Inside a lab at Carnegie Mellon University in Pittsburgh, a robot arm lifts a bottle filled with chemical reagents and carries it over a bank of test tubes, where it dispenses a precise number of drops into each one. The arm swivels, replaces the bottle, swivels again, and picks up another container. Gracelessly, tirelessly, the machine thrums on, carrying out test after test. The experiments are part of an ongoing project to determine the ideal chemical makeup for high-capacity electric car batteries. Soon, machines won’t just run the experiments—they’ll devise them, too.
Over the next few months, an artificial intelligence algorithm will gradually take over the planning of experiments based on the battery test runs. Once fully functioning, this robot graduate student will decide how to modify the concentrations of the ingredients it’s testing. “It’s automating not only the manual part of doing the experiment but also the planning part,” says Brian Storey, the Toyota Research Institute scientist leading the project.
Science has long been considered one of the human activities least likely to be farmed out to robots. That’s changing as sensors, sequencers, and satellites churn out digital information by the terabyte. “We just cannot handle the amount of data anymore,” says Manuela Veloso, who heads Carnegie Mellon’s machine learning department. It’s a daily concern for biotech companies and a wide range of other businesses struggling to make sense of the unprecedented swell of raw information.
AI software designed to identify and sort patterns has been deployed across a wide swath of science, from marine biology (identifying wild dolphin vocalizations from hydrophone recordings) to astronomy (detecting the presence of planets from subtle fluctuations in the brightness of thousands of stars). To discover the Higgs boson, the so-called God particle, an algorithm sifted billions of particle tracks generated within the Large Hadron Collider in Switzerland. AI is fast becoming an essential part of university science curricula.
Automating the process of discovery doesn’t just free up researchers’ time. It could potentially change what sorts of discoveries are made. “I can easily imagine cases in which AI would recommend experiments to try to synthesize a chemical molecule that you wouldn’t think possible, but the AI will be able to do it,” says Barnabás Póczos, a Carnegie Mellon machine learning professor collaborating on the Toyota project.
Unfortunately, generating novel predictions isn’t all that useful by itself. What scientists are after is less what than why—the elegant theoretical formulations that let them understand how the universe works, such as Newton’s first law or E=mc². So far, the neural networks underlying AI software can’t really explain how they arrive at their answers.
Humans, in contrast, are pretty good at that. So in the near term, the most promising approach will be for humans and AI to work together. In February, Dutch publisher Elsevier announced a trial collaboration with software maker Euretos, using AI to assess millions of peer-reviewed scientific articles to suggest hypotheses in the field of biochemistry. Academics will cull these hypotheses online, basing experiments on the most encouraging ones. “The vision is that the discussion becomes a much more automated process,” says Euretos co-founder Arie Baak.
And after that? “People have wondered if you could have the computer automatically figure out the principles underlying physics,” says Toyota’s Storey. “I don’t think we’re going that far out now.”