Unpacking snow

Environmental scientist Marianne Cowherd grew up in Michigan and loved snow. “My favorite thing was having school cancelled and going sledding,” she says. “But I never thought of snow as a water source until moving to California.”

The Department of Energy Computational Science Graduate Fellowship (DOE CSGF) recipient has turned her passion for snow to a pressing environmental and computational challenge. At the University of California, Berkeley, she’s documenting how climate change and fires disrupt existing mountain snowpack prediction systems and how computational tools can supply creative solutions.

Measuring and predicting the snowpack in California and other Western states is critical to water resource planning. The snowpack is what researchers call a “mountain water tower,” a reservoir for millions of people that rely on the melt water.

“In California, more than half of our water comes from snow originally,” says Cowherd, who works with Scott Stephens and Manuela Girotto in UC Berkeley’s environmental science, policy and management department. “So it’s the phase of the water cycle where fire and climate change have the biggest effect.”

The challenge is that the traditional ways of estimating snowpack are breaking down. For example, a key snowpack monitoring tool in the Western United States is a series of a thousand snow pillows. These are sensitive scales placed at ground level that weigh accumulated snow.

Most of the snow pillows reside at medium elevations. But in near-future winters, only rain might fall at these levels because of climate change, with snow at higher elevations. This shift will throw off measurements and predictions.

“The snow monitoring stations were installed in the 1980s with the assumption that the climate would be relatively stationary,” Cowherd explains. “The traditional predictive linear models just can’t keep up with the complexity of how the snowpack is responding to climate change and fires.”

In her fellowship research, Cowherd has harnessed machine learning, a form of artificial intelligence, to capture these nonstationary changes and maintain or even improve predictive accuracy. “Machine learning gives us the flexibility to model those nonlinear relationships and spatial patterns,” she says.

Cowherd used U-Nets, a machine learning program that she trained on 20 years of snow pillow data combined with terrain images and a temperature map. Using the Perlmutter supercomputer at DOE’s National Energy Research Scientific Computing Center, she showed that applying U-Nets improved predictions of the distributed snowpack by 184%. The results were published in Communications Earth and Environment in 2024.

Cowherd has demonstrated that the link between snow pillow monitoring sites and overall basin snowpack estimates is already being skewed by major fires. Studying the area of the 2021 Caldor fire in northern California, she found that forest-cover loss in burned areas resulted in greater early-season snow accumulation but also faster spring melt.

“We quantified the mismatch” between the traditional snow pillow observations before and after the fire, says Cowherd, who presented her team’s results at the Western Snow Conference in April 2024.

Cowherd’s Ph.D. studies have also revealed that water resource planning needs to account for the changing nature of snow droughts, periods of low or no snowpack. In research at Pacific Northwest National Laboratory during her first CSGF practicum, she used an ensemble of climate model simulations, including the DOE’s flagship climate model E3SM, to analyze changes in the frequency and severity of snow droughts from 1850 to 2100. Her research found that compared to the historical benchmark from 1850 to 1900, snow drought frequency more than doubles later this century.

“What we’re seeing is that the drivers of snow droughts are shifting,” Cowherd says of the research published in the journal Environmental Research Letters in 2023. “Historically, we’ve had a balance of dry years and warm years, but the future is dominated by warm droughts, where it’s just too warm for significant snowpack to accumulate, even if precipitation is normal.”

Cowherd has shared her research with the California Cooperative Snow Surveys program which runs the snow pillow network and hopes it will ultimately provide water managers with improved tools for water resource decision making. After completing her Ph.D. in 2025, she hopes to continue playing in the snow in government or academia.

“I’m really passionate about staying in this field of adapting water management systems to climate change,” she says. “There’s so much important work to be done.”