The accurate representation of the ionospheric electron density in 3-dimensions is a challenging problem because of the nature of horizontal and vertical structures on both small and large scales. This paper presents the development of a global three-dimensional (3-D) electron density reconstruction based on radio occultation data during 2006–2019 and neural networks. We demonstrate that the developed model based on COSMIC dataset only is capable of reproducing different ionospheric features when compared to independent datasets from ionosondes and incoherent scatter radars (ISR) in low, middle and high latitude regions. Following some existing modelling efforts based on similar or related datasets and technique we divided the problem into fine resolution grid cells of $5^{\circ }\times 15^{\circ }$ (geographic latitudes/longitudes) followed by development of the neural network subroutine per cell and later combining all the 864 sub-models to compile one global model. This approach has been demonstrated to be appropriate in enabling neural networks to learn, reproduce and generalise local and global behaviour of the ionospheric electron density. Based on ISR data, the 3D model improves maximum electron density of the F2 layer (NmF2) prediction by 10%–20% compared to IRI 2016 model during quiet conditions. For estimation of ionosonde ordinary critical frequency of the F2 layer (foF2) in 2009 at 1200 UT (universal time), the developed 3-D model gives average root mean square error (RMSE) values of 0.83 MHz, 1.06 MHz and 1.16 MHz compared to the IRI 2016 values of 0.92 MHz, 1.09 MHz and 1.01 MHz over the Africa–European, American and Asian sectors respectively making their performances statistically comparable. Compared to ionosonde data, the IRI 2016 model consistently shows a better performance for the hmF2 modelling results in almost all sectors during the investigated periods.

由于小尺度和大尺度水平和垂直结构的性质，在 3 维中准确表示电离层电子密度是一个具有挑战性的问题。本文介绍了基于 2006-2019 年无线电掩星数据和神经网络的全球三维 (3-D) 电子密度重建的发展。我们证明，与来自低、中、高纬度地区的电离探空仪和非相干散射雷达 (ISR) 的独立数据集相比，基于 COSMIC 数据集的开发模型仅能够再现不同的电离层特征。遵循一些基于相似或相关数据集和技术的现有建模工作，我们将问题划分为$5^{\circ }\times 15^{\circ }$（地理纬度/经度），然后为每个单元开发神经网络子程序，然后组合所有 864 个子模型以编译一个全局模型。这种方法已被证明适用于使神经网络能够学习、再现和概括电离层电子密度的局部和全局行为。基于 ISR 数据，与 IRI 2016 模型相比，3D模型在安静条件下将 F2 层 ( NmF2 ) 预测的最大电子密度提高了 10%–20%。用于估计 F2 层的离子探空仪普通临界频率 ( foF2) 在 2009 年 1200 UT（世界时），开发的 3-D 模型给出了 0.83 MHz、1.06 MHz 和 1.16 MHz 的平均均方根误差 (RMSE) 值，而 IRI 2016 的值是 0.92 MHz、1.09 MHz 和 1.01 MHz 分别在非洲-欧洲、美国和亚洲部门上运行，这使得它们的性能具有统计可比性。与离子探空仪数据相比，IRI 2016 模型在调查期间几乎所有部门的hmF2建模结果都显示出更好的性能。