Since ancient times, 'unpredictable weather' has been the most helpless lament of human beings towards the weather. However, the rapid development of artificial intelligence technology has shown scientists the prospect of more accurate predictions of weather and climate. We expect AI models to rival or even surpass physical models in terms of temporal and spatial resolution. ”On October 22, at the 36th Annual Meeting of the Chinese Meteorological Society, which opened in Nanjing, Jiangsu Province, Mu Mu, an academician of the CAS Member and a distinguished professor of Fudan University, pointed out in his report on the conference that the current research on predictability of weather and climate events focuses on clarifying the causes and mechanisms of prediction uncertainty, and AI models may help. At this annual conference, several academicians and scholars exchanged ideas on cutting-edge topics such as weather and climate prediction in the era of artificial intelligence, high-performance computing in the meteorological field, and new technologies for typhoon forecasting. And artificial intelligence is a high-frequency word in the topic. With multiple advantages, AI meteorological models have shown potential in areas such as Pangu, Fengwu, Fuxi... In China, these AI meteorological models with traditional Chinese cultural imprints are now being used in fields such as typhoon prediction and short-term climate forecasting, helping humans improve the accuracy and efficiency of weather forecasting. The use of AI technology to model and forecast various weather and climate events has gradually become a research hotspot. ”Mumu believes that using meteorological models for predictability research presents both opportunities and challenges. The AI weather model has three major advantages. Firstly, it has faster computation time. Traditional physical computing models require about an hour of computation, but AI can calculate it in just a few minutes; Secondly, the forecast accuracy is higher, for example, in the forecast of typhoon paths, the forecast accuracy of some large models is already higher than that of numerical models; The third is the built-in optimization module, which can conduct various predictability studies according to usage needs. ”Mumu stated that this provides opportunities for the application of large-scale models. With the help of AI, scientists and meteorologists have "deciphered" many mysteries of climate change. Mumu introduced that in response to the El Ni ñ o-Southern Oscillation (ENSO) phenomenon, known as the "global climate switch", some scholars have built an AI model based on the Transformer architecture (a deep learning model), and the prediction time has exceeded 18 months, far exceeding traditional dynamic models. In the forecast of Atlantic hurricanes in August this year, the results of large-scale models and dynamic numerical forecasts are becoming increasingly similar, and their judgments on the location of hurricane formation are almost the same. The Mu Mu team has also used large models to identify ENSO sensitive areas located in the central and western Pacific through data on sea surface temperature and offshore wind fields. In the prediction of the invasion of the Kuroshio into the South China Sea, they found the target observation sensitive area by optimizing the modules of the large model. In the view of Yu Yu, Director of the Shanghai Typhoon Research Institute of the China Meteorological Administration, "the global artificial intelligence weather forecasting model has obvious advantages in typhoon path prediction compared to traditional global numerical weather forecasting models." In August of this year, the numerical forecasting team of the institute launched the Shanghai Typhoon Intelligent Model 1.0 version trained on high-resolution typhoon reanalysis data. This model increases the forecast resolution of typhoons to the kilometer level and compresses the forecast time from 64 minutes in previous traditional numerical forecast models to 3 minutes. The integration of AI and traditional solutions breaks the bottleneck of computing accuracy. In the face of rapidly changing weather, the predictive ability of AI also needs to be iterated. Mu Mu believes that there are still two major shortcomings in the current development of meteorological models. The operational logic of a large model involves feeding and outputting data, but the physical mechanism behind the predicted results is still unclear. In addition, the time resolution of the large model is not high enough. For example, some large models can only refresh typhoon data every 6 hours, while traditional numerical models only need a few minutes. "Mu Mu also gave an example that in a study conducted by a foreign research team, an AI model driven by meteorological data was unable to observe the" butterfly effect "commonly seen in traditional numerical models. Mu Mu analyzed that this should be attributed to the AI model's low spatiotemporal resolution and the mismatch between the magnitude of small perturbations and the original magnitude of the training set. Coincidentally, in the opinion of Zheng Weimin, an academician of the CAE Member and a professor of Tsinghua University, AI is still unable to solve the problem of accurate calculation. The imperfections of AI can be compensated for by high-performance computing (HPC). The new generation of weather forecasting should be a combination of HPC and AI. Multiple types of supercomputing are needed to deeply integrate AI with traditional scientific computing solutions, break the bottleneck of computing accuracy, and significantly improve overall computing performance. ”Zheng Weimin said. Numerical simulation of the Earth system is used by scientists to simulate the current state of the Earth and predict future climate change. Zheng Weimin stated that computing power drives the evolution of Earth system simulation accuracy, and conducting research on Earth system simulation also raises higher expectations for HPC. Zheng Weimin suggests that in the future, heterogeneous and multi-core supercomputing systems should be established. Faced with massive computing power, multi million core parallelism will become an important capability for simulating Earth systems. At the same time, it is necessary to establish a supercomputing system with high memory and large bandwidth, and to design computing hardware and algorithms in synergy to effectively combine low precision and high-precision computing. (New Society)