May 2025 at CPN station. Fig. 12Comparison between the TEC values during X1.21 SF from 10th May to 15th May 2025 as contour plots at the CPN station. Comparison between the TEC values during X1.21 SF from 10th May to 15th May 2025 as contour plots at the CPN station. Fig. 13Solar and geomagnetic indices during X2.7 SF event from 16th April 2025 to 16th May 2025. Solar and geomagnetic indices during X2.7 SF event from 16th April 2025 to 16th May 2025. Fig. 14Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, IRI-2020 models during 11th May 2025 and 16th May 2025 at HYDE station. Comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, IRI-2020 models during 11th May 2025 and 16th May 2025 at HYDE station. Figure 4 shows the TEC variations during the X1.86 SF, similar to our first SF event. This SF event occurred on 26th October 2024 from 05:57 UTC to 07:56 UTC, where it reached its peak intensity at 07:17 UTC. The KAN model was input with four parameters, such as (a) SSNs ranging from 122 to 246, (b) Kp values ranging from 0 to 47, (c) Ap values ranging from 0 to 39 and (d) F10.7 ranging from 183 to 253. Around 08:00 UTC, a significant rise in SSN and F10.7 was observed due to the impact of a solar flare, as shown in Fig. 3. On 23rd October, 2024, the IRI-2020–predicted TEC values under-estimated the True TEC values and DBN-Adagrad predicted TEC values over-estimated the True TEC values when compared to other models. Around 08:00 UTC on 24th October 2024, there was a fluctuation in the True TEC value. This event was closely predicted by the DBN-Adagrad model and slightly predicted by the KAN-Adagrad model when compared with other models. On 25th October 2024, the KAN-Adagrad model closely predicted the True TEC values and the DBN-Adagrad model also slightly predicted well, as compared to the KAN (without the optimiser) model and the IRI-2020 model. The IRI-2020 model overestimated the True TEC values continuously when compared to the KAN (without optimiser), DBN-Adagrad and the KAN-Adagrad model on 26th -28th October 2024. On 26th October 2024, a noticeable fluctuation in the True TEC values was observed around 15:00 UTC. Since an X1.86 solar flare occurred on the same day and enhanced solar activity conditions were indicated by SSN values greater than 60 and F10.7 values exceeding 120 sfu, this TEC enhancement may be associated with increased solar activity. Although all models overestimated the fluctuation, the KAN-Adagrad model provided a closer prediction to the observed TEC values during the peak period. On the same day, there was a noticeable fluctuation at the end of the day that was under-predicted by all models. On 28th October 2024, all prediction models overestimated the True TEC values. Overall, for almost all six days, the KAN-Adagrad model and DBN-Adagrad model performed well as compared to other models. The KAN model performed moderately as compared to the IRI-2020 model, whereas the IRI-2020 model under-estimated and over-estimated the True TEC values consistently. Figure 5 shows the comparison of True TEC, DBN-Adagrad, KAN, KAN-Adagrad and IRI-2020 predicted TEC values through contour plots that cover both normal and SF days and illustrate their maximum and minimum values on each contour plot. In Fig. 5, the contour level represents the TEC values in units: TECU, and the colour shading illustrates temporal TEC variations where red and blue colours represent higher and lower TEC values, respectively. The descriptions of colour shading used in Fig. 5 were also employed in Figs. 7, 10, 12, 15, 17, 20, 22, 24 and 26 in this study work, except for different stations, SF intensities, and SF periods. From Fig. 5, it can be seen that the DBN-Adagrad and KAN-Adagrad models follow the True TEC pattern closely on most days. However, on 28th October 2024, a slight overestimation is observed. The IRI-2020 model shows a smoother pattern and does not capture the variations clearly during the solar flare period. Fig. 15Comparison between TEC values of X2.7 SF during 11th May 2025 and 16th May 2025 as contour plots at HYDE station. Comparison between TEC values of X2.7 SF during 11th May 2025 and 16th May 2025 as contour plots at HYDE station. Fig. 16Comparison of the true TEC and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during X2.7 SF from 11th to 16th May 2025 at CPN station. Comparison of the true TEC and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during X2.7 SF from 11th to 16th May 2025 at CPN station. Fig. 17Comparison between TEC values during X2.7 SF from 11th May 2025 to 16th May 2025 as contour plots at CPN station. Comparison between TEC values during X2.7 SF from 11th May 2025 to 16th May 2025 as contour plots at CPN station. Fig. 18Solar and geomagnetic indices during X1.1 SF during 22nd May 2025 and 27th May 2025. Solar and geomagnetic indices during X1.1 SF during 22nd May 2025 and 27th May 2025. Fig. 19Comparison of the true TEC and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during X1.1 SF from 22nd to 27th May 2025 at HYDE station. Comparison of the true TEC and the TEC values predicted by KAN, KAN-Adagrad, and IRI-2020 models during X1.1 SF from 22nd to 27th May 2025 at HYDE station. Fig. 20Comparison between TEC values during X1.1 SF during 22nd May 2025 and 27th May 2025 as contour plots at HYDE station. Comparison between TEC values during X1.1 SF during 22nd May 2025 and 27th May 2025 as contour plots at HYDE station. Fig. 21Comparison of True TEC and the TEC values (KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020) of X1.1 SF event during 22nd -27th May 2025 at CPN station. Comparison of True TEC and the TEC values (KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020) of X1.1 SF event during 22nd -27th May 2025 at CPN station. Fig. 22Comparison between TEC values during X1.1 SF from 22nd May 2025 to 27th May 2025 as contour plots at Chumphon station. Comparison between TEC values during X1.1 SF from 22nd May 2025 to 27th May 2025 as contour plots at Chumphon station. Fig. 23Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during 25th -30th April 2025 at HYDE station. Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during 25th -30th April 2025 at HYDE station. Fig. 24Comparison between Quiet days TEC values during 25th Apr 2025 and 30th Apr 2025 as contour plots at HYDE station. Comparison between Quiet days TEC values during 25th Apr 2025 and 30th Apr 2025 as contour plots at HYDE station. Fig. 25Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during 25th -30th April 2025 at CPN station. Comparison of the True TEC values and the TEC values predicted with KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during 25th -30th April 2025 at CPN station. Fig. 26Comparison between Quiet days TEC values during 25th Apr 2025 and 30th Apr 2025 as contour plots at the CPN station. Comparison between Quiet days TEC values during 25th Apr 2025 and 30th Apr 2025 as contour plots at the CPN station. Table 7 shows the comparison between three statistical parameters, such as Willmott’s Index of Agreement (WIA), Kling-Gupta Efficiency (KGE) and Spearman Correlation Coefficient (SCC) for the four prediction models: KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models over HYDE Station during 23rd -28th October 2024. The table is structured with the first column representing the date of observations, followed by the columns of each statistical parameter corresponding to the models. The variables and statistical parameters listed in Table 7 are consistently used across Tables 7, 8, 9, 10, 11, 12, 13, 14, 15 and 16 in this study, except for different stations, SF intensities, and SF periods. In Table 7, on the day of the SF occurrence (26th October 2024), the KAN-Adagrad and DBN-Adagrad models performed well, achieving higher WIA, KGE, and SCC values than both the KAN and IRI-2020 models. This trend has still been followed throughout the observation period, including the pre-event days (23rd -25th October 2024) and the post-event days (27th -28th October 2024). On October 28, 2024, the KAN-Adagrad showed better performance with higher WIA, KGE, and SCC values [0.8655,0.5563, 0.8183] than other models. The averages of WIA, KGE, and SCC for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9579, 0.7945, 0.9428], [0.9677, 0.8586, 0.9534], [0.9650,0.8573,0.9480] and [0.9156, 0.7228, 0.9172], respectively. This discovery identified that KAN-Adagrad and DBN-Adagrad models gave higher efficiency and correlation, but the KAN-Adagrad model showed slightly better performance