Chuanguan News reporter Wang Guoping
"This is the world's first large-scale seismic wave model with 100 million parameters, which was officially released today." On July 28, the "Diting" seismic wave model jointly developed by the National Supercomputing Center in Chengdu, the Institute of Geophysics of the China Earthquake Administration and Tsinghua University was officially released.
Du Bin, deputy director of the Sichuan Provincial Earthquake Administration, who attended the event, said that Sichuan, as a major economic and energy province with complex geological structure, is one of the regions with the highest level and intensity of seismic activity in China. The release of the "Di Ting" seismic wave model is not only a major breakthrough in earthquake science and technology, but also a positive response to the major strategic needs of the country.
The site of the release of the large model of the seismic waves.
How to practice the "divine beast" of listening to shock?
"'Di Ting' was originally a mythical beast in Chinese myths and legends, and everything in the world can be identified by listening to sounds." Chen Shi, deputy director of the Institute of Geophysics of the China Earthquake Administration, said, "The 'Diting' we developed is a large seismological dataset and a large artificial intelligence model trained on it. ”
At the press conference, Chen Shi explained the origin of "Di Ting": they are all through "listening" to understand and analyze information in nature.
Chen Shi said that with the continuous optimization and upgrading of the mainland earthquake monitoring system and the rapid development of artificial intelligence technology in recent years, seismological research and earthquake prevention and disaster reduction work have also entered the era of seismic wave big data.
So how to train this "divine beast" that can recognize earthquakes?
The first thing to do is to have a sufficient amount of data, which is the "Listening" dataset. Initially, the research team cleaned and desensitized the seismic phase observation reports from 2013 to 2020 and the data from the National Seismographic Network Data Backup Center to establish version 1.0 of the "Diting" dataset. With the expansion of the scale of data, in September 2023, the National Supercomputing Chengdu Center operated by Chengdu Industrial Group and the Institute of Geophysics of the China Earthquake Administration reached a strategic cooperation, and the two sides cooperated to build the "Joint Laboratory for Innovative Application of Earthquake Large Model", and the new generation of "Diting" dataset was officially settled in Chengdu.
The dataset is the basis for large model training, and the size and quality of the dataset directly determine the training effect of the model. At present, the "Di Ting" dataset is not only the first in China, but also one of the largest seismology professional AI training datasets with the most comprehensive sample types and annotations at home and abroad.
According to Chen Shi, the first version of the "Di Ting" seismic wave model was pre-trained based on 23 million seismic event waveforms, while the mainland China seismic network generated about 500,000 annotated event waveforms every year. Based on this calculation, the "Diting" seismic wave model that has just been "born" is equivalent to an "old expert" with about 40 years of experience in earthquake signal recognition.
The second is the design and tuning of algorithms and models. Dr. Liu Chang, an assistant researcher from the Department of Automation at Tsinghua University, said that as the first exploration of seismic wave models in the world, a lot of basic research is needed, otherwise the model training is not sufficient, and the practical application effect is not as expected.
To put it simply, the selection and design of the algorithm directly determine the learning ability and expression ability of the model, and the optimization and improvement of the algorithm can improve the performance and effect of the model. Parameters are the regulators of large model training, which can directly affect the accuracy and stability of the model. In the "Diting" large model, the number of parameters is very large, which also poses a great challenge to the confirmation of the optimal parameters.
"The key to AI algorithms lies in two factors, the amount of training data and the amount of model parameters." Chen Shi used an analogy to explain: the amount of training data can be regarded as the "experience" of the algorithm, and the amount of parameters is the "brain capacity" of the algorithm. In order to effectively memorize and understand the massive amount of seismic data, and fully explore and utilize the information in it, it is necessary to develop a large model with a large "brain capacity" to match it.
What can you do with "Di Ting"?
It is reported that at present, the "Di Ting" seismic wave model has been put into use. Chen Shi said that relying on the massive data of China's seismic network and advanced artificial intelligence technology, it was found that the recognition accuracy and speed of seismic signals can be significantly improved.
Therefore, in the short term, the "Di Ting" seismic wave model can be directly applied to seismic signal recognition, seismic activity monitoring, rapid response to large earthquakes and other fields, which is expected to reduce the work pressure of front-line business personnel of the Earthquake Bureau.
