On April 7, the Shanghai Artificial Intelligence Laboratory, together with the University of Science and Technology of China, Shanghai Jiaotong University, Nanjing University of Information Science and Technology, the Institute of Atmospheric Physics of the Chinese Academy of Sciences, and the Shanghai Central Meteorological Observatory released the large-scale global medium-term weather forecast model "Fengwu" . Based on multi-modal and multi-task deep learning methods, the AI large model "Fengwu" for the first time realized effective forecasting of core atmospheric variables at high resolution for more than 10 days, and surpassed DeepMind's release on 80% of the evaluation indicators. Model GraphCast [1]. In addition, "Fengwu" can generate global high-precision forecast results for the next 10 days in only 30 seconds, which is much more efficient than traditional models.
Ouyang Wanli, the leading scientist of the laboratory, said, "'Fengwu' is named after the 'Xiangfeng Tongwu' in the Qin and Han Dynasties. It is the earliest wind measuring equipment in the world. The large weather forecast model 'Fengwu' not only carries the wisdom of ancient Chinese It also means that the laboratory is committed to making breakthroughs and unremitting explorations in the field of AI for Science represented by meteorology."
Using the "Fengwu" large-scale model, the effective global weather forecast time exceeds 10 days for the first time
1. High-precision, long-term, high-efficiency: AI large-scale model forecasting global weather
How to improve the timeliness and accuracy of weather forecasts has always been a key issue in the industry. With the intensification of global climate change and frequent occurrence of extreme weather in recent years, expectations from all walks of life for the timeliness and accuracy of weather forecasts are increasing day by day. Among meteorological and climate forecasting tasks, global medium-range weather forecasting is one of the most important forecasting tasks. It aims to predict the state of the atmospheric system in the next 14 days. It is not only the basis of the currently widely used integrated weather forecasting system, but also the regional value Background fields and boundary conditions for weather forecasting systems.
In the past few decades, many remarkable achievements have been made in the field of global medium-range weather forecasting. It is only increased by 1 day every 10 years [2], which is difficult to meet the needs of social and economic development.
With the continuous maturity of deep learning technology and framework, large artificial intelligence models represented by ChatGPT and "Scholar" (Intern) have demonstrated excellent capabilities in the fields of natural language and vision, and artificial intelligence has also brought great benefits to fields such as earth science A new research idea.
Bai Lei, a young scientist at the Shanghai Artificial Intelligence Laboratory, said, "'Fengwu' provides a powerful and effective AI framework for global medium-range weather forecasting, and its leadership is reflected in three aspects: forecast accuracy, forecast timeliness, and resource efficiency."
In terms of forecast accuracy, compared with DeepMind's GraphCast, the 10-day forecast error of "Fengwu" is reduced by 10.87%, and compared with the traditional physical model, its error is reduced by 19.4%. In terms of forecast timeliness, according to commonly used international standards, the weather forecast results are available when the z500 ACC is greater than 0.6 [2], which can better guide forecasters to judge the future meteorological development situation. Previously, the best physical model in the world, HRES, had an effective forecast period of up to 8.5 days within this standard range, while "Fengwu" reached 10.75 days based on reanalysis data. In terms of resource efficiency, existing physical models are often run on supercomputers, while the "Fengwu" AI large model can run on a single GPU, and it only takes 30 seconds to generate global high-precision forecast results for the next 10 days.
Paper link: https://arxiv.org/abs/2304.02948
According to meteorological experts, although there are currently some products on the market that provide weather forecast services for the next 15 days, the forecast performance for more than 10 days still has great uncertainty and cannot meet the standard for effective forecasting. Practice has proved that combining observation with numerical forecasting and artificial intelligence can effectively improve the accuracy of numerical forecasting. For the first time, "Fengwu" has increased the validity of the global weather forecast to 10.75 days, which has great operational application value.
2. Multimodality and Multitasking: Deep Learning Drives Geosciences
The AI for Earth joint team of Shanghai Artificial Intelligence Laboratory proposed a deep learning method based on multi-modal and multi-task to build an AI weather forecast model, so as to achieve fast and accurate forecast of global medium-term weather.
During the design and training process of the AI model, the research team found that in the learning process, multiple atmospheric variables interact with each other in the optimization and can be regarded as a multi-task learning problem; atmospheric data has the characteristics of high resolution, high dimension and large volume , making it difficult to directly optimize the multi-step weather prediction results of the model.
"Fengwu" multimodal network structure. Different modalities are processed through different codecs, and multimodal features are fused through a cross-modal fusion module.
"Fengwu" uses multi-modal neural network and multi-task automatic weight balance to solve the problem of the representation and mutual influence of various atmospheric variables. The atmospheric variables it targets include: potential, humidity, zonal wind speed, meridional wind speed, temperature, and land surface. "Fengwu" regards these atmospheric variables as multi-modal information, which can be better processed by using a multi-modal network structure.
From the perspective of multi-task problems, the research team automatically learns the importance of each atmospheric variable, so that multiple atmospheric variables can be better coordinated and optimized. In order to optimize the multi-step prediction results of "Fengwu", the research team proposed a "replay buffer" strategy to reduce autoregressive prediction errors and improve long-term prediction performance.
Prediction results for different atmospheric variables. ACC is an indicator used to measure the effectiveness of the prediction results. The higher the value, the more effective the prediction results (the red line represents "Fengwu", and the black line represents GraphCast).
From the results, the forecasting skill of "Fengwu" is significantly higher than that of GraphCast in the mid-range forecast of 6 to 10 days. Among them, the representative z500 has achieved an effective forecast range of 10.75 days (ACC>0.6), which is the first time that a high-resolution global medium-range weather forecast system can effectively forecast atmospheric variables for more than 10 days.
In the future, the "Fengwu" AI large weather model can complement the traditional physical model. With its excellent performance and precision, it can provide more accurate and practical weather forecast information for production and life, help the digitalization of weather forecast, and provide a better service for agriculture, forestry, animal husbandry and fishery. , aviation and navigation and other industries and public security to provide strong support. It is reported that the AI for Earth team of the Shanghai Artificial Intelligence Laboratory will also apply artificial intelligence methods to a wider range of earth science issues such as meteorology, environment, astronomy, and geology to help "carbon neutrality", disaster prevention and mitigation, and energy security. major need.
references
1. Lam R, Sanchez-Gonzalez A, Willson M, et al. GraphCast: Learning skillful medium-range global weather forecasting[J]. arXiv preprint arXiv:2212.12794, 2022.
2. Bauer, Peter, Alan Thorpe, and Gilbert Brunet. “The quiet revolution of numerical weather prediction.” Nature 525.7567 (2015): 47-55.
Author: TechBeat hardcore broadcast
Illustration by 22 from IconScout
-The End-