Huang et al. (2026) Physically constrained machine learning improves radiation, clouds, and albedo forecasts globally

Identification

• Journal: npj Climate and Atmospheric Science
• Year: 2026
• Date: 2026-09-21
• Authors: Qiusheng Huang, Xiaohui Zhong, Junyu Cai, Linfan Zhou, Zijian Zhu, Wenxu Qian, Lei Chen, Libo Wu, Dazhi Yang, Hao Li
• DOI: 10.1038/s41612-026-01550-1

Research Groups

• Fudan University's Artificial Intelligence Innovation and Incubation Institute
• Shanghai Innovation Institute
• Shanghai Academy of Artificial Intelligence for Science
• FuXi Intelligent Computing Technology Co., Ltd.
• Electric Power Research Institute, State Grid Zhejiang Electric Power Co., Ltd.
• School of Data Science, Fudan University
• Institute for Big Data, Fudan University
• MOE Laboratory for National Development and Intelligent Governance, Fudan University
• Harbin Institute of Technology's School of Electrical Engineering and Automation

Short Summary

This study presents FuXi-RTM, a physics-guided machine learning model that improves radiation, clouds, and albedo forecasts globally by incorporating a differentiable radiative transfer model. The results show that FuXi-RTM outperforms state-of-the-art physics-based models for shortwave radiation and cloud-related variables.

Objective

• To develop a machine learning-based weather forecasting model that explicitly integrates physically grounded and learnable representations of clouds, radiation, and surface albedo.
• To evaluate the performance of FuXi-RTM against reanalysis data, independent ground-based station observations, and satellite-derived radiation products.

Study Configuration

• Spatial Scale: Global scale, with a focus on shortwave radiation and cloud-related variables.
• Temporal Scale: 10-day forecasts, with initializations at 12 UTC between January 1, 2024, and February 28, 2025.

Methodology and Data

• Models used:
  • FuXi-RTM: A physics-guided machine learning model that incorporates a differentiable radiative transfer model.
  • FuXi-base: A fully data-driven model without radiative constraints.
  • FuXi-base+: A variant of FuXi-base with an additional branch to output radiative fluxes, but without backpropagating radiative transfer gradients through the atmospheric state.
  • ECMWF HRES: The world-leading physics-based high-resolution forecasts from the European Centre for Medium-Range Weather Forecasts (ECMWF).
• Data sources:
  • ERA5 reanalysis data
  • Independent ground-based station observations from the Baseline Surface Radiation Network (BSRN)
  • Satellite-derived radiation products from the Clouds and the Earth's Radiant Energy System (CERES) project

Main Results

• FuXi-RTM outperforms FuXi-base, FuXi-base+, and ECMWF HRES for shortwave radiation variables (SSRD, SSR, and FDIR).
• FuXi-RTM consistently achieves the best performance for cloud cover variables (TCC, LCC, MCC, and HCC) throughout the 10-day forecasts.
• FuXi-RTM reduces both systematic bias and random error variability for radiation-related variables.

Contributions

• This study presents a practical way to incorporate nonlinear physical processes into machine learning-based weather forecasting systems by using differentiable surrogates as training constraints that link physical process errors to the forecast state.
• The results demonstrate the potential value of FuXi-RTM for renewable energy forecasting, particularly in improving solar power forecasts.

Funding

• This study was funded by the Shanghai Innovation Institute and the MOE Laboratory for National Development and Intelligent Governance, Fudan University.

Citation

@article{Huang2026Physically,
  author = {Huang, Qiusheng and Zhong, Xiaohui and Cai, Junyu and Zhou, Linfan and Zhu, Zijian and Qian, Wenxu and Chen, Lei and Wu, Libo and Yang, Dazhi and Li, Hao},
  title = {Physically constrained machine learning improves radiation, clouds, and albedo forecasts globally},
  journal = {npj Climate and Atmospheric Science},
  year = {2026},
  doi = {10.1038/s41612-026-01550-1},
  url = {https://doi.org/10.1038/s41612-026-01550-1}
}