Old Version
Special Report

Brain Trust

As AI accelerates the efficiency of research, experts cast doubt on its ability to trigger another scientific revolution – and take their jobs

By Peng Danni Updated Aug.1

AI technology is penetrating almost all sectors of research, from labs to libraries (Photo by VCG)

FengWu, an AI weather forecast model released in April by Shanghai AI Laboratory and its partners, is capable of reducing erroneous predictions for 10- day forecasts by 19.4 percent compared to traditional models.  

Named after China’s earliest meteorological instrument from the Han Dynasty (202 BCE-220 CE), FengWu outperformed GraphCast, the latest AI model from Google’s DeepMind, on 80 percent of indicators concerning accuracy and efficiency, Chinese news website The Paper reported on April 7.  

The new model extended medium-range weather forecasts from 10 to 10.75 days. It can calculate a 10-day forecast in 30 seconds, half the time of GraphCast.  

When it comes to AI applications to science, this is just the tip of the iceberg.  

“These innovations show the long-lasting influence AI is now only beginning to have on scientific discovery, from protein structures to climate modeling and gravitational waves detection to understanding the universe,” read an article from science data institute Dataconomy on November 9, 2022.  

To ride the tide of AI applications, China’s Ministry of Science and Technology joined the National Natural Science Foundation of China (NSFC) in March for “AI for Science,” a nationally funded project aiming to further integrate AI in fields such as mathematics, physics, chemistry and astronomy, the Xinhua News Agency reported on March 27.  

Many scientists welcome the move, confident that despite AI’s strengths, it still lacks the capability of deep, abstract and critical thinking that humans bring to the advancement of science. 

Quick Combos 
Material scientist Liu Miao told NewsChina that he never would have expected to complete more than four research projects during his doctoral program 10 years ago. Today with AI, Liu and his team can screen hundreds of thousands of possible element compounds with the click of a mouse.  

In 2018, Liu and his teammates created a material database named Atomly to explore possible chemical compounds.  

“With a substantial database and rapid computing abilities, scientists can find their intended chemical compounds and learn their possible physical natures,” Liu said.  

For instance, choosing oxygen and titanium in the Atomly database shows 280 possible compounds, including their mechanical properties.  

“It’s like fishing with a huge net instead of a single hook,” Liu said.  

On March 8, assistant professor Ranga Dias and his team at the University of Rochester in the US announced their discovery of a new room-temperature superconductor made from hydrogen, nitrogen and lutetium, a rare earth element.  

The announcement made a huge splash in academia.  

However, 13 days later Liu and his team challenged the discovery, arguing in a paper that Atomly failed to formulate any stable structures with the three elements in more than 1,500 possible compounds. “It was unbelievable that we could challenge the discovery so quickly,” Liu said.  

AI technology is penetrating almost all sectors of research, from laboratories to university libraries.  

An AI created by the University of Liverpool invented a photocatalyst through 688 experiments conducted over eight days, Nature reported in July 2020.  

In Shenzhen, XtalPi, a biopharma-ceutical firm established in 2014 by three Chinese chemists who graduated from the Massachusetts Institute of Technology, has applied a blend of quantum physics and AI to lab experiments and data analysis for researching molecular drugs or facilitating automatic drug synthesis.  

Using Xtalpi’s AI algorithm, Pfizer confirmed its design of the Covid-19 specific medicine Paxlovid was optimal in terms of its crystal structure, an arrangement of ions, atoms and molecules in a solid substance.  

Xtalpi shortened the process from several months to six weeks, CN-Healthcare reported on September 22, 2022.  

“The AI can calculate with 400 computers, explore thousands of reaction conditions and screen numerous catalysts while running 24 hours a day,” Wen Shuhao, co-founder of Xtalpi, told NewsChina. “It’s simultaneous and large-scale operation is far from the reach of any laboratory technicians.”  

AI has not only been employed in scientific research but also in coding and writing. OpenAI’s rollout of ChatGPT boosts efficiency in searching, editing, grouping and summarizing literature across a wide spectrum of disciplines. “It’s like a versatile student of liberal arts particularly good at writing,” Xu Bo, director of the Institute of Automation, the Chinese Academy of Sciences, told NewsChina.  

Although ChatGPT cannot read an article as incisively as a well-trained scholar, it is much more efficient than humans in data selection and analysis, Pei Jianfeng, a research fellow at the Center for Quantitative Biology, Peking University, told NewsChina. “China can take advantage of ChatGPT’s strong language processing capability to beef up its knowledge and data systems, both of which have not gotten enough attention over the past few decades,” Pei said. 

