As Madonna famously remarked, we live in a material world. Our species has called historical periods for bronze, stone, and iron. Modern society relies on breakthrough materials like lithium-ion batteries and solar cells.
 |
| Can AI really change our material world? |
So, when Google DeepMind researchers announced in November that their artificial intelligence technology had identified more than 2 million new crystalline minerals, the announcement made global headlines. The business described the findings as "an order-of-magnitude expansion in stable materials known to humanity".
Separately, other Berkeley researchers revealed that their automated laboratory had developed 41 unique compounds in less than three weeks, which were cross-referenced with Google DeepMind's database. Only the dullest intellect could fail to envision a gleeful future: lines of robotic arms manufacturing sparkling new AI-designed materials to address great challenges such as clean energy.
However, since the two papers were published in Nature, the glitter has faded slightly. Earlier this month, materials scientists claimed Google DeepMind had oversold its accomplishment. In March, researchers questioned whether the 41 chemicals described by the Berkeley team were truly unique. Both Google DeepMind and the Berkeley team told the Financial Times that they stood by their separate papers.
The uproar comes as the company's co-founder, Sir Demis Hassabis, has cautioned against artificial intelligence hype and "grifting". The clash demonstrates that, while AI has the potential to alter, businesses and institutions are trying to balance optimism with overselling.
Coming up with new materials is typically a costly and time-consuming process of trial and error. DeepMind's AI technique, known as Graph Networks for Materials Exploration, or Gnome, provides a computational shortcut for one type of compound, inorganic crystals. It generates new candidate crystals from existing libraries of known structures and employs artificial intelligence to iterate toward structurally stable molecules.
Of the 2.2 million unique elements discovered, the business deemed 380,000 stable enough to include in a database. However, Anthony Cheetham and Ram Seshadri of the University of California, Santa Barbara, stated this month in the journal Chemistry of Materials that the analysis revealed "scarce evidence for compounds that fulfill the trifecta of novelty, credibility, and utility." Cheetham recently told me that "AI has a great future in materials science, but . . . [that] some of the accomplishments to date have been overhyped".
A Google DeepMind spokeswoman stated that the study simply claimed innovation and stability, not any specific qualities and that the critique appeared to be based on different terminologies. He noted that further studies would shed light on the chemicals' characteristics.
The other AI-related publication being criticized in materials science centers on work done by researchers at the University of California, Berkeley, and the Lawrence Berkeley National Lab. They've created an automated lab that uses AI-guided robots to mix and characterize novel substances. The Berkeley team announced that its "A-lab" had created 41 unique molecules, relying in part on the Google DeepMind database.
However, Robert Palgrave, a chemist at University College London, and other Princeton University sceptics stated last month that the Berkeley claim was unfounded on two grounds: the AI offered compounds that were already known and were later unable to detect that they lacked originality. Gerbrand Ceder, the head of the Berkeley team, says: "We stand by the results in our paper that the A-lab succeeded in autonomously developing and demonstrating synthesis recipes for compounds for which it had no prior information, which is a remarkable achievement."
Palgrave admits that novelty is a subjective term: chips with extra salt are still chips, while caramel with added salt is seen differently by many people. He goes on to say that scientists often agree on material novelty.
In other words, this is a large-scale cultural confrontation. There is also an unavoidable mismatch in quality: early AI forays into materials science may be spectacular by AI standards, but they are still far from the degree of human competence in the topic.
We don't know how long that mismatch will last: researchers hope to conjure up a high-temperature superconductor with a well-defined user prompt. If and when that day arrives, we will still live in a tangible world, although one that we did not wholly create.