? Release of "Ten Application Challenges of New Generation Domestic E-Class Supercomputing System"
According to reports, the research and development of a new generation of tens of billions (E-class) high-performance computers is an important deployment of the country in the next generation of information technology, which will effectively promote the innovation and development of the national information technology industry, and the degree of autonomy is much higher than other supercomputing platforms. At the same time, the scale and performance have been greatly improved. The purpose of the release of the top ten application challenges is to give full play to the powerful computing power of a new generation of E-class high-performance computers, develop key technologies and application software adapted to domestic supercomputing systems, and build a new ecosystem of domestic E-class supercomputing applications.
Challenge 1: Full-scale fusion simulation of magnetic confinement fusion reactor (artificial small sun)
Controlled fusion energy is considered as one of the most effective means to deal with energy and environmental challenges in the future. In the design of magnetic confinement fusion device, the most important problem is how to confine the plasma with higher density and temperature at lower cost.
At present, it is generally believed that one of the main mechanisms to limit the performance of constraints is the so-called microscopic instability generated by boundaries and cores. In previous studies, due to multi-time scale problems, these instabilities are usually described by simplified models such as cyclotron dynamics or magnetohydrodynamics, and the plasma core and boundary need to be modeled separately.
On the new generation of domestic E-class computers, it will be possible to use the basic model of classical plasma to directly simulate the plasma of the whole device of magnetically confined fusion reactor with the resolution of ion cyclotron radius, without distinguishing the boundary from the core. With the help of geometric algorithm, the reliability of the simulation results of the long-term evolution of the system can be guaranteed, and the unstable process inside it can be reproduced more accurately and consistently, and the mechanism of improving the confinement performance of the magnetic confinement fusion device can be sought.
In addition, the plasma dynamics simulation of the whole device can also obtain a more realistic device-scale plasma evolution model, which will better guide the plasma design of the future magnetic confinement fusion reactor and provide strong support for controlled fusion research and fusion energy development.
Challenge 2: Computational Fluid Dynamics Simulation of Full-scale Spacecraft on a Billion Grid
In recent years, the numerical study of complex flow problems of near-space vehicles has played an important role in understanding the mechanism of high-speed flow at high altitude. The flight envelope of adjacent spacecraft covers continuous basin, sliding basin and transitional basin, and there are complex unsteady multi-field coupling phenomena such as aerodynamic heat, rarefied non-equilibrium effect, multi-body separation under high dynamic pressure, chemical reaction and plasma.
The previous generation of supercomputers can't simulate the unsteady multi-field coupling of near-space vehicles that meet the requirements of accuracy and efficiency across river basins in terms of computing power and architecture design, but the new generation of domestic E-class supercomputers is expected to break through the computing bottleneck in theory, thus realizing high-precision full-scale simulation of over 10 billion grids.
On the one hand, the research on unsteady multi-field coupling simulation of inter-basin near space vehicles can help us fully understand the complex coupled flow phenomenon of aircraft in high-altitude and high-speed flight, and identify the flow mechanism and its influence on aircraft. On the other hand, it can expand the ability to build a numerical wind tunnel on a new generation of supercomputers, and provide a carrier for the design of various space launch vehicles and domestic large aircraft, thus better serving China's strategic development and construction.
Challenge 3: Dynamic Simulation of Digital Cell Superbillion Atomic System
In understanding the mysteries of life, especially the biological mechanism of cells, the virtual experiment of all-atom molecular dynamics simulation plays an increasingly prominent role.
Cells are full of all kinds of biological macromolecules and small molecules, and they are always in a dynamic crowded environment, which will have a great impact on the diffusion, aggregation, conformational changes and chemical reactions of biomolecules. Therefore, it will be an important means to simulate the whole atomic and molecular dynamics of biological systems on a cellular scale, but it is difficult for current computers to provide enough computing power to realize it.
