image: ProNet integrates AI directly into traditional atmospheric equations. This hybrid approach accurately predicts air pollution and corrects multi-day errors without violating the laws of physics, minimizing the unfeasible guesses often made by pure AI systems.
Credit: ©Science China Press
Every day, millions of people rely on air quality forecasts to protect their health, while cities depend on them to trigger emergency pollution responses and emission controls. Yet delivering forecasts that are both pinpoint-accurate and scientifically explainable has long posed a daunting dilemma. Traditional chemical transport models simulate atmospheric physics and chemistry by calculating governing equations but are severely constrained by unpredictable human emissions. Meanwhile, purely data-driven artificial intelligence (AI) has advanced rapidly, but pure AI models lack physical common sense: they frequently produce erratic, physically unfeasible concentration swings, fail during unprecedented weather extremes, and operate as opaque black boxes that cannot diagnose why errors occur.
To break this trade-off, a research team led by Prof. Zifa Wang and Dr. Huangjian Wu from the Institute of Atmospheric Physics (IAP) at the Chinese Academy of Sciences, collaborating with Peking University, the China National Environmental Monitoring Center, and international partners, has developed ProNet—a physics-AI tightly coupled forecasting framework.
Rather than treating AI as an external add-on or a complete replacement for traditional submodules, ProNet embeds neural networks directly into the mathematical equations that govern the atmosphere. In plain terms, the AI dynamically handles the complex, localized uncertainties of chemical reactions and human emissions, while strict physical equations govern how the wind transports the pollution.
This physical constraint strictly prevents the system from making erratic, impossible guesses—a common problem known as "AI hallucinations." The real-world results are striking. When evaluated across 338 Chinese cities over a full year, ProNet’s 7-day forecasts matched or even beat the 1-day forecasts manually adjusted by human experts. Furthermore, during a severe spring dust storm, ProNet cut forecast errors in Beijing by 62%, proving its reliability during highly unpredictable weather extremes where pure AI models typically fail.
Beyond predictive accuracy, ProNet functions as a scientific diagnostic lens. By mapping forecast errors across space and time, the system performs quantitative error attribution. Remarkably, using only routine ground-level PM2.5 measurements, ProNet isolated component-specific uncertainty patterns and revealed systematic parameterization biases in current atmospheric models—pinpointing an overestimation of secondary nitrate formation and an underestimation of winter sulfate formation. These findings independently corroborate empirical results from the multinational East Asia multi-model intercomparison study (MICS-Asia III).
"Traditional AI in atmospheric science has largely been a black-box pursuit of accuracy," explains Prof. Zifa Wang, corresponding author of the study. "With ProNet, we show that AI can become an investigative tool. By embedding neural networks into physical equations, we preserve physical realism while using routine data to diagnose where our chemical understanding falls short." Dr. Huangjian Wu, first author of the paper, adds: "This bridges operational forecasting directly with scientific discovery, transforming everyday monitoring data into a feedback loop for refining Earth system models."
Journal
National Science Review
Method of Research
Computational simulation/modeling