than other models, because it had higher averages of WIA, KGE, and SCC. Table 7Performance evaluation during X1.86 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad23-10-20240.99090.99380.95500.99680.90400.92510.86320.97650.97480.98260.92960.986124-10-20240.97940.99180.94690.99670.81200.87380.83070.94490.98610.99220.92250.986125-10-20240.98760.99340.96230.99010.87290.96500.89620.89290.97740.98960.94000.983526-10-20240.95850.96850.90050.96910.79760.89120.75170.87380.94090.94870.86350.950427-10-20240.98650.99340.92110.99370.90300.93990.77210.95540.97650.98870.90870.984328-10-20240.84420.86550.80780.84380.47730.55630.22310.50020.80090.81830.93910.7974 Average 0.9579 0.9677 0.9156 0.9651 0.7945 0.8586 0.7228 0.8573 0.9428 0.9533 0.9172 0.9480 Table 7Performance evaluation during X1.86 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad23-10-20240.99090.99380.95500.99680.90400.92510.86320.97650.97480.98260.92960.986124-10-20240.97940.99180.94690.99670.81200.87380.83070.94490.98610.99220.92250.986125-10-20240.98760.99340.96230.99010.87290.96500.89620.89290.97740.98960.94000.983526-10-20240.95850.96850.90050.96910.79760.89120.75170.87380.94090.94870.86350.950427-10-20240.98650.99340.92110.99370.90300.93990.77210.95540.97650.98870.90870.984328-10-20240.84420.86550.80780.84380.47730.55630.22310.50020.80090.81830.93910.7974 Average 0.9579 0.9677 0.9156 0.9651 0.7945 0.8586 0.7228 0.8573 0.9428 0.9533 0.9172 0.9480 Performance evaluation during X1.86 SF (HYDE station). TEC prediction comparison during X1.86 SF at CPN station Figure 6 shows the comparison of the true TEC values and the TEC values predicted by KAN, the KAN-Adagrad, DBN-Adagrad and the IRI-2020 models. On 24th October 2024, there was a sudden increase in the input parameters Kp (Kp > 5), SSN (SSN > 60), Ap (Ap > 30) and F10.7 (F10.7 > 120 sfu), which indicated a disturbed day due to the impact of X1.86 SF. From Fig. 6, we observed that there was a rapid fluctuation in the TEC values from 06:00 UTC to 21:00 UTC, where it reached its peak TEC value at 15:00 UTC. This event was under-estimated by the IRI-2020 and KAN-Adagrad models as compared to the KAN model. On 25th October 2024, the IRI-2020 model under-predicted when compared to KAN, KAN-Adagrad and DBN-Adagrad models. On 26th October 2024, during the occurrence of the X1.86 solar flare event, a TEC enhancement was observed around 12:00 UTC. This TEC enhancement may be associated with the occurrence of the X1.86 solar flare, together with increased solar activity indicated by F10.7 > 120 sfu and SSN > 60. This enhancement was under-estimated by all four models, while the IRI-2020 model showed poor performance. On 27th October 2024, the KAN-Adagrad model and DBN-Adagrad performed well as compared to the KAN model (as an over-estimated case) and the IRI-2020 model (as an under-estimated case). On 28th October 2024, with TEC peaks around 06:00 UTC to 13:00 UTC, all four prediction models under-estimated the True TEC values due to the rapid fluctuations in the input parameters (Kp, SSN, Ap, and F10.7). It can be summarised for X1.86 SF (our first SF event) during 23rd-28th October 2024 at CPN station that the KAN-Adagrad and DBN-Adagrad models performed better than other models, like HYDE station, in Subsection 3.1. The IRI-2020 model continuously under-estimated for all six days of this X1.86 SF, whereas the KAN-Adagrad and DBN-Adagrad models provided closer agreement with the observed TEC values. Meanwhile, the KAN model showed moderate performance compared with the KAN-Adagrad and DBN-Adagrad models. Figure 7 shows the comparison of True TEC, DBN-Adagrad, KAN, KAN-Adagrad and IRI-2020 predicted TEC values through contour plots that cover both normal and SF days. From the contour plots during 23rd -28th October 2024 at CPN station, the KAN-Adagrad’s and DBN-Adagrad’s contour plots resembled the contour plot of True TEC like HYDE station, indicating its better performance. Table 8Performance evaluation during X1.86 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad23-10-20240.86430.93950.45880.95270.71080.7786-0.07850.87660.82870.95650.31480.942624-10-20240.90660.93560.48440.96620.75210.7486-0.11840.79890.87040.94350.27910.958325-10-20240.91380.97270.51400.97970.78020.82280.04520.86600.89740.97300.44200.970426-10-20240.91280.93610.48920.95770.70420.8139-0.00820.87980.88000.90520.29910.922627-10-20240.91530.93390.55370.94200.83370.85190.25490.88520.86000.88090.45560.928728-10-20240.91200.94750.45990.92110.78130.71510.01340.64200.90260.96870.35670.9522 Average 0.9041 0.9442 0.4933 0.9533 0.7604 0.7885 0.0181 0.8248 0.8732 0.9380 0.3579 0.9458 Table 8Performance evaluation during X1.86 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad23-10-20240.86430.93950.45880.95270.71080.7786-0.07850.87660.82870.95650.31480.942624-10-20240.90660.93560.48440.96620.75210.7486-0.11840.79890.87040.94350.27910.958325-10-20240.91380.97270.51400.97970.78020.82280.04520.86600.89740.97300.44200.970426-10-20240.91280.93610.48920.95770.70420.8139-0.00820.87980.88000.90520.29910.922627-10-20240.91530.93390.55370.94200.83370.85190.25490.88520.86000.88090.45560.928728-10-20240.91200.94750.45990.92110.78130.71510.01340.64200.90260.96870.35670.9522 Average 0.9041 0.9442 0.4933 0.9533 0.7604 0.7885 0.0181 0.8248 0.8732 0.9380 0.3579 0.9458 Performance evaluation during X1.86 SF (CPN station). Table 8 shows the comparison between the statistical performance metrics, such as WIA, KGE, and SCC, for the four prediction models from 23rd October 2024 to 28th October 2024 of the X1.86 SF event at the CPN station. Table 8 presents the same details as Table 7, with a different station. On the day of the SF existence (26th October 2024), the DBN-Adagrad model performed well with higher WIA, KGE, and SCC values as compared to those of other models. On 23rd October 2024 (the first study day), the WIA, KGE and SCC values for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.8643, 0.7108, 0.8287], [0.9395, 0.7786, 0.9565], [0.9527, 0.8766, 0.9426] and [0.4588, -0.0785, 0.3148], respectively. These statistical parameters show that the performance of the DBN-Adagrad model is slightly higher than both the KAN, DBN-Adagrad and IRI-2020 models, and this trend was followed throughout the observation days. The averages of WIA, KGE, and SCC for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models are [0.9041, 0.7604, 0.8732], [0.9442, 0.7885, 0.9380], [0.9533, 0.8248, 0.9458] and [0.4933, 0.0181, 0.3579], respectively. These performance indices indicated that the DBN-Adagrad model at the CPN station also behaved as the best model among all four models. TEC prediction during X1.21 SF on 13th May 2025 The X1.86-class solar flare that happened on May 13, 2025, initiated from the Sun’s active region AR 4086, which was located on the western edge of the solar disk. The event began at approximately 15:25 UTC, reached its peak at 15:38 UTC, and ended around 15:44 UTC. Due to its X-class magnitude, the flare caused a moderate increase in ionization level and anomalies in ionospheric TEC. TEC, Solar and Geomagnetic variations during X1.21 SF Figure 8 shows the TEC values over HYDE and CPN stations along with the solar and geomagnetic indices, including F10.7 solar radio flux, Ap index, SSN, and Kp index from 15th April 2025 to 16th May 2025. This observation period covers the X1.21 SF event on 13th May 2025. The F10.7 solar radio flux ranged from 112 to 178 sfu, and the SSN values varied between 53 and 212, indicating variable solar conditions. A geomagnetic disturbance was observed around 16th–17th April 2025, where the Ap index reached approximately 177 nT and the Kp index exceeded 7, indicating a strong geomagnetic storm. Strong enhancements of TEC values were observed at both stations during this disturbed period. On 13th May 2025, the X1.21 SF event produced moderate TEC enhancements at both HYDE and CPN stations, associated with the enhancement in solar radiation during the solar flare. TEC prediction comparison during X1.21 SF at HYDE station Figure 9 shows the comparison of True TEC values and the TEC values predicted by KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models of X1.21 SF on 13th May 2025 at HYDE station. The X-axis denotes the dates from 10th May 2025 to 15th May 2025 (3 days before and 2 days after the solar flare occurrence) along with the universal time in Hrs [UTC]. The Y-axis represents the TEC values in units: TECU. On 10th May 2025 (the first study day of X1.21 SF), there were rapid fluctuations in the input parameters (Kp, SSN, and F10.7) as shown in Fig. 8, indicating a moderately disturbed day. In spite of the rapid fluctuations, all prediction models performed well, where both KAN and KAN-Adagrad models slightly overestimated, but DBN-Adagrad slightly predicted well as compared to other models and the IRI-2020 model slightly underestimated. On 11th May 2025, a slight increase in the True TEC values was observed during the peak period. This TEC enhancement may be associated with solar activity, indicated by F10.7 values exceeding 120 sfu (F10.7 > 120 sfu). During this period, all models under-estimated the True TEC values, whereas the IRI-2020 model showed greater underestimation as compared to other models. On 12th May 2025, with a TEC peak around 09:00 UTC, the KAN-Adagrad model and DBN-Adagrad model predicted the True TEC values more closely than the other models. On 13th -14th May 2025, with no rapid TEC variations, the KAN, DBN-Adagrad and KAN-Adagrad models performed better than the IRI-2020 model with TEC underestimation. On 15th May 2025, with a TEC peak around 10:00 UTC, the KAN-Adagrad model predicted the True TEC values most closely among all models. Overall, the KAN-Adagrad and DBN-Adagrad models generally performed better than the KAN and IRI-2020 models throughout the study period. The KAN model showed moderate performance, whereas the IRI-2020 model under-predicted the TEC values in most situations. These results indicate that the KAN-Adagrad and DBN-Adagrad models provided closer agreement with the observed TEC values and improved prediction capability compared to the IRI-2020 model during the X1.21 solar flare event. Table 9 illustrates the comparison of statistical parameters such as WIA, KGE and SCC of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models for the data obtained from the HYDE station. The details in Table 9 are the same as those in Table 7, except for the occurrence date and severity of the SF event. On 10th May 2025 (the first study day), the WIA, KGE and SCC values for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.986, 0.8872, 0.9174], [0.9774, 0.8517, 0.9661], [0.9685,0.8090,0.9504] and [0.9370, 0.7050, 0.9591], respectively. These results indicated that the KAN model provided higher predictive efficiency, and the KAN-Adagrad model gave higher correlation with the True TEC on that day. Likewise, on the day of SF existence (13th May 2025), the KAN model performed surprisingly better than other models, in which it showed higher performance metrics than both KAN-Adagrad, DBN-Adagrad and IRI-2020 models. On the other hand, on 15th May 2025, KAN-Adagrad has predicted better than other models. Particularly, the averages of the statistical metrics (WIA, KGE, and SCC) for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models owned the highest values interchangeably, i.e., [0.9777, 0.8845, 0.9387], [0.97, 0.8897, 0.9334], [0.9736,0.8817,0.9390] and [0.9423, 0.7256, 0.9469], respectively. Thus, the KAN model predicted overall well. The results showed that the KAN model had better efficiency and correlation to predict the TEC values than other predictive models (KAN-Adagrad, DBN-Adagrad and IRI-2020) in this SF event. Figure 10 demonstrates the comparison between the True TEC and the TEC predicted using the KAN-Adagrad, KAN, DBN-Adagrad and IRI-2020 models using contour plots for the dates from 10th May 2025 to 15th May 2025, covering both normal and SF days at HYDE station. The TEC contour plot obtained from KAN-Adagrad and DBN-Adagrad showed a slight difference from the True TEC contour plot on 11th May 2025, whereas those on other days matched closely with the True TEC contour plot. Table 9Performance evaluation during X1.21 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad10-05-20250.98600.97740.93700.96850.88720.85170.70500.80900.91740.96610.95910.950411-05-20250.95540.94790.92570.96970.84270.89080.66710.91430.86610.85130.93380.900912-05-20250.97950.97400.96250.98360.92450.91810.75650.92470.95300.94350.96520.959113-05-20250.97680.95030.94730.94200.90560.83390.74310.87140.95740.88000.92350.862614-05-20250.98920.98060.95470.98900.91360.88610.76280.90760.96520.96960.96000.976515-05-20250.97920.98980.92660.98890.83350.95760.71880.86320.97300.98960.94000.9843 Average 0.9777 0.9700 0.9423 0.9736 0.8845 0.8897 0.7255 0.8817 0.9387 0.9333 0.9469 0.9390 Table 9Performance evaluation during X1.21 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad10-05-20250.98600.97740.93700.96850.88720.85170.70500.80900.91740.96610.95910.950411-05-20250.95540.94790.92570.96970.84270.89080.66710.91430.86610.85130.93380.900912-05-20250.97950.97400.96250.98360.92450.91810.75650.92470.95300.94350.96520.959113-05-20250.97680.95030.94730.94200.90560.83390.74310.87140.95740.88000.92350.862614-05-20250.98920.98060.95470.98900.91360.88610.76280.90760.96520.96960.96000.976515-05-20250.97920.98980.92660.98890.83350.95760.71880.86320.97300.98960.94000.9843 Average 0.9777 0.9700 0.9423 0.9736 0.8845 0.8897 0.7255 0.8817 0.9387 0.9333 0.9469 0.9390 Performance evaluation during X1.21 SF (HYDE station). TEC prediction comparison during X1.21 SF at CPN station Figure 11 shows the comparison of the true TEC values and the TEC values predicted by KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models over the CPN station. On 10th May 2025, there was a notable fluctuation in the input parameters (Kp, SSN, and F10.7) as well as TEC values. On this day, all four prediction models performed poorly, where the IRI-2020 model under-predicted completely. Meanwhile, the KAN, DBN-Adagrad and the KAN-Adagrad models performed better, but they overestimated during the TEC peak period. On 11th May 2025, a slight increase in the True TEC values was observed during the peak period. This TEC enhancement may be associated with increased solar activity, indicated by F10.7 values exceeding 120 sfu (F10.7 > 120 sfu). This peak TEC value around 12:00 UTC was closely predicted by the KAN model. Meanwhile, the KAN-Adagrad and DBN-Adagrad models performed moderately and the IRI-2020 model provided poor performance. On 12th May 2025, there were random rises and dips in the input parameters (Kp and Ap indices) as depicted in Fig. 8; hence, both the KAN models (without and with Adagrad optimiser) and DBN-Adagrad overestimated. On 13th May 2025, there was a sudden fluctuation from 09:00 to 14:00 UTC. On 13th–14th May 2025, slight fluctuations were observed in the TEC values, which may be associated with the occurrence of the X1.21 solar flare event. During these True TEC fluctuations, the IRI-2020 model gave very poor performance as compared to other models (KAN, KAN-Adagrad and DBN-Adagrad models). Meanwhile, both KAN with Adagrad optimiser, KAN without optimiser and DBN-Adagrad (DBN-Adagrad model under-estimate during 14th May, 2025) models were over-estimated, but the KAN with Adagrad optimiser and DBN-Adagrad predicted closely on these two days as compared to the KAN model and IRI-2020 model, especially on 14th May 2025. On 15th May 2025, there were rapid fluctuations in the input parameters (Kp and Ap indices), hence, there were noticeable and random fluctuations in TEC values. On this day, the KAN-Adagrad and DBN-Adagrad models predicted closely to the True TEC as compared to the other two models (KAN and IRI-2020 models). Overall, we can infer that the IRI-2020 model performed poorly with under-prediction for all days of this SF event, and the KAN model predicted closely to the True TEC as compared to other models in certain situations. Both KAN models and DBN-Adagrad performed better than the IRI-2020 model, in which the KAN-Adagrad model generally provided closer agreement with the observed TEC values than the KAN and DBN-Adagrad models during this event. Figure 12 demonstrates the comparison between the True TEC and the TEC predicted using the KAN-Adagrad, KAN, DBN-Adagrad and IRI-2020 models using contour plots for the dates from 10th May 2025 to 15th May 2025, covering both normal and SF days at the CPN station. The TEC contour plot obtained from KAN-Adagrad and DBN-Adagrad showed a slight difference from the True TEC contour plot on 12th May 2025, whereas those on other days matched closely with the True TEC contour plot. Table 10Performance evaluation during X1.21 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad10-05-20250.94150.94090.57530.92450.83760.82070.10950.79660.93650.89130.58260.863511-05-20250.90670.91680.52340.91250.81910.8101-0.06920.70300.80170.80000.43900.879112-05-20250.90390.88780.54700.96050.77130.76730.09230.87530.84090.81650.60570.924313-05-20250.94640.95580.56250.96000.69530.75090.10890.82450.94610.97740.58780.976514-05-20250.90320.93870.56350.93140.77250.85590.06700.84060.70430.76780.57830.762615-05-20250.86280.91060.54490.94580.74110.80590.02010.89400.64170.86000.52520.9296 