Water, machine learning to understand it better.
The density of water and the liquid-liquid phase transition: new insights from quantum chemistry and machine learning.
Water has puzzled scientists for decades. Over the past 30 years or so, they have theorized that, when cooled to a very low temperature like -100°C, water might be able to separate into two liquid phases of different densities. Like oil and water, these phases don’t mix and may help explain other strange behaviors of water, such as the fact that it becomes less dense as it cools. However, it is nearly impossible to study this phenomenon in the laboratory because water crystallizes into ice quickly at such low temperatures. Now, new research from the Georgia Institute of Technology uses machine learning models to better understand the phase changes of water, opening new avenues for a better theoretical understanding of various substances. With this technique, the researchers found strong computational evidence to support the liquid-liquid transition of water (some types of phase transitions are solid-liquid, solid-gas, liquid-gas transitions) that can be applied to real systems that use water to function.
“We’re doing this with very detailed quantum chemistry calculations that try to get as close to the physics and real physical chemistry of water as possible,” said Thomas Gartner, an assistant professor in Georgia Tech’s School of Chemical and Biomolecular Engineering. “It’s the first time we’ve been able to study this transition with this level of precision .”
The research is presented in the article “Liquid-Liquid Transition in Water From First Principles,” published in the journal Physical Review Letters, with co-authors from Princeton University.
To better understand how water interacts, the researchers ran molecular simulations on supercomputers , which Gartner likened to a virtual microscope . “If you had an infinitely powerful microscope, you could zoom down to the level of individual molecules and watch them move and interact in real time,” she said. “That’s what we’re doing, creating almost a computational film.”
Researchers analyzed the motion of molecules and characterized the structure of the liquid at different water temperatures and pressures , simulating the phase separation between high- and low-density liquids. They collected a lot of data, ran some simulations for a year, and kept refining their algorithms to get more accurate results. Even a decade ago it wouldn’t have been possible to run simulations this long and detailed, but today machine learning offers a shortcut. The researchers used a machine learning algorithm that calculated the energy of the interactions between the molecules of water. This model computed much faster than traditional techniques, allowing simulations to proceed much more efficiently.
Machine learning isn’t perfect, so these lengthy simulations also improved the accuracy of predictions. The researchers were careful to test their predictions with different kinds of simulation algorithms. When multiple simulations gave similar results, the accuracy of the predictions was validated. “One of the challenges of this work is that there aren’t a lot of data to deal with, because it’s a nearly impossible problem to study experimentally,” Gartner said. “We’re really pushing the envelope, and that’s another reason why it’s so important that we try to do it using different computational techniques.”
Beyond the water
Some of the conditions tested by the researchers were extremes that probably don’t exist directly on Earth, but could potentially be present in various aquatic environments in the solar system, from the oceans of Europa to the water in the center of comets. But these findings could also help researchers better explain and predict water’s strange and complex physical chemistry, informing water use in industrial processes , developing climate models , and more.
According to Gartner, the work is even more generalizable. Water is a well-studied research area, but this methodology could be extended to other hard-to-simulate materials, such as polymers, or complex phenomena such as chemical reactions, Gartner said. “We now have this very powerful new computational technique, but we don’t yet know what the limitations are, and there is a lot of room to advance the field.”