MLP-Based FPGA Model for Ionospheric TEC Prediction over a Low-Latitude Region during G3 and G4 Geomagnetic Storms that Occurred in 2026 and Comparison with IRI-2020 Model

Abstract This study investigates ionospheric Total Electron Content (TEC) anomalies before, during, and after two geomagnetic storm periods in 2026 over a low-latitude region. Apart from the TEC anomaly analysis, this paper investigates the performance of a hardware-accelerated multi-layer perceptron (MLP) based Field Programmable Gate Array (FPGA) model coupled with an AdamW optimizer for predicting TEC during storm periods January 19, .2026 to January 21, 2026 and March 22, 2026, using TEC data obtained from the BAKO station, along with solar and geomagnetic indices such as SSN, F10.7, Dst, Ap, and KP obtained from the Omniweb which capture the primary physical drivers of ionospheric variability. Multiple statistical parameters, including RMSE, MAE, R2, MAPE, SMAPE, and NRMSE, were used to assess the performance of the MLP-based FPGA model. Performance evaluation parameters indicate that the developed model achieves an RMSE of 3.43 TECU, an MAE of 2.69 TECU, a MAPE of 20.03% and an R2 of 0.96 during the January 2026 storm. Similarly, the model achieves an RMSE of 4.37 TECU, an MAE of 3.43 TECU, a MAPE of 21.35% and an R2 of 0.97 during the March 2026 storm. The proposed MLP-FFNN (Feed-Forward Neural Network) architecture was validated against a Gated Recurrent Unit (GRU) model, demonstrating higher predictive accuracy while avoiding the heavy DSP and memory overhead of recurrent gating mechanisms. The MLP-FFNN model effectively captures TEC variability, especially under disturbed ionospheric conditions. This hardware architecture utilizes a sequential Multiply-Accumulate (MAC) state machine combined with Q16.16 fixed-point BRAM (Block Random Access Memory) Look-Up Tables for a non-linear activation function. This highly efficient design achieves real-time inference latency of 0.44–0.45 ms while consuming merely 2% of the available DSP slices and less than 0.3 Watts of power. Across the two events, noticeable TEC, solar and geomagnetic deviations were observed during both storm periods. The findings highlight the potential of integrating an AI algorithm with ultra-efficient physical hardware to predict TEC for low-latency space weather forecasting.