Average 0.9107 0.9251 0.5528 0.9391 0.7728 0.8018 0.0548 0.8223 0.8119 0.8522 0.5531 0.8893 Table 10Performance evaluation during X1.21 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad10-05-20250.94150.94090.57530.92450.83760.82070.10950.79660.93650.89130.58260.863511-05-20250.90670.91680.52340.91250.81910.8101-0.06920.70300.80170.80000.43900.879112-05-20250.90390.88780.54700.96050.77130.76730.09230.87530.84090.81650.60570.924313-05-20250.94640.95580.56250.96000.69530.75090.10890.82450.94610.97740.58780.976514-05-20250.90320.93870.56350.93140.77250.85590.06700.84060.70430.76780.57830.762615-05-20250.86280.91060.54490.94580.74110.80590.02010.89400.64170.86000.52520.9296 Average 0.9107 0.9251 0.5528 0.9391 0.7728 0.8018 0.0548 0.8223 0.8119 0.8522 0.5531 0.8893 Performance evaluation during X1.21 SF (CPN station). Table 10 illustrates the comparison of the statistical metrics, such as WIA, KGE, and SCC, among the four prediction models (e.g., KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020) for the observation period from 10th May to 15th May 2025 at the CPN station. The contents of Table 10 also mirrored those of Table 7, except for the different studied SF severity and duration. On the day of SF existence (13th May 2025), the WIA, KGE, and SCC values were [0.9464, 0.6953, 0.9461], [0.9558, 0.7509, 0.9774], [0.9600, 0.8245, 0.9765] and [0.5625, 0.1089, 0.5878] for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models, respectively. These values illustrated clearly that DBN-Adagrad and KAN models provide better performance than other models (KAN-Adagrad and IRI-2020 models). The averages of WIA, KGE, and SCC for KAN, KAN-Adagrad, and IRI-2020 models were [0.9107, 0.7728,0.8119], [0.9251, 0.8018, 0.8522], [0.9391, 0.8223, 0.8993] and [0.5528, 0.0548, 0.5531], respectively. Overall, the performance of the DBN-Adagrad model performed better than the KAN model, KAN-Adagrad and IRI-2020 models because it had better predictive efficiency and correlation. TEC prediction during X2.7 SF on 14th May 2025 The X2.7 solar flare that happened on May 14, 2025, initiated from the Sun’s active region AR 4087, which was located on the northeastern edge of the Sun. The event began at approximately 08:04 UTC, reached its peak at 08:19 UTC, and ended around 08:31 UTC. Due to its X-class magnitude, the flare hit the sunlit side of Earth, causing an increase in ionization level and anomalies in ionospheric TEC over Europe, Asia and the Middle East. TEC, solar and geomagnetic variations during X2.7 SF Figure 13 represents the TEC values over HYDE and CPN stations along with the solar and geomagnetic indices, including F10.7 solar radio flux, Ap index, SSN, and Kp index from 16th April 2025 to 16th May 2025. The F10.7 solar radio flux values ranged from 112 to 176 sfu, and the SSN values varied between 53 and 212, indicating highly active solar conditions during this period. A significant geomagnetic disturbance was observed around 16th–17th April 2025, where the Ap index reached approximately 177 nT and the Kp index exceeded 7, suggesting a strong geomagnetic storm event. This disturbance caused noticeable variations in the TEC values at both HYDE and CPN stations. TEC prediction comparison during X2.7 SF at HYDE station Figure 14 illustrates the comparison between the observed TEC (True TEC) and the TEC values predicted by KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during X2.7 SF on 14th May 2025 at HYDE station. The X-axis represents the period from 11th May 2025 to 16th May 2025 in Hrs [UTC], while the Y-axis denotes the TEC values in units: TECU. From the observation in Fig. 13, a sudden increase in True TEC was observed on 11th May 2025 between 10:00 UTC and 13:00 UTC. The variation of True TEC was found due to the SF effect (SSN > 60 and F10.7 > 120 sfu). On the day of the SF event (14th May 2025), the F10.7 value exceeded the threshold value (> 120 sfu), indicating a moderate solar flare effect. At the same time, the Kp and Ap values were above their threshold levels, suggesting the presence of strong geomagnetic activity during this period, as shown in Fig. 13. On 11th May 2025, both KAN and KAN-Adagrad models performed better than the IRI-2020 model and DBN-Adagrad model; however, they slightly under-predicted the True TEC between 10:00 UTC and 12:00 UTC. Between 12th May 2025 and 15th May 2025, the KAN model under-predicted the True TEC values at the end of the day and over-predicted during the TEC peak hours (from 11:00 UTC to 16:00 UTC). On 13th May 2025, all four prediction models did not perform well, particularly during the TEC peak period and at the end of the day. On the SF event day (14th May 2025), both KAN-Adagrad, DBN-Adagrad and KAN models performed significantly better than the IRI-2020 model, although their slight over-predictions were observed between 05:00 UTC and 08:00 UTC. On 16th May 2025, the KAN-Adagrad and DBN-Adagrad performed better than other models and the KAN model slightly over-predicted the True TEC values during the TEC peak period. Overall, the results indicated that the KAN-Adagrad and DBN-Adagrad consistently provided better performance than other models and the KAN model illustrated moderate performance. Meanwhile, the IRI-2020 model did not perform well, as a result, it generally under-predicted the True TEC values during this X2.7 SF. Figure 15 illustrates the comparison between the True TEC and the TEC predicted using the KAN-Adagrad, KAN, DBN-Adagrad and IRI-2020 models through contour plots of X2.7 SF (the studied third SF event) for the dates from 11th May 2025 to 16th May 2025, covering both normal and SF days. The comparison is based on the maximum and minimum TEC values across the study period. According to the contour plots, the KAN-Adagrad and DBN-Adagrad contour plots resembled the contour plot of True TEC, indicating its better performance. Table 11 depicts the comparison between the statistical performance metrics: WIA, KGE, and SCC for all four prediction models: KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models of X2.7 SF event during 11th May 2025 and 16th May 2025 at HYDE station. The details in Table 11 are the same as in Table 7, except for different SF severity and duration. On the day before the SF existence (13th May 2025), the KAN model predicted better than other models. On 14th May 2025 (the day of the SF occurrence), the performance metrics of WIA, KGE, and SCC for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9892, 0.9136, 0.9652], [0.9806, 0.8861, 0.9696], [0.9869,0.9181,0.9748] and [0.9547, 0.7628, 0.96], respectively. On the day after the SF event (15th May 2025), KAN-Adagrad showed comparable performance with higher WIA, KGE, and SCC than KAN, DBN-Adagrad and IRI-2020 models. The averages of WIA, KGE, and SCC of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were respectively [0.9732, 0.8820, 0.9367], [0.9724, 0.8985, 0.9383], [0.9803,0.9086,0.9460] and [0.9429, 0.7265, 0.9462]. Thus, on the whole, the DBN-Adagrad model also performed well as compared to other models, because it expressed higher efficiency and correlation with the True TEC under this X2.7 SF event. Table 11Performance evaluation during X2.7 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad11-05-20250.93290.94500.92570.97730.85010.86380.66710.91660.81480.87130.93380.922612-05-20250.97950.97540.96250.98290.92450.91990.75650.86480.95300.94350.96520.963513-05-20250.97680.95060.94730.94890.90560.83510.74310.85480.95740.88000.92350.878314-05-20250.98920.98060.95470.98690.91360.88610.76280.91810.96520.96960.96000.974815-05-20250.97920.99090.92660.99120.83350.95240.71880.93660.97300.99040.94000.989616-05-20250.98150.99170.94090.99430.86460.93380.71090.96080.95650.97480.95440.9470 Average 0.9732 0.9724 0.9430 0.9803 0.8820 0.8985 0.7265 0.9086 0.9367 0.9383 0.9462 0.9459 Table 11Performance evaluation during X2.7 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad11-05-20250.93290.94500.92570.97730.85010.86380.66710.91660.81480.87130.93380.922612-05-20250.97950.97540.96250.98290.92450.91990.75650.86480.95300.94350.96520.963513-05-20250.97680.95060.94730.94890.90560.83510.74310.85480.95740.88000.92350.878314-05-20250.98920.98060.95470.98690.91360.88610.76280.91810.96520.96960.96000.974815-05-20250.97920.99090.92660.99120.83350.95240.71880.93660.97300.99040.94000.989616-05-20250.98150.99170.94090.99430.86460.93380.71090.96080.95650.97480.95440.9470 