A new generation of E-class supercomputers made in China can theoretically realize the molecular dynamics simulation of digital cells with billions or billions of atoms, which makes it possible to conduct accurate simulation at the cell level in the future. Full-atom molecular dynamics simulation of cells will enable us to observe the micro-dynamic processes inside and outside the whole cell and its biological macromolecules with high temporal and spatial resolution in virtual experiments, and help us fully and profoundly understand the important challenges of life science such as how COVID-19 and others invade cells and how cells exchange material information.
The knowledge gained through digital simulation will also play a fundamental role in the future research and development of new drugs and the protection of life and health.
Challenge 4: Sub-kilometer fine numerical weather forecast at convective scale.
Small-scale and fast-developing severe convective weather systems are often difficult to predict, which is easy to cause destructive disasters and poses a great threat to the operation and management of big cities.
With the shortening of the time-space scale of weather system, the chaos of the atmosphere becomes more and more obvious, and the uncertainty of forecast intensifies, which brings great challenges to the fine weather forecast. At present, the timeliness of early warning based on detection technology and the prediction of system evolution are often insufficient.
Based on the powerful computing power of a new generation of domestic E-class supercomputers, by integrating ultra-high resolution simulation, ensemble forecasting and fast circulation technology, the triggering, evolution and extinction of strong convective weather can be predicted 0-6 hours in advance, which provides continuous and probabilistic prediction for a single convective system and improves the prediction accuracy of local heavy precipitation, hail, sudden gale and tornado caused by strong convective weather.
Challenge 5: Screening virtual drugs with tens of billions of high-efficiency Qualcomm.
The discovery and optimization of lead structure, as the research core of new drug discovery stage, often takes several years and hundreds of millions of dollars, which is the key technical bottleneck of drug research and development. Therefore, how to generate new molecules and optimize the key properties of molecules (such as biological activity, drug formation, safety and selectivity) are two key issues that affect the success or failure of drug molecular design.
It is estimated that the available chemical space range is 10 23 ~ 10 60, and the number of small molecules in some mature databases reaches one billion. How to intelligently generate molecules in such a huge chemical space, search for the rapid evolution of structures and predict their properties is a huge challenge for drug screening.
At present, relatively rough methods such as molecular docking can usually be used for drug screening. Previous supercomputers can quickly screen out billions of small molecules, and then use more accurate free energy perturbation calculation and other methods to make more accurate evaluation and analysis.
The powerful computing power provided by the new generation of E-class supercomputers can support the rapid screening of tens of billions of small drug molecules, and with more advanced algorithms, the efficiency of virtual drug screening can be improved by dozens or even hundreds of times. At the same time, efficient drug screening can also be combined with the discovery of effective components of traditional Chinese medicine to promote the modernization of research and development of traditional Chinese medicine.
Challenge 6: General artificial intelligence super-large-scale pre-training model
Deep neural network is the leading field of a new generation of artificial intelligence, which has been successfully applied to computer vision and natural language processing, and has achieved outstanding results.
With the enrichment and development of application scenarios, the traditional domain model training and application paradigm based on labeled data sets are increasingly unsuitable for the development and popularization of artificial intelligence applications. The emergence of self-supervised learning technology based on unlabeled data and large-scale pre-training model with good comprehensive ability and strong general ability has pushed data-driven deep learning technology and general artificial intelligence to a new development stage.
In recent years, experts and enterprises in the field of computer and artificial intelligence have completed the development of multi-modal pre-training model with parameters of 1.75 trillion on existing high-performance computers. The new generation of E-class supercomputers makes it possible to train and apply models with larger support parameters and stronger versatility (exceeding 10 trillion or even 10 trillion).
On the one hand, the development and deployment of large-scale pre-training model will accelerate the landing of humanoid robots; On the other hand, having a universal model as the foundation will greatly reduce the difficulty of transforming data in subdivided fields into intelligent models, effectively promote the construction of artificial intelligence application infrastructure, and enhance the industrial modernization, digital economic development and digital governance capabilities of intelligent society.
Challenge 7: High-resolution sky survey image processing of fast and ultra-large-scale observation data
Measuring neutral hydrogen in the sky is one of the important scientific goals of the 500-meter spherical radio telescope (FAST). By detecting the distribution of neutral hydrogen in Hubble volume, it provides support for the research in frontier scientific fields such as the origin and evolution of the universe, dark matter and dark energy.
