A Diffusion-Based Machine Learning Model for Forecasting the Bottomside Ionosphere

Updating an ionospheric model to reflect real-time conditions (assimilation) is challenging, especially when the available data we want to incorporate is of several different types. We present a new assimilative model based on latent diffusion transformers, a form of deep neural network that uses progressive denoising, attention, and variational autoencoders to sample from a conditional probability distribution — in this case, the distribution of plausible global maps of critical frequency (foF2), F2 density peak height (hmF2), maximum usable frequency (MUF), and vertical total electron content (TEC), given a time of day, time of year, and a collection of recent observations from ionosondes or global navigation satellite system (GNSS) receivers. One challenge when training diffusion models is the need for "clean examples" in the desired output format, i.e. global maps of ionospheric characteristics. Since available maps do not have fully contiguous spatial coverage, we're forced to train on synthetic data sampled from the International Reference Ionosphere (IRI-2020). We add coherent, semi-realistic random perturbations to the IRI predictions; this data augmentation process allows the model to estimate the spatial and temporal covariance of the "noisy" real-world phenomena that the "smooth" IRI lacks. Ongoing research shows promise in "rejection-sampling fine-tuning" for bringing real-world observations into the training loop. In this process, after the model is bootstrapped on synthetic training data, multiple forecasts are carried out for a given set of real-world initial conditions with varying noise seeds, and the forecast that most closely matches the real-world followup observations is used as additional training data. The model's forecast accuracy over a period of 0 to 24 hours isevaluated against space-based Global Navigation Satellite System radio occultation (GNSS-RO) and ground-based Global Navigation Satellite System total electron content (GNSS TEC) observations from Spring 2026.