Average 0.9732 0.9724 0.9430 0.9803 0.8820 0.8985 0.7265 0.9086 0.9367 0.9383 0.9462 0.9459 Performance evaluation during X2.7 SF (HYDE station). TEC prediction comparison during X2.7 SF at CPN station Figure 16 shows the comparison of the True TEC and the TEC values predicted by KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models during X2.7 SF at the CPN station. From the observation, a sudden increase in True TEC was noticed between 12:00 UTC and 14:00 UTC on 11th May 2025, which could be attributed to the solar flare activity (SSN > 60 and F10.7 > 120 sfu as shown in Fig. 13). On 11th May 2025, all four prediction models under-forecasted the True TEC values during the TEC peak period; however, both KAN and KAN-Adagrad models performed slightly better than the DBN-Adagrad and IRI-2020 models. On 12th May 2025, the KAN-Adagrad, DBN-Adagrad and KAN models showed better performance with the observed TEC as compared to the IRI-2020 model, although they slightly over-predicted the True TEC values between 12:00 UTC and 15:00 UTC. On 13th May 2025, both KAN models over-predicted the True TEC during the TEC peak hours (10:00 UTC to 12:00 UTC), while the IRI-2020 model showed significant under-prediction and DBN-Adagrad showed slightly over-prediction during the peak period. On the day of the SF existence (14th May 2025), the KAN, DBN-Adagrad and KAN-Adagrad models also provided better performance than the IRI-2020 model, although some fluctuations were slightly over-predicted during certain intervals. On 14th -15th May 2025, noticeable fluctuations were observed in the TEC values, which may be associated with the occurrence of the X2.7 solar flare event together with increased solar activity conditions indicated by SSN and F10.7 values exceeding their threshold levels. During 15th -16th May 2025, the KAN model slightly over-predicted the True TEC values, particularly during the TEC peak period and at the end of the day. Overall, the results suggested that the KAN-Adagrad and DBN-Adagrad models could capture the sudden TEC fluctuations more effectively, while the KAN model illustrated moderate performance. In contrast, the IRI-2020 model consistently under-predicted the True TEC values under this X2.7 SF event. The comparison between the True TEC and the TEC values forecasted by KAN-Adagrad, DBN-Adagrad, Kan and IRI-2020 model using contour plots is depicted in Fig. 17. The contour plots represent TEC values from 11th May 2025 to 16th May 2025, including both quiet and SF days. From the contour plots, the KAN-Adagrad model over-predicted on 13th May 2025, whereas it matched closely with the true TEC contour plot on other days. Table 12 shows the comparison between the performance indices (WIA, KGE, and SCC) of all four prediction models (i.e., KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models) during the X2.7 SF event at CPN station between 11th May 2025 and 16th May 2025. The details in Table 12 are identical to those in Table 7, except for different SF intensity and duration. On 12th May 2025 (two days before SF existence), the DBN-Adagrad model performed better than other models. On the day of SF occurrence (14th May 2025), the WIA, KGE, and SCC values of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9205, 0.8470, 0.7574], [0.9387, 0.8559, 0.7678], [0.9337, 0.8526, 0.7922] and [0.5635, 0.0670, 0.5783], respectively. This value shows that the KAN-Adagrad performs slightly better than the other models on the solar flare day, with higher WIA and KGE values. The averages of WIA, KGE, and SCC of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9165, 0.7549, 0.8312], [0.9318, 0.7974, 0.8652], [0.9477, 0.8444, 0.9101] and [0.5415, 0.0316, 0.5179], respectively. Overall, our study results showed that the DBN-Adagrad model performed well and gave better correlation than the KAN, KAN-Adagrad and IRI-2020 models. Table 12Performance evaluation during X2.7 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad11-05-20250.90390.92180.52340.94880.80160.8171-0.06920.80800.76520.81130.43900.927012-05-20250.90160.88780.54700.93390.80450.76730.09230.81280.82610.81650.60570.883513-05-20250.95250.95580.56250.95710.72350.75090.10890.83190.96000.97740.58780.965214-05-20250.92050.93870.56350.93370.84700.85590.06700.85260.75740.76780.57830.792215-05-20250.88980.91060.54490.92050.78660.80590.02010.78530.77390.86000.52520.918316-05-20250.93070.97570.50750.99230.56640.7874-0.02930.97560.90430.95830.37130.9748 Average 0.9165 0.9318 0.5415 0.9477 0.7549 0.7974 0.0316 0.8444 0.8312 0.8652 0.5179 0.9101 Table 12Performance evaluation during X2.7 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad11-05-20250.90390.92180.52340.94880.80160.8171-0.06920.80800.76520.81130.43900.927012-05-20250.90160.88780.54700.93390.80450.76730.09230.81280.82610.81650.60570.883513-05-20250.95250.95580.56250.95710.72350.75090.10890.83190.96000.97740.58780.965214-05-20250.92050.93870.56350.93370.84700.85590.06700.85260.75740.76780.57830.792215-05-20250.88980.91060.54490.92050.78660.80590.02010.78530.77390.86000.52520.918316-05-20250.93070.97570.50750.99230.56640.7874-0.02930.97560.90430.95830.37130.9748 Average 0.9165 0.9318 0.5415 0.9477 0.7549 0.7974 0.0316 0.8444 0.8312 0.8652 0.5179 0.9101 Performance evaluation during X2.7 SF (CPN station). TEC prediction during X1.1 SF on 25th May 2025 at HYDE station The X1.1 solar flare that happened on May 25, 2025, initiated from the Sun’s active region AR 4098, which was located on the southwest quadrant of the Sun. The event began at approximately 01:40 UTC, reached its peak at 01:52 UTC, and ended around 01:57 UTC. This flare hit the sunlit side of the Earth, causing an increase in ionization level in the D region of the ionosphere. TEC, solar and geomagnetic variations during X1.1 SF Figure 18 shows the TEC values over HYDE and CPN stations along with the solar and geomagnetic indices, including F10.7 solar radio flux, Ap index, SSN, and Kp index for the period from 27th April 2025 to 27th May 2025. The TEC values at both stations exhibited regular diurnal variations with significant disturbances during this period. The F10.7 solar radio flux ranged from 120 to 165 sfu, indicating moderate solar activity. The SSN values varied between 32 and 128, reflecting variable solar conditions. A notable geomagnetic disturbance was observed around 17th–18th May 2025, where the Ap index reached approximately 90 nT and the Kp index exceeded 6, indicating a moderate geomagnetic storm. The quiet reference day of 28th April 2025 is clearly visible in Fig. 19, with both Ap and Kp values remaining low (Kp < 3), confirming undisturbed ionospheric conditions. On 25th May 2025, the X1.1 SF event produced a noticeable enhancement in TEC values at both HYDE and CPN stations. TEC prediction comparison during X1.1 SF at HYDE station Figure 19 presents a comparison between the True TEC and the TEC values predicted by four studied prediction models (i.e., KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models) during X1.1 SF (the studied fourth SF event) on 25th May 2025 at HYDE station. The X-axis represents the period from 22nd May 2025 to 27th May 2025 along with the universal time in Hrs [UTC]. Meanwhile, the Y-axis represents the TEC values in units: TECU. According to the observation, a noticeable slight increase in True TEC was observed around 09:00 UTC on 26th May 2025, where it was attributed to the X1.1 SF event (SSN > 60 and F10.7 > 120 sfu). On 22nd May 2025, the KAN, DBN-Adagrad and KAN-Adagrad models performed better than the IRI-2020 model, although they slightly over-predicted the True TEC values between 09:00 UTC and 11:00 UTC. Likewise, the IRI-2020 model also showed stronger over-prediction on 22nd May 2025. On 23rd May 2025, we observed noticeable slight fluctuations during the peak period, which may be associated with increased solar activity indicated by SSN and F10.7 values exceeding their threshold (SSN > 60 and F10.7 > 120 sfu). During this period, the KAN-Adagrad and DBN-Adagrad models showed slightly over-prediction only between 09:00 UTC and 11:00 UTC, whereas both KAN models and the DBN-Adagrad model performed relatively better than the IRI-2020 model on the remaining days. Between 24th May 2025 and 27th May 2025, the KAN-Adagrad and DBN-Adagrad models showed the best performance among the four studied models, whereas the KAN model showed moderate performance