A team of researchers based in the Lawrence Berkeley National Laboratory is working to train a computer to potentially discover new weather patterns.

The Deep Learning for Climate project, led by campus earth and planetary science doctoral student Prabhat, has been using artificial intelligence to find patterns of extreme weather such as hurricanes and cyclones for the last three years. By looking at known weather patterns, the group will train a neural network — a computer program inspired by the human brain and designed to continuously improve itself — to be able to quantify how climate might change in the future.

“(The networks are) pretty computer intensive,” said Michael DeWeese, a campus associate professor of physics and neuroscience. “There are neural networks now that no other kind of algorithmic approach can match, so it’s worth the effort and it’s worth the computation power.”

According to Prabhat, simulations have already been written using decades of scientific work describing how the atmosphere, land and ocean would change over time. Prabhat’s team’s goal is to look through these simulations to find patterns of extreme weather that climate scientists have criteria for but also discover climate patterns that scientists do not yet have criteria for.

“We’re going to find those patterns and quantify them, but also there might be new types of weather patterns that might only happen in the future and have not yet been observed in the past,” Prabhat said.

According to Daniel Kammen, a campus professor of energy, climate is a highly complex interaction between global trends and local conditions. As such, greater data monitoring and other analysis will make models more and more relevant to decision making, such as in determining whether the recent Hurricane Harvey was caused by global warming.

The group is most interested in learning how the climate will change in the future, according to Prabhat. They are looking to answer questions such as whether hurricanes will hit land more often in the future, and whether they will be more intense when they do.

To do this, Prabhat’s team will focus on analyzing simulations that scientists have already made, which, he said, are commonplace and have simulated up to the end of the 21st century. The group has no current plans to find a way to actually predict climate, however, and Prabhat added that creating artificial intelligence to predict the weather is much more difficult, even suggesting such a project makes him uncomfortable.

“We have to start asking really hard questions,” said Prabhat. “Are the solutions or the projections the AI system makes, are those physically consistent? … If I’ve never encountered a certain situation before, I do not have the training data for that. How do the AI system possibly make predictions?”