Limited by the field of view, the telescope can only cover a limited sky area at a time, and the neutral hydrogen survey can last for several years. The accumulated observation data need to be spliced and fused to obtain a complete high-resolution survey image. In the process of neutral hydrogen measurement data processing, gridding is the most intensive part of calculation and I/O, which is the bottleneck restricting the efficiency of neutral hydrogen measurement data processing and result output.
The data processing ability of a new generation of E-class supercomputer, combined with high-performance grid algorithm, can cope with PB-class neutral hydrogen survey observation data, thus providing a powerful boost for the national heavyweight 500-meter spherical radio telescope (FAST) to "produce results early, produce more results, produce big results and produce good results" and promote major discoveries in the basic and frontier fields of astronomy.
Challenge 8: global seismic full waveform inversion
Seismic full waveform inversion is the highest resolution imaging method at present, which is a powerful tool to study the internal structure and dynamic evolution of the earth, and can also provide key support for mineral resources and oil and gas exploration.
In recent ten years, experts in the field of earthquake science have realized the full waveform inversion of low-frequency elastic waves at regional scale on the last generation of supercomputers.
A new generation of E-class supercomputer made in China will be able to study the propagation simulation and waveform imaging of high-frequency viscoelastic seismic wave fields, including the attenuation characteristics of seismic waves. On the one hand, global scale high-frequency viscoelastic seismic waveform inversion can obtain high-precision imaging results inside the earth, deepen our understanding of the formation and evolution of plate tectonics, subduction zones and orogenic belts, on the other hand, it can provide seismological evidence of material and energy exchange in various circles (middle and lower crust, lithosphere, asthenosphere, etc.) inside the earth. ), which provides a basis for studying the deep mineralization and volcanic/seismic activities of the earth, and helps human beings to understand the geological evolution and the formation and development of terrestrial planets more comprehensively.
Challenge 9: Dynamic simulation of 100 billion neurons in the whole brain
In recent years, neuroscience research has obtained a lot of data about brain structure and activity to understand the working mechanism of the brain. For the analysis and reproduction of advanced brain functions such as motor control and thinking, it is urgent to establish a simulated neural network platform on the scale of human brain.
There are 86 billion neurons in the human brain. In the past decade, experts in the field of computational neuroscience have simulated the brain circuits of the previous generation of supercomputers. The new generation E-class supercomputer can theoretically simulate the whole brain circuit including cerebral cortex, cerebellum and basal ganglia.
On the one hand, the simulation and research of human brain circuits can help us understand the advanced functions of brain thinking and develop brain-like artificial intelligence algorithms. On the other hand, the pathogenesis of brain diseases such as Parkinson's disease and Huntington's disease will be further analyzed and verified. In addition, using the simulated brain model to build a neuromodal robot based on impulsive neural network can improve the perception and decision-making level of the existing robot system.
Challenge 10: full-resolution global mesoscale ocean numerical simulation
Climate change is a major challenge to global sustainable development, and it is also one of the most challenging scientific problems faced by the scientific community. The ocean is an important part of controlling the seasonal, seasonal, interannual and interdecadal changes of the climate system.
In recent years, with the rapid development of ocean observation, many mechanisms of ocean mesoscale and sub-mesoscale processes have been continuously revealed, and the characteristics of ocean multi-scale interaction have become clearer, which also puts forward higher requirements for numerical simulation of ocean circulation. Resolution mesoscale and submicroscale processes and their interactions with the atmosphere have become an important research direction.
10 years, scientists all over the world have made unremitting efforts in this direction, improving the global ocean simulation to the resolution of sub-mesoscale vortex (2km) with partial resolution, while scientists in China have independently developed a global ocean model of 3-5km, which can basically distinguish the mesoscale processes in the open ocean. A new generation of E-class supercomputers can simulate mesoscale processes with full resolution, help scientists fully understand the multi-scale interaction process and the cascade process of ocean energy, and further improve the simulation ability of ocean circulation and the whole climate system.
Editor/Fan Hui