AI Transformer Forecasts Nuclear Radiation Using Atmospheric Diffusion.

A new AI model, NRFormer+, has been developed to provide highly accurate nationwide forecasts of nuclear radiation. This is a critical advancement given the persistent public health and environmental risks associated with radiation, especially in the wake of events like the Fukushima accident. Existing forecasting methods struggle with the complex nature of radiation data, which is influenced by non-stationary time series, uneven monitoring station distribution, and the intricate interplay with meteorological conditions. NRFormer+ is a spatio-temporal Transformer that uniquely addresses these challenges. It combines non-stationary temporal attention and density-adaptive spatial attention with a novel atmospheric diffusion module. This module explicitly estimates how weather conditions, such as wind and temperature, drive radiation dispersion, injecting this crucial physical understanding directly into the neural network. The model has demonstrated state-of-the-art accuracy across multiple datasets and baselines, significantly improving predictions for sudden radiation changes while maintaining efficient inference.