Research

I work on deep learning for the physical sciences, with a particular focus on weather and climate modeling and on accelerating scientific simulations. My collaborators and I have published in Nature, Science, PNAS, Physical Review Letters and NeurIPS.

For a complete list, see my Google Scholar profile (40,000+ citations, h-index 29).

AI for weather and climate

I conceived and led the Neural General Circulation Models project at Google Research. NeuralGCM is the first AI-based model to improve on traditional physics-based 15-day weather forecasts and atmosphere-only climate simulations. It combines a differentiable atmospheric dynamical core written in JAX with learned parameterizations for sub-grid physics, and is trained end-to-end on reanalysis data.

The work was published in Nature (2024) with 16 co-authors, and received press coverage in Bloomberg and MIT Technology Review, among others. The open-source model has since been used to produce state-of-the-art monsoon-onset forecasts sent to 38 million farmers in India, in collaboration with the University of Chicago and the Indian Ministry of Agriculture.

On the technical side, I personally implemented many of the key components, including model and data parallelism scaling to 256 TPUs.

• Neural general circulation models for weather and climate. D Kochkov, J Yuval, …, S Hoyer. Nature (2024).
• WeatherBench 2: A benchmark for the next generation of data-driven global weather models. S Rasp, S Hoyer, et al. Journal of Advances in Modeling Earth Systems (2024).
• Learning skillful medium-range global weather forecasting. R Lam et al. Science (2023).

AI for fluids

For several years I led a research program on accelerating fluid simulations using deep learning and Google TPUs. The program produced six peer-reviewed publications (two in PNAS) and over 2,000 citations.

• Machine learning–accelerated computational fluid dynamics. D Kochkov, …, S Hoyer. PNAS (2021).
• Learning data-driven discretizations for partial differential equations. Y Bar-Sinai*, S Hoyer*, J Hickey, MP Brenner. PNAS (2019).

AI for physics

I have published research applying machine learning to a range of problems across the physical sciences, including drug discovery, microscopy, nanophotonics, quantum chemistry, structural engineering, flood modeling and fundamental physics.

• Lagrangian neural networks. M Cranmer, S Greydanus, S Hoyer, et al. arXiv (2020).
• Kohn–Sham equations as regularizer: building prior knowledge into machine-learned physics. L Li, S Hoyer, et al. Physical Review Letters (2021).
• Free-form diffractive metagrating design based on generative adversarial networks. J Jiang, D Sell, S Hoyer, J Hickey, J Yang, JA Fan. ACS Nano (2019).
• Neural reparameterization improves structural optimization. S Hoyer, et al. NeurIPS workshop (2019).

Fundamental AI research

At Google, I collaborated with researchers at DeepMind and Google Brain and published in top machine learning conferences.

• Efficient and modular implicit differentiation. M Blondel et al. NeurIPS (2022).
• The Cramér distance as a solution to biased Wasserstein gradients. M Bellemare et al. arXiv (2017).

Tools for scientific computing at scale

Alongside specific research projects, I have contributed to general-purpose tools for working with large scientific datasets, including Xarray, NumPy and JAX. See Software for the open source projects.

• Array programming with NumPy. CR Harris et al. Nature (2020).
• Xarray: N-D labeled arrays and datasets in Python. S Hoyer, J Hamman. Journal of Open Research Software (2017).

Earlier: quantum dynamics

Before joining industry I completed a Ph.D. in theoretical physics at UC Berkeley with Birgitta Whaley, working on quantum dynamics in photosynthetic light harvesting (2008–2013).

• Limits of quantum speedup in photosynthetic light harvesting. S Hoyer, M Sarovar, KB Whaley. New Journal of Physics (2010).
• Faster transport with a directed quantum walk. S Hoyer, DA Meyer. Physical Review A (2009).