with slight under- and over-predictions during the TEC peak periods and at the end of the day. On the SF existence (25th May 2025), all four prediction models over-predicted the True TEC values from 10:00 UTC to 16:00 UTC. On 26th May 2025, the KAN-Adagrad, DBN-Adagrad and IRI-2020 models slightly under-predicted the True TEC, while the KAN model showed slight over-prediction during the TEC peak period (10:00 UTC to 13:00 UTC). Overall, the results showed that both KAN and KAN-Adagrad models and the DBN-Adagrad model provided better performance than the IRI-2020 model with significant deviations in several intervals. Figure 20 compares the True TEC values with KAN, DBN-Adagrad, KAN-Adagrad and IRI-2020 predicted TEC values using a contour plot. The contour plots are plotted to demonstrate the TEC variations from 22nd May 2025 to 27th May 2025, covering quiet and SF days. As for the contour plots, we observed that the contour plot of the KAN-Adagrad and DBN-Adagrad models also matched closely with that of True TEC. Table 13 represents the comparison between the performance metrics: WIA, KGE, and SCC for the four studied prediction models (i.e., KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models) during 22nd -27th May 2025 at HYDE station. The details in Table 13 are the same as in Table 7, except for the different studied SF intensity and duration. On the day before the SF event (24th May 2025), the highest WIA and SCC values obtained from KAN-Adagrad and DBN-Adagrad models were 0.9882, 0.9892 and 0.9809,0.9887, respectively, while the highest KGE of 0.9413 was discovered using the KAN model. The results showed that both KAN predicted with higher efficiency, and the DBN-Adagrad model predicted with higher correlation on 24th May 2025. These trends have been followed throughout the pre-event days. On 25th May 2025 (the day of the SF existence), the WIA, KGE and SCC values for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9608, 0.8427, 0.9609], [0.9551, 0.7505, 0.9704], [0.9511,0.7324,0.9739] and [0.9112, 0.6488, 0.9904], respectively. The results identified that KAN predicted with higher efficiency, and KAN-Adagrad and DBN-Adagrad models predicted with higher correlation on 25th May 2025. On 26th May 2025, the KAN model performed better than other models in terms of WIA and KGE. On 27th May 2025, the DBN-Adagrad model performed well compared to other models. The averages of WIA, KGE, and SCC for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9768, 0.9175, 0.9581], [0.9784, 0.8763, 0.9607], [0.9728,0.8870,0.9529] and [0.9019, 0.6309, 0.9519], respectively. The results indicated that both KAN and KAN-Adagrad models performed better than the other models for this X1.1 SF event. Table 13Performance evaluation during X1.1 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad22-05-20250.98710.99430.71550.99810.94350.9305-0.03720.98270.98700.99390.98000.983523-05-20250.98230.98350.92250.97140.94730.94700.75950.94370.96260.96520.92170.933924-05-20250.97340.98820.95680.98920.94130.93940.80530.93960.94610.98090.94430.988725-05-20250.96080.95510.91120.95110.84270.75050.64880.73240.96090.97040.99040.973926-05-20250.98970.96510.94660.93530.93580.79170.76620.78890.96260.88430.94700.856527-05-20250.96750.98410.95900.99150.89460.89840.84280.93470.92960.96960.92780.9809 Average 0.9768 0.9784 0.9019 0.9728 0.9175 0.8763 0.6309 0.8870 0.9581 0.9607 0.9519 0.9529 Table 13Performance evaluation during X1.1 SF (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad22-05-20250.98710.99430.71550.99810.94350.9305-0.03720.98270.98700.99390.98000.983523-05-20250.98230.98350.92250.97140.94730.94700.75950.94370.96260.96520.92170.933924-05-20250.97340.98820.95680.98920.94130.93940.80530.93960.94610.98090.94430.988725-05-20250.96080.95510.91120.95110.84270.75050.64880.73240.96090.97040.99040.973926-05-20250.98970.96510.94660.93530.93580.79170.76620.78890.96260.88430.94700.856527-05-20250.96750.98410.95900.99150.89460.89840.84280.93470.92960.96960.92780.9809 Average 0.9768 0.9784 0.9019 0.9728 0.9175 0.8763 0.6309 0.8870 0.9581 0.9607 0.9519 0.9529 Performance evaluation during X1.1 SF (HYDE station). TEC prediction comparison during X1.1 SF at CPN station Figure 21 shows the comparison between the True TEC and the TEC values predicted by KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models of X1.1 SF (the studied fourth SF event) during 22nd -27th May 2025 at CPN station. From the observation, the sudden fluctuations in True TEC observed during the study period were mainly associated with SF activity (SSN > 60 and F10.7 > 120 sfu). On 22nd May 2025, both proposed KAN models and DBN-Adagrad performed well as compared to the IRI-2020 model. However, during the TEC peak period (from 10:00 UTC to 14:00 UTC), the sudden TEC fluctuations were over-predicted by KAN-Adagrad, DBN-Adagrad and KAN models. On 23rd May 2025, all four studied prediction models under-predicted the True TEC values during the TEC peak hours (09:00 UTC to 13:00 UTC); however, both KAN models and DBN-Adagrad performed slightly better than the IRI-2020 model. On 24th May 2025, both KAN models and DBN-Adagrad slightly over-predicted the TEC values around 13:00 UTC, but they provided better performance than the IRI-2020 model. On solar flare event day (25th May 2025), the KAN-Adagrad and DBN-Adagrad models provided better performance than the other two models, and the KAN model showed strong under-prediction towards the end of that day. After the solar flare event, the KAN-Adagrad and DBN-Adagrad models maintained better performance as compared to the other two models. On 26th May 2025, noticeable rapid fluctuations in the True TEC values were observed, particularly during the peak period. These fluctuations may be associated with the X1.1 solar flare, along with SSN and F10.7 values exceeding their threshold levels (SSN > 60 and F10.7 > 120 sfu). During this period, all four models under-predicted the True TEC values. In contrast, the KAN model showed slight over-prediction on 27th May 2025, particularly during the TEC peak period around 15:00 UTC. Overall, the results suggested that the IRI-2020 model consistently under-predicted the True TEC values, and the KAN model provided moderate performance throughout the study period. Moreover, the KAN-Adagrad and DBN-Adagrad models gave the best performance among all four studied prediction models, although some TEC fluctuations were not accurately predicted. Figure 22 compares the true TEC and the KAN, DBN-Adagrad, KAN-Adagrad and IRI-2020 model-predicted TEC values using contour plots. The contour plots are plotted for TEC variations from 22nd May 2025 to 27th May 2025, including both SF and normal days. The contour plot allowed the identification of TEC variations in several areas at specific hours of each day. We found that the contour plot of the KAN-Adagrad and DBN-Adagrad model-predicted TEC values also matched closely with the contour plot of true TEC on this X1.1 SF event. Table 14 highlights the comparative performance based on statistical metrics such as WIA, KGE, and SCC of X1.1 SF event at CPN station during 22nd May 2025 and 27th May 2025 for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models. The details in Table 14 are similar to those in Table 7, except for the different studied SF intensity and duration. On 24th May 2025 (a day before the SF event), the KAN model performed the best; however, on the day of the SF event (25th May 2025), the performance of the KAN-Adagrad model was better than that of the KAN, DBN-Adagrad and IRI-2020 models. On the SF event day, the WIA, KGE, and SCC of KAN-Adagrad, KAN, and IRI-2020 models were [0.9782, 0.8971, 0.9522], [0.8276, 0.5791, 0.8235], [0.9418, 0.8732, 0.9278] and [0.5701, 0.1946, 0.6274], respectively. The averages of WIA, KGE, and SCC for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.8657, 0.6688, 0.8091], [0.9270, 0.8012, 0.8739], [0.9087, 0.7839, 0.8458] and [0.5114, 0.0330, 0.3529], respectively. The results identified that the KAN-Adagrad model generally provides better performance than other models, because it provided higher predictive accuracy and correlation. Table 14Performance evaluation during X1.1 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad22-05-20250.86770.93500.48790.89560.75010.8116-0.05080.78020.71650.87390.28960.792223-05-20250.89410.93700.53290.96600.77330.7716-0.01130.84080.86520.93220.42790.957424-05-20250.93740.88820.52150.87310.78860.69590.07270.63240.84870.78610.32710.788725-05-20250.82760.97820.57010.94180.57910.89710.19460.87320.82350.95220.62420.927826-05-20250.86760.93730.48900.88830.71470.84750.02640.79020.81480.89910.36660.850427-05-20250.79970.88610.46700.88710.40690.7832-0.03340.78670.78610.80000.08200.7583 