Deep learning can almost perfectly predict how ice forms

Researchers have used deep learning to more precisely than ever model how ice crystals form in the atmosphere. Their paper, published this week in PNAS, points to the potential to significantly increase the accuracy of weather and climate forecasts.

The researchers used deep learning to predict how atoms and molecules behave. First, models were trained on small-scale simulations of 64 water molecules to help them predict how electrons in atoms interact. The models then replicated those interactions on a larger scale, using more atoms and molecules. It is this ability to accurately simulate electron interactions that allowed the team to accurately predict physical and chemical behavior.

“The properties of matter arise from how electrons behave,” said Pablo Piaggi, a researcher at Princeton University and the study’s lead author. “Explicitly simulating what happens at that level is one way to capture much richer physical phenomena.”

It is the first time this method has been used to model something as complex as ice crystal formation, also known as ice nucleation. This is one of the first steps in cloud formation, where all the precipitation comes from.

Xiaohong Liu, a professor of atmospheric science at Texas A&M University who was not involved in the study, says that half of all precipitation events — be it snow, rain or sleet — start out as ice crystals, which then enlarge and result in precipitation. If researchers could more accurately model ice nucleation, it could give the weather forecast a big boost overall.

Nucleation of ice is currently predicted based on laboratory experiments. Researchers collect data on ice formation under various laboratory conditions, and that data is fed into weather forecasting models under comparable real-life conditions. This method works well enough sometimes, but is often inaccurate due to the large number of variables involved in the actual weather conditions. If even a few factors differ between the lab and the real world, the results can be quite different.

“Your data is only valid for a certain region, temperature, or kind of lab environment,” Liu says.

Predicting ice nucleation based on how electrons interact is much more accurate, but it is also very computationally expensive. It requires researchers to model at least 4,000 to 100,000 water molecules, and even on supercomputers, such a simulation can take years. Even that could only model the interactions for 100 picoseconds or 10-10 seconds — not long enough to observe the ice nucleation process.

However, using deep learning, researchers were able to perform the calculations in just 10 days. The duration was also 1000 times longer – still a fraction of a second, but just enough to see nucleation.

Of course, more accurate models of ice nucleation alone won’t make the forecast perfect, Liu says, because it’s only a small but crucial part of weather modeling. Other aspects are also important: understanding how water droplets and ice crystals grow, for example, and how they move and interact under different conditions.

Still, the ability to more accurately model how ice crystals form in the atmosphere would greatly improve weather forecasts, especially those regarding whether and how much it is likely to rain or snow. It can also help predict climate by improving the ability to model clouds, which affect the planet’s temperature in complex ways.

Piaggi says future research could model ice nucleation when there are substances like smoke in the air, potentially improving the accuracy of models even more. Thanks to deep-learning techniques, it is now possible to use electron interactions to model larger systems for longer periods of time.

“That essentially opened up a new field,” Piaggi says. “It is already playing and will play an even greater role in simulations in chemistry and in our simulations of materials.”