AI in China 
In 2022, the number of Chinese AI firms, including the producers of smart chips, sensors and intelligent connected vehicles, reached 4,200, with AI core industries scaled up by 18 percent to 508 billion yuan (US$71.88b), according to the China Academy of Information and Communication Technology.  

The industry is expected to boost interdisciplinary scientific research through multi-scale modeling, high-precision emulation and differential equations, Xu said, adding that AI for Science is providing AI modeling and algorithms to major scientific programs.  

Nowadays, scientific research in China cannot be financed easily without employing AI, a theoretical scientist told NewsChina. “The absence of AI in research may engender strong skepticism for a scientist’s work efficiency and comprehension ability,” said the expert, who requested anonymity.  

AI was first adopted in chemical and pharmaceutical research in 2012, Pei told NewsChina. In 2015, following a yearlong effort, Pei and his team published the country’s first dissertation on AI’s application to drug designs, bringing AI and science integration to the forefront.  

In November 2020, three years before the AI for Science program, the NSFC established a department of interdisciplinary studies that prioritized AI application in science as its first major research program, Tang said.  

The same year, the industry burgeoned with start-ups dedicated to computing and modeling, which laid an important foundation for AI’s deep learning capability beyond academia. 

“The integration between AI and science actually highlights the role of engineering, which has contributed a lot to major scientific breakthroughs such as AlphaFold’s prediction of protein structure,” Pei pointed out. According to him, though the independent discoveries in specific fields remain important, interdisciplinary cooperation will be increasingly adopted in the AI era to take on more complex and pressing problems.  

AI provides new approaches to address conundrums such as atmospheric simulation and seismic surveillance, Ouyang Wanli, lead scientist of Shanghai AI Laboratory told NewsChina. “Take FengWu for example. Two to three years ago, we wouldn’t have expected to find solutions to the most difficult parts of medium-range weather forecasting,” Ouyang said. 

Human Advantages 
In July 2022, AlphaFold announced its predictions on 214 million protein structures among more than 1 million species, including 98.5 percent of human proteins. Until then, scientists had discovered only 17 percent of the structures in a human protein sequence. This stark contrast highlights the coming dominant role of AI in research.  

“When a technology is able to override a threshold, it brings brand new changes to research both on what to study and how to work it out. This is what AI for Science is ramping up for,” Wen said.  

According to Xu, AIs like ChatGPT are able to inspire researchers to work on more innovative designs and with brand new approaches.  

Meanwhile, Tang said that deep learning will not only lead AI to solve specific scientific problems but also the theorems behind them.  

However, there are still many who do not buy into the notion that AI will trigger a new scientific revolution.  

In a 2009 article for Nature, Nobel Prize winning physicist Philip W. Anderson wrote: “Even if machines did contribute to normal science, we see no mechanism by which they could create a Kuhnian revolution and thereby establish new physical law,” referring to American philosopher Thomas Kuhn’s argument that scientific breakthroughs occur in sudden paradigm shifts rather than linear accumulations of knowledge.  

AI has its limits in chemistry and pharmaceuticals as well. Despite successes like AlphaFold’s predictions of protein structures, it has yet to make any equal breakthroughs in chemistry and medicine, Pei said.  

A pharmaceutical expert told NewsChina that biological systems are too complicated to be simulated by AI modeling. For instance, when involving AI in evaluating drug effects, precision is sacrificed.  

That is why AI models have yet to create any meaningful results for drug development through big data. Furthermore, pharmaceutical manufacturers cannot provide enough data to meet AI’s voracious demand for deep learning.  

On April 18, the stock price of MIT-based AI pharmaceutical firm Relay Therapeutics tumbled by 36 percent when its anti-tumor suppressant RLY-2608 fell short of expectations compared to initial clinical trials.  

Unlike ChatGPT, which is dependent on reinforcement learning, a process for machines to learn from user feedback, tagging protein sequences requires huge amounts of experiments, Guo Chunlong, CEO of Beijing-based Shuimu BioSciences, told NewsChina.  

Even though scientists are split on whether AI can trigger another scientific revolution, many believe their jobs are safe for now.  

Despite AI’s outstanding abilities, Wen believes that humans are still unique in their insightful and critical thinking. “Machines will never ask why one plus one equals two, but humans will,” he said.

Pictured is the rendered structure of a human protein by AlphaFold, an AI system developed by DeepMind, displayed at EMBL’s European Bioinformatics Institute (EMBL-EBI) on July 22, 2021. In July 2022, AlphaFold announced its predictions for 214 million protein structures, including 98.5 percent of human proteins (Photo by VCG)