Meng Chen from Aufei Temple
By Quantum Bit | QbitAI
AI for Science, when did this concept catch fire?
A trigger point, in July 2021, AlphaFold2 open-sourced and predicted 98.5% of human protein structures in one breath, showing the world the huge potential of AI to solve real-world problems in the field of scientific research.
At the end of the year, when major organizations release their year in review and future predictions, AI for Science and its acronym, AI4S, are in the forefront of people's minds in one exposure after another.
On the one hand, data-driven is touted as the next paradigm for scientific research. On the other hand, scientific research applications are also seen as the new battleground for AI landing.
Now that almost a year has passed since AlphaFold2, how is AI4S progressing?
In addition to DeepMind, which is backed by Google and has a lot of money, continues to shine, producing results such as AI-assisted mathematicians proving theorems and AI-controlled nuclear reactors, there is another trend that deserves attention.
There has been an explosion of open source tools and innovations generated based on open source tools, and AI4S research has expanded to include more fundamental problem areas.
To study water, there are more than ten phase transitions of water in the interval from zero temperature and pressure to 2400 K and 50 GPa predicted with first-nature principle accuracy by means of DP-GEN depth potential energy generation tool.
This result was featured in the top physics journal Physical Review Letters and selected as an "Editors' Suggestion".
To study fire, there is DeepFlame, an open source project that combines heterogeneous parallelism with AI gas pedals to build hydrodynamic computational tools for combustion reactions.
The first version of this project has been released and has been adapted to several domestic ARM architecture chips.
More recently, there has been research to build bridges between first-principles data and machine learning potential functions based on a series of open-source frameworks.
The ceiling of molecular dynamics simulations in terms of scale and accuracy is significantly improved.
The problems that these new research methods and open source tools are designed to solve are usually not close to the lives of the general public and lack the influence of celebrity companies, so they often fail to ignite the conversation and hit the hot topics.
However, for the researchers concerned, these are important issues that affect the direction and efficiency of their research.
Beyond the spotlight, a group of people actively exploring the integration of machine learning and physical modeling are gathering in an open source community called DeepModeling.
DeePMD-kit is the earliest and most influential project in the community.
It raises the limit of molecular dynamics to the scale of one billion atoms while maintaining high accuracy through a combination of machine learning, high-performance computing, and physical modeling.
This achievement was awarded the Gordon Bell Prize in 2020, which is known as the Nobel Prize in supercomputing, and was also selected as one of the top ten technological advances in China in 2020, together with "Chang'e 5 achieved the first automatic lunar surface sample return" and "artificial sun".
By now, the DeepModeling community has grown into an open source platform across multiple disciplines.
Other projects developed include ABACUS, which focuses on condensed matter simulation calculations, DeePKS-kit, a general machine learning framework for building accurate and efficient density generalized models, and FEALPy, a library of algorithms for the numerical solution of partial differential equations.
Just this recent month, they have also launched four new projects intensively.
These are DeepFlame, the aforementioned combustion reaction fluid computing platform, DMFF, a production-grade differentiable force field computation engine, dflow, a framework designed to co-create AI4S and scientific computing workflows in the cloud-native era, and AI4Science101, a pedagogical documentation project that hopes to help all those interested in AI4S gain a quick understanding of the field.
The main driving force behind the rapid development is a new type of research institute, which can be seen from its name as being made for AI4S.
AI for Science Institute, Beijing (AISI).
The Institute for AI for Science
AISI is a young institute, only officially established in September 2021, but has already made quite a name for itself.
In addition to his research work and pushing to build the DeepModeling open source community, he also pushed to found the new Journal of Machine Learning (JML).
JML hopes to be an ideal platform for academic communication in the early stages of AI for Science's development, in addition to traditional disciplinary journals and conferences in the field of machine learning.
Why is this institute focusing on the direction of AI for Science?
We also need to start from the president, Academician Weinan E.
Weinan E has been engaged in applied mathematics and scientific computing research for many years, and was elected as a member of the Chinese Academy of Sciences in 2011.
Since the 1980s, he has been promoting multi-scale models to solve multi-body problems, drug and material design, turbulence and non-Newtonian fluid dynamics, among other challenges.
In these problems, however, there is a long-standing "dimensional catastrophe".
The dimensional catastrophe was first proposed by Richard Bellman, the founder of dynamic programming.
means that as the number of dimensions grows, the amount of computation required to analyze data in high-dimensional spaces grows exponentially.
For example, to achieve the same sampling density in a high-dimensional space, the number of sampling points required increases exponentially, which makes it difficult to adapt classical computational methods to the study of complex problems.
Because of a chance attempt, Wei-Nan E's group obtained an acceleration of 5-6 orders of magnitude when trying to use AI-assisted molecular dynamics simulations, which made Wei-Nan E start to realize that deep learning is the perfect tool to solve the dimensional catastrophe, because the essence of deep neural networks is in approximating high-dimensional functions.
For example, image recognition is the conversion of image content into high-dimensional vectors through feature extraction, and AlphaGo plays Go by solving equations that satisfy the optimal strategy in a high-dimensional space.
The foundation of AI for Science is the ability to apply deep learning to a variety of scientific problems.
Guided by this idea, Eweinan led a team in 2017 to propose a deep potential energy(Deep Potential)Molecular dynamics methods, using deep learning methods to calculate the potential energy functions of the interactions between hundreds of millions of atoms.
