TGR-foF2: TEC-Driven Graph-Based Spatial Reconstruction of Continuous foF2 Fields with Physics-Guided Learning

Abstract

Spatial reconstruction of the ionospheric F2-layer (foF2) critical frequency is essential for high-frequency communication and ionospheric modeling, yet remains challenging due to the sparse distribution of ground-based ionosondes. This paper proposes a graph attention network framework that reconstructs spatially continuous foF2 fields by treating global navigation satellite system (GNSS) total electron content (TEC) as a direct spatial covariate, bypassing intermediate conversion steps. The framework is named TEC-driven graph-based spatial reconstruction of continuous foF2 fields (TGR-foF2). Each ionosonde station is modeled as a graph node with co-located TEC, coordinates, and solar-geomagnetic features; three physically motivated edge types (geographic proximity, TEC similarity, geomagnetic latitude conjugacy) govern inter-station message passing; and a physics-guided output head encodes the first-order physical scaling between foF2 and the square root of TEC while learning residual corrections. Evaluated via leave-one-station-out cross-validation across eight Australian stations using 2020 full-year data, TGR-foF2 achieves a mean root mean squared error (RMSE) of 0.840 MHz, outperforming inverse distance weighting (0.912 MHz, −7.9%), IRI-2020 (0.976 MHz, −13.9%), and polynomial TEC-to-foF2 conversion (1.227 MHz, −31.5%). Ablation experiments confirm that TEC features, graph topology, and the attention mechanism are the three most impactful components. The strong correlation between station isolation distance and RMSE (r = 0.886, p = 0.003) identifies network sparsity as the primary remaining limitation.