Abstract:
AI for weather and climate has arrived, but new scientific insight has lagged behind. To get the most out of these tools, we need to stop treating them solely as faster forward simulators and start taking advantage of what AI architectures can uniquely do, such as running extreme events "on rewind" to directly sample antecedent conditions. Furthermore, the era of "bigger is better" is ending, and the field is starting to reward smart over big. This makes explaining model decision-making even more critical. While forecasts tell us what a model can do, and post-hoc explainable AI tells us where it looks, training tells us how it thinks. We don't have to open the black box after the fact; we can watch it being built. In this talk, I will show how creative task design and investigations into training dynamics allow us to both uncover new climate science and improve current state-of-the-art approaches.
Bio: Dr. Elizabeth Barnes is the Dalton Family Chair in Environmental Data Science & Sustainability, Professor of Computing & Data Sciences, and Professor of Earth & Environment at Boston University, which she joined in 2025.
Her group's research focuses on understanding Earth system variability, predictability, and change across time and space, with an emphasis on developing and implementing artificial intelligence tools in a way that reflects scientific reasoning to improve intrinsic interpretability. An overarching research goal of her group is to responsibly harness AI to anticipate human-Earth system futures in support of a thriving society in the decades ahead.
Bio:
Dr. Elizabeth Barnes is the Dalton Family Chair in Environmental Data Science & Sustainability, Professor of Computing & Data Sciences, and Professor of Earth & Environment at Boston University, which she joined in 2025.
Her group's research focuses on understanding Earth system variability, predictability, and change across time and space, with an emphasis on developing and implementing artificial intelligence tools in a way that reflects scientific reasoning to improve intrinsic interpretability. An overarching research goal of her group is to responsibly harness AI to anticipate human-Earth system futures in support of a thriving society in the decades ahead.
Summary
AI for
Tailored prediction (liaise between human-earth system components)
Post–processing of climate/weather data
Improve physics-based models
Merging observations and model data
Climate change communication
Deep-learning weather/climate emulators (Focus of talk)
Emulators allow new experiments, not just faster older experiments
Simulate thousands of possible futures
cBottle generates samples of weather states, not time series forecasts
Generates entire time series as a single sample
Can condition samples on a starting and ending condition
Starting: roll out forward in time
Ending: roll out backwards to see which scenarios can lead to the target state
Example: look at Superstorm Sandy and which storm trajectories may have led to landfall at that location
Example: possible trajectories of a given heat wave
Weather to climate:
Disaster preparedness
Property insurance
Infrastructure Planning
Are the joint statistics right?
E.g. Hot-Dry compound event frequency
Comparing numeric models to diverse AI models shows that these joint statistics are still not being accurately reproduced
Currently, not even as accurate as using the climatological averages for these statistics
Observation: the distribution of different forecasts from AI models looks a lot more like year 2000 than future climates (since that’s where the training data comes from)
Challenge: AI models can easily memorize the short data records we have in the recorded climate era, which creates challenges for evaluating them and generalizing outside of the training set
Explainable and interpretable AI: XAI
Why?
Estimate confidence in a model’s prediction
Helps to fine-tune and optimize models
Learn new science
Analysis
Insert lagged ocean boundary conditions to make climate predictions for subsequent time
Then look back into the network to understand which regions of space/time to make good/bad predictions
Very interpretable with simple feed forward networks
But with complex autoregressive networks (e.g. Graphcast) it's hard to interpret
Looking at gradients across rollouts of autoregressive models
Look at the temperatures at a given location and then surf the gradients backwards to identify what the model is most sensitive to
E.g. temperatures in Massachusetts depend on weather systems to the west (directions of the dominant winds)
Understanding what happens during model training
Approach:
Train the model on the full dataset
Make a forecast and watch it
Repeat for each epoch to see how to forecast evolves
Example: track AI model to forecast atmospheric river 7 days ahead
Between checkpoint 0 and 25 the model gets increasingly accurate
At checkpoint 27 the model seems to forget
Then it relearns the dynamics by checkpoint 87
Forgetting phenomenon happens often across different target metrics
Analysis: look whether perturbing an input impacts multiple metrics
Indicates that the model learns cross-metric correlations
Sometimes models forget and never recover
Detection of unlearning makes it possible to intervene on the model and avoid it
Where the field is stuck
Smarter, not bigger: dataset is constrained by limited availability of historical data
AI slop is an immediate threat: trust is huge
Interpretability is very important
Calibration of extreme event frequencies is difficult to assess