AI for Climate Science Seminar Series: Beyond Prediction - Answering Climate What-If Questions from Observations

Join us for the next AI for Climate Science Seminar!

After a short summer break, we are back in September with a guest talk by Dingling Yao from the Institute of Science and Technology Austria.

Dingling Yao is an ELLIS PhD candidate at the Institute of Science and Technology Austria, advised by Francesco Locatello, and currently a visiting researcher at Caltech with Anima Anandkumar. Her research lies at the intersection of causal representation learning, scientific machine learning, and climate science. She studies how AI systems can move beyond predictive shortcuts to recover meaningful physical structure and answer reliable what-if questions from observational data. She is a 2025 Google PhD Fellow and a 2022 Amazon–MPI Science Hub Fellow.

For online participation, a registration is necessary.

Please register here.

Title:
Beyond Prediction: Answering Climate What-If Questions from Observations

Abstract:
Machine learning models can now forecast complex scientific systems with remarkable accuracy. But science often asks a harder question than what happens next?: What would happen if the system were different? Answering such what-if questions from observational data is fundamentally challenging because machine learning models are designed to capture patterns in their training data and, by default, need not generalize to unseen scenarios.
I will first illustrate this challenge through the Perception–Physics Paradox. Using tropical cyclones as a case study, I will show that vision foundation models can predict physical quantities from satellite imagery while failing to represent the underlying physical state—especially for intense storms, precisely where scientific reliability matters most. I will then discuss how combining causal representation learning with dynamical-system identification can help recover scientifically meaningful latent parameters from observed trajectories, enabling downstream causal analyses of real-world sea-surface temperatures. Finally, I will discuss how these ideas extend to large-scale scientific emulators, and how incorporating scientific structure may help such models reason more reliably about conditions that are not represented in their training data.
Together, these works point toward a path beyond pattern matching: building scientifically grounded AI systems that can use observational data not only to predict the world as it is, but also to reason reliably about how it could be.