In the long run, seismology is an observational science, and major breakthroughs often come from a deep understanding of observational data. The more comprehensive the understanding of observations and the greater the ability to integrate observations, the closer we are to breakthroughs in the scientific problems of seismology. At present, traditional methods and small and medium-sized models cannot make full use of hundreds of terabytes and petabytes of seismic observation data, and these data contain many important seismological scientific problems, and only large models can dig deep into these "treasures". Therefore, the "Diting" seismic wave model is expected to bring a major breakthrough to earthquake science research.
From the perspective of application fields, the prospect of the "Di Ting" seismic wave model is also very exciting.
"In the future, the application scenarios of this model can also be used in many fields such as mine earthquake monitoring, shale gas extraction, urban underground space structure detection, and seabed seismic monitoring." Wang Jianbo, executive deputy director of the National Supercomputer Chengdu Center, said.
Take, for example, the search for oil and gas. At present, more than 95% of the world's oil and gas field discoveries mainly rely on seismic exploration. Wang Jianbo said that when seismic waves propagate in different media, the intensity, morphology and other characteristics are different, and the "Diting" seismic wave model can deduce whether there is oil and gas underground by learning the waveform characteristics of the oil storage area.
Chen Shi said that the first version of the "Di Ting" seismic wave model completed the exploration of the complete process from pre-training to fine-tuning, and initially demonstrated the performance better than the small and medium-sized models, which explored the way and accumulated experience for the subsequent full mining and giving full play to the advantages of the seismic wave model.
What else can supercomputing do besides "listening"?
"The development of seismic wave models is not only an inevitable trend in the field of artificial intelligence seismology, but also the commanding heights of science and technology in this field." Chen Shi said, "After having sufficient data and computing power, we quickly organized a team to carry out relevant research work as soon as possible, striving to fill this important gap and contribute to earthquake scientific research and earthquake prevention and disaster reduction." ”
Guo Li, Deputy Secretary of the Party Committee of Chengdu Data Group and Chairman of Chengdu Supercomputer Center Operation Management Co., Ltd., said that the research and development of the "Diting" seismic wave model not only greatly promotes the development, testing and application of artificial intelligence algorithms in the field of seismology, but also provides solid technical support for the intelligent development of earthquake monitoring and forecasting business.
As a senior practitioner who has been working in the field of supercomputing for a long time, Wang Jianbo believes that the "Di Ting" seismology dataset has officially settled in the National Supercomputing Chengdu Center Earthquake Large Model Innovation and Application Joint Laboratory, which is a new attempt to deeply integrate professional data and large-scale computing power in the vertical field, and integrate software and hardware.
"The release of the 'Di Ting' seismic wave model is of great significance for breaking through the performance bottleneck of small and medium-sized seismic wave models and improving the intelligent processing ability and information mining level of seismic big data." Wang Jianbo said that the strong computing power and technical service capabilities of the National Supercomputer Chengdu Center can not only meet the customized service needs of scientific research units such as the Institute of Geophysics of the China Earthquake Administration, but also provide a solid platform guarantee for the development of future industries such as artificial intelligence and robotics.
At present, the National Supercomputing Center in Chengdu has formed characteristic computing power applications in many fields. Among them: in the field of artificial intelligence large model, the National Supercomputing Chengdu Center has cooperated with Chengdu Xiaoduo Technology to develop a large language model for the vertical e-commerce customer service industry developed based on large language model technology, which can provide intelligent service and marketing integration solutions for e-commerce enterprises. In May this year, the model successfully passed the national generative AI service filing.
National Supercomputer Chengdu Center.
In the field of disaster prevention and mitigation, the Chengdu Institute of Mountain Disaster and Environment, Chinese Academy of Sciences, together with the National Supercomputer Chengdu Center, has carried out a project research on "Key Technologies and Demonstration of Green Regulation of Mountain Disaster Risks under Climate Change Conditions", built a mountain disaster risk simulation and danger forecasting platform, and broke through the bottleneck of refined and accurate disaster early warning and forecasting. In addition, the National Supercomputer Center in Chengdu also cooperates with the Sichuan Meteorological Department to provide accurate weather forecasting services.
"The National Supercomputer Chengdu Center will continue to accelerate the application and development of artificial intelligence technology in disaster prevention and mitigation, urban governance and other fields, and give full play to our strategic supporting role as a major country." Wang Jianbo said.
Courtesy of the National Supercomputer Chengdu Center