Megan
Stanley

I've spent much of my recent career building models that predict complex scientific systems. But prediction and explanation aren't the same objective. I'm interested in how we build AI systems that can move beyond prediction to discover parsimonious explanations: forming hypotheses, designing discriminating experiments, revising models from evidence, and communicating the resulting understanding to humans. I'm interested in the model architectures and training methods that make those capabilities possible.

I came to machine learning from physics and have spent close to a decade doing AI research in industry. At Microsoft Research I was a first author on Aurora, a foundation model for the Earth system published in Nature (2025), contributed to Skala, a deep learning model for quantum chemistry, and worked on few-shot learning and molecular representation. I'm now at the Ellison Institute of Technology, building foundation models for biological sequences: proteins, genetics, and whole-cell behaviour.

I enjoy formulating ideas, and I do the technical work to test them myself: designing models, writing code, and running experiments.

The systems I want to build maintain an internal model of their domain and improve it through interaction with data, tools, simulation, and experiment. Biology is a case in point: to be modellable at all it needs both language and verification in the lab. The representations these systems learn about the world fascinate me in their own right. And whatever modelling paradigm we ultimately choose, we should build systems that are transparent, with true explanation accessible to humans.