Schematic of the Δ-depth potential energy method from Phys. Rev. Lett. 120 (14), 143001
DeePMD, which was later awarded the Gordon Bell Prize, and the DeepModeling open source community were developed on top of this.
Ewenan's thoughts and actions for the later establishment of the Institute also started from the end of 17.
His student and later Vice President of AISI, Zhang Linfeng, considered Mr. E not only as a scholar but also as a flag bearer.
In FY18, Weinan E gathered scholars from many directions at Peking University to discuss AI for Science.
This is probably the first time that the topic of AI for Science has been discussed on a large scale in the world.
Later, he also called for "science as the next main battlefield of AI" on several national and international occasions.
Including an opinion piece entitled "The Dawn of a New Era in Applied Mathematics" published in the Journal of the American Mathematical Society, which gained wide attention.
He returned to China from Princeton in September 2020 and began assembling a team to prepare for the establishment of the Institute.
One year later, the AISI Institute was finally officially launched.
In July this year, Ewenan was also invited to speak at the quadrennial International Congress of Mathematicians(ICM)He will be giving a one-hour presentation at the Fields Medal event.
Only 21 mathematicians in the world have received this honor, and Weinan E will be the third person from mainland China.
In his new role as President of AISI, he will also continue his call for the promotion of AI for Science to mathematicians around the world.
Why do we need to operate as an institute when we already have an open source community?
First, the new paradigm brought about by AI for Science requires a tight combination of machine learning, high-performance computing, and physical models, one without the other.
This requires a name to bring people of all backgrounds together.
In general, there is always a shortage of engineers in schools, and companies have difficulty recruiting a large number of scientists.
The Independent Institute, then, has become a more effective frontier for conducting AI for Science work.
With the support of a physical research institute, it is also easier for the open source community, a non-entity organization, to recruit members, attract investment, and publish research results.
Since its establishment, AISI has gathered a core group of members from top universities, research institutions and enterprises in cross-disciplinary disciplines at home and abroad.
At the current stage, AISI aims at AI for Science infrastructure building and frontier scientific problem exploration, to build platform-based tools for scientific development.
If we look at the longer term, AISI also hopes to move the entire scientific research from a "smallholder workshop model" to a platform-based "Android model".
A new model of platform-based research
How to understand the "smallholder model" of research?
On the one hand, it means that scientists are working alone in their own labs with closed-source ancestral code that is difficult to modify and also prone to monopoly dominance.
On the other hand, it also refers to the lack of automated processes in research, which requires a large number of experienced personnel to do manual interventions.
Weijie Sun, a strategic development consultant at AISI, believes that researchers in the traditional model are simply ...... too miserable.
The Internet and AI industries have been "taking off" in recent years, with practitioners enjoying a clear division of labor and a well-established infrastructure of development tools.
In contrast researchers are still dealing with inefficient work environments and collaboration models.
The "Android model" can be understood as a large-scale production of scientific research, drawing on the successful experience of Linux, Android and other platforms, using open source to bring a snowball effect, gathering talent, data, algorithms and application scenarios, to accelerate scientific innovation.
Compared with AI in security, medical and other scenarios where the demand is more concentrated, AI for Science has a more dispersed demand in various disciplines and will experience a longer chain.
So in this "Android model", both the open source community to provide tools, research institutes to do innovation, but also the need for companies to develop these innovations into products.
For example, Deep Potential Technology, a company founded by core members of AISI, has created a micro-scale industrial design platform using the molecular dynamics method DeePMD.
Among them, the drug design platform Hermite has been used by a number of drug company developers, and not long ago also successfully replicated AlphaFold2 to create the open source Uni-Fold, integrated into the platform to solve the data source of protein structures.
As a result, DeepModeling open source community, AISI Research Institute, Deep Potential Technology and other companies, as well as more partners in the academia and industry, together form a complete ecological chain from innovation to implementation under the AI for Science platform-based research model.
So why hasn't there been a big open source platform in traditional scientific computing before AI for Science?
AISI Vice President Linfeng Zhang believes there are three reasons for this.
First, for historical reasons, ancestral codes play a key role, despite the fact that communication and connectivity between scientists has become extremely easy in the Internet era.
Then there is the nature of research looking at innovation and the academic evaluation system does not encourage the formation of a platform, publishing in top journals will bring great prestige to the researcher, but it is difficult for the developers of open source tools to benefit directly.
Finally, there are the new changes that AI has brought to the field of scientific computing.
With the involvement of AI will inevitably require a new infrastructure, including both the three elements of AI data, algorithms and computing power, but also the models and tool chains built on top of this.
And the natural intersection of AI practitioners and IT staff also brings the gene of platform thinking to AI4S.
Of course, there is still a need for specific people to promote this, it is impossible to say that the entire field one day we all awakened to the spontaneous formation of a platform.
The DeepModeling open source community, as well as the AISI Institute, are meant to be the group that acts first.
Whether it is the study of specific cross-cutting issues or the more ambitious drive for paradigm change in scientific research, more fellow travelers are needed.
AISI is currently looking for talented and creative scholars with excellent research skills in cross-cutting areas to join us.
Recruitment areas include electronic structures, molecular dynamics, computationally assisted materials design, computational fluid dynamics, combustion algorithms, high performance scientific computing, etc. You can click the link below or read the original article to learn more.
AISI Beijing Institute of Scientific Intelligence.
DeepModeling open source community.
- End -
Quantum Bits QbitAI - Headlines Signed
Follow us to be the first to know the cutting-edge technology news