For years, scientists have suspected that, at the molecular level, water is two different liquids ‪— a denser one and a less-dense one ‪— that are constantly switching places. Catching real molecular evidence of this microscopic transformation has been hard. But now, with help from artificial intelligence, researchers say they’ve finally found it.

“It’s hard to imagine — here is just one water, right?” said Xiao Cheng Zeng, a physical chemist at the City University of Hong Kong and co-author of the new study, told Live Science while holding a water bottle in the air. That puzzle sent him digging through scientific literature, where he found the possible explanation: the two-state hypothesis. “That got my attention. We have literature to talk about it but no evidence.”

The findings, published in the journal Nature Physics, could not only prove this long-sought molecular change is real, but also help to explain dozens of water’s weird behaviors. Most liquids become denser as they cool, but water behaves differently; it becomes denser until about 4 degrees Celsius, then starts to expand, which is why ice floats. Water also resists temperature changes better than similar liquids and has a viscosity that decreases under certain pressures. Scientists have documented various anomalies related to water and suspect they may be interconnected.

The two-state model is an attempt to be that unifying explanation. Zeng has been studying water since his postdoc days in the late 1990s, when he worked on liquid freezing. The two-state hypothesis itself came onto his radar later — around 2006, when he first encountered it at scientific conferences. But for years, he set it aside as too difficult to tackle directly. That changed roughly around 2016, as researchers began reporting experimental evidence that supercooled water could split into distinct high-density and low-density forms.

Around two and a half years ago, Zeng handed the problem to Liwen Li, a postdoctoral researcher in his lab. Rather than repeating the conventional approaches other groups had already struggled with, Li suggested the use of “unsupervised deep learning” — AI trained to spot patterns in data without being told what to look for. “So AI [is] forced to learn — to use [its] knowledge to create, to explore,” Zeng told Live Science.

The team ran massive molecular dynamics simulations, using the GROMACS simulation package. They tracked how hundreds of thousands of water molecules moved and interacted and generated tens of millions of data points. “Traditionally, you may need a lot of students to figure that out. … With computers and AI, it took [Li] maybe a year and a half,” Zeng said. Without AI, he estimated, the same analysis might have taken closer to a decade.