Maike Sonnewald

Where, and why, is the ocean's circulation governed by certain physical balances? Where do distinct marine ecological communities live? How do we find structure in noisy, high-dimensional data?

The ocean's circulation regulates Earth's heat, carbon, and sea level, and its ecosystems sustain over three billion livelihoods. However, determining boundaries to build understanding and monitoring tools is often mapped by hand and expert judgement. I introduced unsupervised ML to identify ocean dynamical regimes objectively, which revealed a Southern Ocean “supergyre,” and built SAGE and NEMI to let statistically significant structure emerge from convoluted data with a principled validation pipeline that works when there is no ground truth.

• Supergyre, Nature Communications Earth & Environment (2023)
• Marine eco-provinces, Science Advances (2020)
• Ocean dynamical regions, Earth & Space Science (2019)
• NEMIGithub

Can we infer the unobserved ocean interior from the surface we see everywhere? Has a network or numerical model genuinely gained dynamical understanding, or does it only look right in its output fields?

Neural networks are suspiciously good predictors on hard, sparse-data problems, but by default do not divulge this insight, and cannot show they are trustworthy. This is deeply problematic for deployment in support of high-stakes decisions. I treat networks not as an oracle but as a guide to discovery, by building methods to check, or help it learn, domain-consistent patterns, then use this to uncover new science. THOR infers interior ocean dynamics from the surface; OceanCBM builds physics-derived concepts into the architecture; and my fidelity-verification framework can catch models being confidently wrong.

• THOR, Journal of Advances in Modeling Earth Systems (2021)
• Bayesian neural networks for ocean dynamics, Journal of Advances in Modeling Earth Systems (2022)
• OceanCBM (preprint, in review at NeurIPS 2026)
• THORGithub

The adoption of foundation models is unfolding at breathtaking pace. Where does a massive AI model's prediction hold? Can we verify foundation models?

I am carrying fidelity verification into the foundation-model era, probing where a large model's physics is reliable so the next generation of scientific models can either be trustworthy by construction, or have fidelity verified after training. My work deliberately extends from ocean systems to coupled ocean-atmosphere systems used for weather prediction and more. I see incredible potential, but also high stakes, as these tools improve much faster than our scientific understanding.

• Interpretability challenges in a foundation model for continuum dynamics, ICLR 2026 workshop
• Machine learning for climate physics, Annual Review of Condensed Matter Physics (2025)
• Neural Earth system modelling, Nature Machine Intelligence (2021)