Storms, Waves, and a Century of Simulated Weather
The Faculty of Computing & Data Sciences offers a place where students and researchers can explore questions that would be impossible to answer with observations in a single discipline. The team working with Dr. Elizabeth (Libby) Barnes, Dalton Family Chair in Environmental Data Science & Sustainability and Professor of Computing & Data Sciences and of Earth & Environment at Boston University, is combining cutting-edge AI with traditional climate science, bringing together advanced modeling, machine learning, and real-world forecasting. Whether studying tropical cyclones, extreme rainfall, or past climates stretching back hundreds of years, researchers here have the opportunity to test bold ideas, work with enormous datasets, and help shape the next generation of climate science. The Barnes Group is tackling some of the hardest questions in weather and climate: why storms form, how atmospheric waves interact, and what hidden patterns emerge over decades or even centuries.
The age‑old question of how convectively coupled Kelvin waves contribute to tropical cyclones has long been set back by a simple problem: we just don’t have a long enough modern observational record to see the full story. In Mu‑Ting Chien’s work with the deep learning climate emulator ACE2, that limitation eases, turning a few decades of measurements into a century‑scale laboratory where the relationship between waves, atmospheric moisture, and low level rotation can be explored in far greater detail.
A convectively coupled Kelvin wave is an eastward moving equatorial atmospheric wave that travels together with organized tropical convection and rainfall. The coexistence of both convectively coupled Kelvin waves and tropical cyclones is rare in a short observational record. A few decades of satellite data and reanalyses are not enough to map out robust statistical links between them. A climate emulator like ACE2 provides new insights by generating long and consistent climate simulations (for ~100 years) so researchers can accumulate enough “events” to see patterns that would be impossible to see in observations alone. Traditional climate models can currently do this, but they are expensive to run and often struggle to represent tropical cyclones and equatorial waves at coarse resolution.
ACE2 is trained on high quality reanalyses data (i.e., observations assimilated in numerical models) so that it can act as a quick substitute: it learns how key climate quantities typically respond to each other, then reproduces that behavior much faster than a full physics-based climate model. The representation of the mean climate in ACE2 stays within the range of behavior we already see in existing models and observations, and checks against familiar patterns like how often tropical cyclones form and how tropical waves behave.
For someone unfamiliar with the concept, a climate emulator operates similarly to a traditional climate model in the sense that it produces the temporal evolution of atmospheric conditions for decades. However, instead of solving the full fluid dynamics equations from scratch every time, an emulator learns the input/output relationship, showing patterns of sea‑surface temperature and atmospheric conditions evolve in time, which come directly from a large archive of simulations and observations. Once trained, it can be run cheaply and repeatedly.
That makes emulators powerful for generating large ensembles and testing hypotheses. Researchers can instantly run many “what if” scenarios where Kelvin waves, background moisture, or other components are dialed up or down, and see how often tropical cyclones form under each combination. However, the limitation is that emulators are tuned to represent the mean climate states; in order to answer specific questions and examine specific phenomena, we need to closely verify that the emulator is suitable for the purpose. If verified, the ACE2 model can be very powerful in testing hypotheses and enhance our physical understanding of weather and climate variability.
The flip side? “The vertical resolution is relatively coarse, so it's only 8 vertical levels. So if you want to look at the detailed vertical structure, it could be a bit limited,” Chien says. It’s great for the big-picture ideas like overall storm counts and wave patterns, but might miss the details inside a storm, like how moisture evolves at different heights.
She puts it simply: “I don’t think this is a replacement of physics-based climate models because the AI emulators are still very novel and we still need a lot of tests. So, I think in the future, we have to go hand in hand, both the physics simulation and the AI emulator.” Physics models keep the guardrails with their hardwired laws of nature, and AI speeds things up for massive “what if” runs that current, real data can’t touch.
This approach excels at unraveling other elusive climate connections, such as interactions among equatorial waves or the drivers behind extreme rainfall events. As Chien observes, “Extreme precipitation is usually a compound effect of different conditions, but it's very difficult to separate all these effects from observation.” By generating vast ensembles of simulated years, emulators provide the statistical power needed to isolate these effects.
She cautions against blanket application, however: “My study is just on Kelvin waves and TC. I'm definitely not saying that we can use this emulator for everything. Before we really use the emulator to represent each of the phenomena, there has to be some validation.” Certain processes will always require direct observational or physics based checks.
In practice, the method holds promise for forecasting: “This suggests that this model can be used for subseasonal prediction. We can also run it just for, like, 2 or 3 weeks and predict how likely the tropical cyclone is going to form.” Starting from current conditions, it offers a 2 to 3 week outlook on cyclone risks or heavy rain, aiding preparation amid uncertainty.
For weather forecasters or policymakers interested in adopting these AI emulators, Chien is clear: don’t blindly trust AI; verify that ACE2 matches real observations and traditional models on basics like storm seasons, tropical wave patterns, and known physical relationships. The AI emulators can improve subseasonal forecasts with careful examination.
Chien is now turning to paleoclimate, using emulators to reconstruct periods like the Little Ice Age (1350-1550), where El Nino-like sea temperatures correlated with reduced global cyclone activity. “It's very hard to extract the relationship just through a few decades of modern observational data. So that's why I go back to the past for like a few 100s of years,” she explains. By validating against proxy data such as sediment records, she builds confidence for extrapolating to future scenarios without direct observations. “I’m very excited to expand my research to the past millennium.”
-- Shriya Jonnalagadda (CDS'28), Data Science Research Communications Intern