Graph models break down on irregular networks: how physics saves radiation forecasting

78% of Japan’s radiation monitoring stations are located in less than 6% of the country’s territory, clustered around Fukushima. This violates the basic assumptions of standard graph models: an algorithm trained on such data simply does not know what is happening outside the dense network.

NRFormer+, a spatio-temporal Transformer for nationwide radiation background forecasting, addresses this practical limitation. The architecture combines non-stationary temporal attention with density-adaptive spatial attention, and then adds an atmospheric diffusion module. This module evaluates how meteorological factors—wind, temperature, and humidity—influence radiation scattering and embeds the physical signal into the network as an architectural prior.

The idea is not to “give the model more data,” but to make it respect the physics of atmospheric transport. Purely data-driven approaches fail to capture how radiation co-evolves with the weather, because observations alone do not explain the physics of the process. Incorporating diffusion as a prior is an attempt to bridge this gap between the data and the underlying mechanism.

Result: NRFormer+ outperforms 13 baseline models on two datasets, reducing the MAE during sudden changes by 19.1% compared to the strongest baseline, with comparable output latency. Sudden changes are precisely what is critical for emergency response, rather than for routine monitoring.

The code and datasets are publicly available, making it possible to verify the claimed accuracy on real-world data rather than on synthetic benchmarks.

Source: cs.AI updates on arXiv.org