Average 0.8657 0.9270 0.5114 0.9087 0.6688 0.8012 0.0330 0.7839 0.8091 0.8739 0.3529 0.8458 Table 14Performance evaluation during X1.1 SF (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad22-05-20250.86770.93500.48790.89560.75010.8116-0.05080.78020.71650.87390.28960.792223-05-20250.89410.93700.53290.96600.77330.7716-0.01130.84080.86520.93220.42790.957424-05-20250.93740.88820.52150.87310.78860.69590.07270.63240.84870.78610.32710.788725-05-20250.82760.97820.57010.94180.57910.89710.19460.87320.82350.95220.62420.927826-05-20250.86760.93730.48900.88830.71470.84750.02640.79020.81480.89910.36660.850427-05-20250.79970.88610.46700.88710.40690.7832-0.03340.78670.78610.80000.08200.7583 Average 0.8657 0.9270 0.5114 0.9087 0.6688 0.8012 0.0330 0.7839 0.8091 0.8739 0.3529 0.8458 Performance evaluation during X1.1 SF (CPN station). TEC prediction during a quiet day on 28th April 2025 at HYDE station Figure 23 shows the comparison between the observed TEC values and the values predicted by the KAN, KAN-Adagrad, DBN-Adagrad, and IRI-2020 models. The X-axis represents the date from 25th to 30th April 2025 along with Universal Time (UTC in hours), while the Y-axis represents TEC values in TECU. From this period, 28th April 2025 is considered as a geomagnetically quiet day, since the Kp index is less than 3 and the Ap index is less than 15 nT. This indicates stable ionospheric conditions. The solar parameters, such as sunspot number (SSN = 123) and F10.7 (155.7 sfu), also remain nearly constant during this period, showing no significant variation in solar activity. From the results, it is observed that the KAN, KAN-Adagrad, and DBN-Adagrad models perform better than the IRI-2020 model. These models are able to capture even small variations in TEC values caused by minor solar radiation effects. In contrast, the IRI-2020 model tends to underestimate the TEC values throughout the observation period. Overall, the results clearly indicate that the KAN, KAN-Adagrad, and DBN-Adagrad models provide more accurate TEC predictions compared to the IRI-2020 model under quiet ionospheric conditions. Figure 24 compares the true TEC and the KAN, DBN-Adagrad, KAN-Adagrad and IRI-2020 model-predicted TEC values using contour plots. The contour plots are plotted for TEC variations from 25th April 2025 to 30th April 2025. The contour plot allowed the identification of TEC variations in several areas at specific hours of each day. We found that the contour plot of the KAN-Adagrad and DBN-Adagrad model-predicted TEC values also matched closely with the contour plot of true TEC on this quiet day (28th April 2025) in the HYDE station. Table 15 illustrates the comparison of statistical parameters such as WIA, KGE and SCC of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models for the data obtained from the HYDE station. On 27th April 2025, the WIA, KGE and SCC values for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9953, 0.9697, 0.9861], [0.9959, 0.9720, 0.9870], [0.9957,0.9642,0.9878] and [0.9134, 0.6566, 0.9355], respectively. These results indicated that the KAN, KAN-Adagrad and DBN-Adagrad models provided higher predictive efficiency, and the KAN-Adagrad model gave higher correlation with the True TEC on that day. Likewise, on 28th April 2025, the KAN model, KAN-Adagrad, and DBN-Adagrad performed better than the IRI-2020 model. The averages of the statistical metrics (WIA, KGE, and SCC) for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.8657, 0.6688, 0.8091], [0.9270, 0.8012, 0.8739], [0.9087,0.7839, 0.8458] and [0.5114, 0.0330, 0.3529], respectively. Thus, the KAN-Adagrad model predicted overall well. The results showed that the KAN-Adagrad model had better efficiency and correlation to predict the TEC values than other predictive models (KAN, DBN-Adagrad and IRI-2020) on this Quiet day. Table 15Performance evaluation during Quiet day (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad25-04-20250.99610.99030.89570.99370.95460.91820.63190.95180.99040.98000.92320.973026-04-20250.99660.99500.88740.99580.95980.94060.65260.97150.98610.98000.90220.978327-04-20250.99530.99590.91340.99570.96970.97200.65660.96420.98610.98700.93550.987828-04-20250.99110.98910.86660.98530.98090.93190.63210.96400.96610.97300.85650.967029-04-20250.98170.99510.88370.99190.90010.93060.66520.93130.96780.98960.87910.987030-04-20250.99460.99300.91280.98870.97360.93840.69850.91720.98260.98350.92170.9878 Average 0.8657 0.9270 0.5114 0.9087 0.6688 0.8012 0.0330 0.7839 0.8091 0.8739 0.3529 0.8458 Table 15Performance evaluation during Quiet day (HYDE station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad25-04-20250.99610.99030.89570.99370.95460.91820.63190.95180.99040.98000.92320.973026-04-20250.99660.99500.88740.99580.95980.94060.65260.97150.98610.98000.90220.978327-04-20250.99530.99590.91340.99570.96970.97200.65660.96420.98610.98700.93550.987828-04-20250.99110.98910.86660.98530.98090.93190.63210.96400.96610.97300.85650.967029-04-20250.98170.99510.88370.99190.90010.93060.66520.93130.96780.98960.87910.987030-04-20250.99460.99300.91280.98870.97360.93840.69850.91720.98260.98350.92170.9878 Average 0.8657 0.9270 0.5114 0.9087 0.6688 0.8012 0.0330 0.7839 0.8091 0.8739 0.3529 0.8458 Performance evaluation during Quiet day (HYDE station). TEC prediction during a quiet day on 28th April 2025 at CPN station Figure 25 shows the comparison between the observed TEC values and the values predicted by the KAN, KAN-Adagrad, DBN-Adagrad, and IRI-2020 models. From this period, 28th April 2025 is considered as a geomagnetically quiet day, since the Kp index is less than 3 and the Ap index is less than 15 nT, indicating stable ionospheric conditions. The solar parameters, such as the sunspot number (SSN = 123) and F10.7 (155.7 sfu), also remain nearly constant during this period, showing no significant variation in solar activity. From the results, it is observed that the KAN, KAN-Adagrad, and DBN-Adagrad models perform better than the IRI-2020 model. These models are able to capture even small variations in TEC values caused by minor solar radiation effects. However, during the 26th–27th April 2025 period, the observed TEC reaches a maximum value (200 TECU), but this enhancement is not accurately predicted by any of the models. Meanwhile, the IRI-2020 model tends to underestimate the TEC values throughout the observation period. Overall, the results indicate that the KAN-Adagrad and DBN-Adagrad models provide more accurate TEC predictions compared to the KAN and IRI-2020 models under quiet ionospheric conditions. Figure 26 compares the true TEC and the KAN, DBN-Adagrad, KAN-Adagrad and IRI-2020 model-predicted TEC values using contour plots. The contour plots are plotted for TEC variations from 25th April 2025 to 30th April 2025. The contour plot allowed the identification of TEC variations in several areas at specific hours of each day. We found that the contour plot of the KAN-Adagrad and DBN-Adagrad model-predicted TEC values also matched closely with the contour plot of true TEC on this quiet day (28th April 2025) in the CPN station. Table 16 illustrates the comparison of statistical parameters such as WIA, KGE and SCC of KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models for the data obtained from the CPN station. On 27th April 2025, the WIA, KGE and SCC values for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9065, 0.7407, 0.8843], [0.9671, 0.8016, 0.9296], [0.9618, 0.8023, 0.9443] and [0.5447, 0.0260, 0.6496], respectively. These results indicated that the KAN-Adagrad and DBN-Adagrad models provided higher predictive efficiency than the KAN model and the IRI-2020 model. Likewise, on 28th April 2025, KAN-Adagrad and DBN-Adagrad performed better than KAN and the IRI-2020 model. The averages of the statistical metrics (WIA, KGE, and SCC) for KAN, KAN-Adagrad, DBN-Adagrad and IRI-2020 models were [0.9348, 0.7720, 0.8790], [0.9631, 0.8659, 0.9074], [0.9574, 0.8547, 0.9065] and [0.5545, 0.0625, 0.6752], respectively. Thus, the KAN-Adagrad model predicted overall well. The results showed that the KAN-Adagrad model had better efficiency and correlation to predict the TEC values than other predictive models (KAN, DBN-Adagrad and IRI-2020) on this Quiet day. Table 16Performance evaluation during Quiet day (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad25-04-20250.95690.97430.54810.97380.80020.93220.10110.92890.92960.91910.68870.927826-04-20250.90020.93670.57890.93060.75740.79900.08570.75370.85040.85830.77570.862627-04-20250.90650.96710.54470.96180.74070.80160.02600.80230.88430.92960.64960.944328-04-20250.92890.94480.58240.94580.78720.78600.06880.85570.79830.85390.66960.845229-04-20250.96910.98110.52320.97800.84630.96100.01500.91060.91740.94960.63220.948730-04-20250.94710.97480.55010.95440.70000.91550.07850.87710.89390.93390.63570.9104 Average 0.9348 0.9631 0.5545 0.9574 0.7720 0.8659 0.0625 0.8547 0.8790 0.9074 0.6752 0.9065 Table 16Performance evaluation during Quiet day (CPN station).DATEWIAKGESCCKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-AdagradKANKAN-AdagradIRI-2020DBN-Adagrad25-04-20250.95690.97430.54810.97380.80020.93220.10110.92890.92960.91910.68870.927826-04-20250.90020.93670.57890.93060.75740.79900.08570.75370.85040.85830.77570.862627-04-20250.90650.96710.54470.96180.74070.80160.02600.80230.88430.92960.64960.944328-04-20250.92890.94480.58240.94580.78720.78600.06880.85570.79830.85390.66960.845229-04-20250.96910.98110.52320.97800.84630.96100.01500.91060.91740.94960.63220.948730-04-20250.94710.97480.55010.95440.70000.91550.07850.87710.89390.93390.63570.9104 Average 0.9348 0.9631 0.5545 0.9574 0.7720 0.8659 0.0625 0.8547 0.8790 0.9074 0.6752 0.9065 Performance evaluation during Quiet day (CPN station). Overall performance metrics comparison of four models for TEC prediction Figure 27 compares the average WIA, KGE, and SCC values of the KAN, KAN-Adagrad, DBN-Adagrad, and IRI-2020 models under four selected X-class solar flare events and quiet day conditions at the HYDE and CPN stations. In terms of WIA, the KAN-Adagrad model shows the best performance, with values ranging from 0.9251 to 0.9931. This indicates a close match between the predicted and observed TEC values across most conditions. The DBN-Adagrad model also performs well, with values between 0.9087 and 0.9927, followed by the KAN model, which ranges from 0.8657 to 0.9918. In contrast, the IRI-2020 model shows significantly lower values, particularly at the CPN station, where it drops to around 0.4933, indicating weaker agreement with observed TEC values. Similarly, for KGE, the KAN-Adagrad model maintains stable and relatively high values, ranging from 0.8012 to 0.9396 during quiet periods and reaching up to 0.8985 during solar flare events. The DBN-Adagrad model shows comparable performance, with values between 0.7839 and 0.95, while the KAN model ranges from 0.6688 to 0.9565. However, the IRI-2020 model again shows poor performance, especially at the CPN station, where the value decreases to 0.0330, indicating its difficulty in capturing rapid ionospheric variations. For SCC, both KAN-Adagrad and KAN models demonstrate strong correlation with the observed TEC values. The KAN-Adagrad model ranges from 0.8522 to 0.9822, and the KAN model from 0.8091 to 0.9798. The DBN-Adagrad model also maintains good correlation, with values between 0.8458 and 0.9802. In comparison, the IRI-2020 model shows lower correlation overall, particularly at the CPN station, where the value drops to around 0.3529. Overall, the results indicate that the KAN-Adagrad model consistently provides the most accurate predictions across different conditions. The DBN-Adagrad and KAN models also show reliable performance. In contrast, the IRI-2020 model performs comparatively poorly, especially during solar flare events and at the low-latitude CPN station. These observations suggest that the use of the Adagrad optimiser enhances the prediction capability of the KAN and DBN model for TEC variations. Fig. 27Comparison of WIA, KGE, and SCC values of all models at HYDE and CPN under solar flare and quiet conditions. Comparison of WIA, KGE, and SCC values of all models at HYDE and CPN under solar flare and quiet conditions. Statistical comparison of the models using the friedman test The Friedman test was used to compare the performance of the four models across four different solar flare events and the quiet day over two different stations. The calculated p-values are shown in Table 17. Since all p-values are much smaller than 0.05, the null hypothesis is rejected. Also, it indicates that there is a significant difference among the models for all selected events and stations. More details about the Friedman test and the formula used for calculating the p-values are described in detail by Rainio et al. (2024)34. Table 17Friedman Test p-values for Model Comparison Across Solar Flare and Quiet-day Events.EVENTSTATIONp-valuesSolar Flare (26-10-2024)HYDE3.75E-12Solar Flare (26-10-2024)CPN5.52E-08Solar Flare (13-05-2025)HYDE9.51E-08Solar Flare (13-05-2025)CPN3.95E-17Solar Flare (14-05-2025)HYDE7.10E-31Solar Flare (14-05-2025)CPN7.34E-17Solar Flare (25-05-2025)HYDE5.07E-24Solar Flare (25-05-2025)CPN1.45E-22Quiet DayHYDE1.12E-28Quiet DayCPN1.09E-28 Table 17Friedman Test p-values for Model Comparison Across Solar Flare and Quiet-day Events.EVENTSTATIONp-valuesSolar Flare (26-10-2024)HYDE3.75E-12Solar Flare (26-10-2024)CPN5.52E-08Solar Flare (13-05-2025)HYDE9.51E-08Solar Flare (13-05-2025)CPN3.95E-17Solar Flare (14-05-2025)HYDE7.10E-31Solar Flare (14-05-2025)CPN7.34E-17Solar Flare (25-05-2025)HYDE5.07E-24Solar Flare (25-05-2025)CPN1.45E-22Quiet DayHYDE1.12E-28Quiet DayCPN1.09E-28 Friedman Test p-values for Model Comparison Across Solar Flare and Quiet-day Events. Conclusion This paper presented the prediction of ionospheric TEC during four solar flare (SF) events using the Kolmogorov-Arnold Network (KAN) with and without the Adaptive Gradient Algorithm (Adagrad) optimiser and compared the results with the DBN-Adagrad model, IRI-2020-predicted TEC values, and the observed TEC at Hyderabad (HYDE), India and Chumphon (CPN), Thailand. The developed models were trained using chronologically arranged datasets and validated through the PSIS-LFO Cross-Validation approach. Four solar flare events, namely X1.86 (26 October 2024), X1.21 (13 May 2025), X2.7 (14 May 2025), and X1.1 (25 May 2025), were analysed in this study. The obtained results demonstrated that both KAN-Adagrad and DBN-Adagrad generally achieved higher predictive accuracy than the KAN and IRI-2020 models for most of the investigated events. During the X1.86 flare event, KAN-Adagrad produced the highest performance at HYDE (WIA = 0.9677, KGE = 0.8586, SCC = 0.9533), while DBN-Adagrad slightly showed better performance than KAN-Adagrad at CPN with WIA, KGE, and SCC values of 0.9533, 0.8248, and 0.9458, respectively. For the X1.21 flare event, KAN showed the best performance at HYDE (WIA = 0.9777, KGE = 0.8845, SCC = 0.9387), whereas DBN-Adagrad provided better agreement with the observed TEC values at CPN (WIA = 0.9391, KGE = 0.8223, SCC = 0.8893). Similarly, during the X2.7 flare event, both DBN-Adagrad and KAN-Adagrad showed good predictive capability at HYDE, with DBN-Adagrad achieving slightly higher statistical performance (WIA = 0.9803, KGE = 0.9086, SCC = 0.9459) than KAN-Adagrad (WIA = 0.9724, KGE = 0.8985, SCC = 0.9383). At CPN, DBN-Adagrad again showed the best overall performance, with WIA, KGE, and SCC values of 0.9477, 0.8444, and 0.9101, respectively. For the X1.1 flare event, both KAN and KAN-Adagrad performed well at HYDE, while KAN-Adagrad achieved the highest predictive accuracy at CPN (WIA = 0.9270, KGE = 0.8012, SCC = 0.8739). The quiet-day analysis further demonstrated the effectiveness of the proposed models, where KAN-Adagrad consistently achieved the best overall performance at both HYDE and CPN stations. In general, the incorporation of the Adagrad optimiser improved the predictive capability of the KAN architecture, resulting in higher agreement, efficiency, and correlation with the observed TEC values. The consistently high WIA and SCC values obtained across different solar flare intensities indicate that the proposed KAN-Adagrad framework is capable of effectively capturing the temporal variability of TEC under both disturbed and quiet ionospheric conditions. Overall, the results suggest that KAN-Adagrad provides a reliable and effective approach for TEC prediction, especially under disturbed space weather conditions such as solar flares. These findings suggest that KAN-based approaches can be effectively used for TEC prediction during solar flare events. However, the proposed model requires further validation using TEC data obtained from different latitude GNSS stations, during different phases of the solar cycle and other solar flare events, to assess its robustness and generalizability.