Ferran Alet | Google DeepMind

Ferran Alet

I am a Staff Research Scientist and Weather Science Lead at Google DeepMind in London. I completed my PhD in Computer Science at MIT CSAIL, advised by Leslie Pack Kaelbling, Tomás Lozano-Pérez, and Josh Tenenbaum. Before MIT, I completed the double degree in Mathematics and Engineering Physics at UPC - CFIS in Barcelona as valedictorian.

Research: I work on machine learning for Science and Sustainability. My team develops AI weather forecasting systems which achieve state-of-the-art predictions for everyday and extreme weather, including tropical cyclones. I've also worked on AI4Math, meta-learning, algorithmic discovery, and encoding&discovering physics inductive biases.

Twitter / Email / CV / Google Scholar / LinkedIn / GitHub

Featured In • Wired • The New York Times • The Washington Post • BBC News • Financial Times • New Scientist • Quanta Magazine • The Guardian • MIT Tech Review • Bloomberg • La Vanguardia • Diari ARA • TV3

Selected Research

	Operational Tropical Cyclone Forecasting with AI Ferran Alet, Thomas R. Andersson, Ilan Price, Stratis Markou, Andrew El-Kadi, Dominic Masters, Amy Li et al. Nature*, 2026 Paper (Nature) DeepMind Blog Wired (quoted) New Scientist (quoted) Fast Company (quoted) NYT (quoted) The Guardian Nature News La Vanguardia (interview) We present WeatherNext Cyclones, an operational machine learning model forecasting tropical cyclone tracks, intensity, and structure worldwide, providing up to an extra day of advance warning compared to state-of-the-art numerical physics engines.
	Probabilistic Weather Prediction with Machine Learning (GenCast) Ilan Price, Álvaro Sánchez-González, Ferran Alet, Thomas R. Andersson et al. Nature, 2024 Paper (Nature) The New York Times The Washington Post Financial Times The Guardian MIT Tech Review Ars Technica Nature News A generative diffusion model for global ensemble weather forecasting up to 15 days out, outperforming the operational European Centre for Medium-Range Weather Forecasts (ECMWF ENS) across 97% of target variables and extreme weather events.
	Learning Skillful Medium-Range Global Weather Forecasting (GraphCast) Remi Lam, Álvaro Sánchez-González, Matthew Willson, Peter Wirnsberger, Meire Fortunato, Ferran Alet, Shreya Ravuri* et al. Science, 2023 🏆 MacRobert Award Winner 2024 🏅 Science Breakthrough of the Year Runner-up Paper (Science) The Washington Post The New York Times BBC News Financial Times The Guardian Wired The Verge El País (interview) La Vanguardia (interview) Diari ARA (interview) TV3 A machine learning weather forecasting system based on graph neural networks operating on icosahedral multi-mesh representations, generating 10-day global forecasts in under a minute with higher accuracy than ECMWF HRES on over 90% of verification variables.
	Tailoring: Encoding Inductive Biases by Optimizing Unsupervised Objectives at Prediction Time Ferran Alet, Maria Bauza, Kenji Kawaguchi, Nurullah Giray Kuru, Tomás Lozano-Pérez, Leslie Pack Kaelbling NeurIPS, 2021 • Spotlight at Physical Inductive Biases Workshop Paper (arXiv) 15-minute Talk We optimize unsupervised objectives tailored to each specific test input at prediction time. By adapting where we act, we bypass generalization gaps and can enforce a wide spectrum of physical and domain-specific inductive biases.
	Meta-Learning Curiosity Algorithms Ferran Alet, Martin Schneider, Tomás Lozano-Pérez, Leslie Pack Kaelbling ICLR, 2020 Paper (OpenReview) Code (GitHub) MIT News (interview) VentureBeat TechCrunch By meta-learning domain-specific computer programs rather than neural network weights, we dramatically increase out-of-distribution generalization, discovering novel curiosity exploration strategies in simple toy environments that transfer to complex visual domains.
	Graph Element Networks: Adaptive, Structured Computation and Memory Ferran Alet, Adarsh K. Jeewajee, Maria Bauza, Alberto Rodriguez, Tomás Lozano-Pérez, Leslie Pack Kaelbling ICML, 2019 • Oral Presentation (Long Talk) Paper (arXiv) Oral Talk Code (GitHub) We introduce Graph Element Networks, combining graph neural networks and attention to learn adaptive computational meshes that map continuous physical functions to functions, with applications across continuum mechanics, Poisson problems, and neural scene representation.

Other Papers & Preprints

	Skillful Joint Probabilistic Weather Forecasting from Marginals (WeatherNext 2) Ferran Alet, Ilan Price, Andrew El-Kadi, Dominic Masters, Stratis Markou, Thomas R. Andersson, Jack Stott, Remi Lam, Matthew Willson, Álvaro Sánchez-González, Peter Battaglia arXiv preprint arXiv:2506.10772, 2025 Paper (arXiv) Code (GitHub) Bloomberg (quoted) The Guardian The Verge We present WeatherNext 2, a probabilistic global weather forecasting model trained to predict joint multi-variable weather distributions from marginal loss formulations, demonstrating state-of-the-art fidelity on extreme event forecasts and commodity/energy metrics.
	Discovery of Unstable Singularities Y. Wang, M. Bennani, J. Martens, S. Racanière, S. Blackwell, $\dots$, Ferran Alet et al. arXiv preprint arXiv:2509.14185, 2025 Paper (arXiv) Quanta Magazine Feature Using machine learning and non-convex optimization to search for candidate blow-up singularities and hidden glitches in the partial differential equations governing fluid dynamics.
Gemini	Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities Gemini Team (incl. Ferran Alet) Technical Report, 2025 Report (arXiv) Reuters TechCrunch Wired Next-generation multimodal reasoning and frontier agentic evaluation across scientific domains.
Diffusion	Gemini Diffusion Gemini Diffusion Team (incl. Ferran Alet) Google DeepMind Blog, 2025 DeepMind Blog VentureBeat TechCrunch Advancing generative diffusion architectures for multi-domain foundation models.
	Functional Risk Minimization Ferran Alet, Clement Gehring, Tomás Lozano-Pérez, Kenji Kawaguchi, Joshua B. Tenenbaum, Leslie Pack Kaelbling arXiv preprint arXiv:2412.21149, 2024 Paper (arXiv) We analyze fundamental contradictions in standard empirical risk minimization for continuous function spaces, deriving an axiomatic framework for loss functions defined directly over function spaces.
	Omnipush: Accurate, Diverse, Real-World Dataset of Pushing Dynamics with RGB-D Video Maria Bauza, Ferran Alet, Yen-Chen Lin, Tomás Lozano-Pérez, Leslie Pack Kaelbling, Phillip Isola, Alberto Rodriguez International Journal of Robotics Research (IJRR), 2022 • (Conference version in IROS, 2019) Paper (IJRR) Project Website Code (GitHub) MIT News A diverse benchmark dataset of 250 objects pushed 250 times each with calibrated RGB-D tracking, establishing the first precision benchmark for probabilistic meta-learning and physical dynamics modeling.
	Noether Networks: Meta-Learning Useful Conserved Quantities Ferran Alet, Dylan Doblar, Allan Zhou, Joshua B. Tenenbaum, Kenji Kawaguchi, Chelsea Finn NeurIPS, 2021 Paper (arXiv) Website Code (GitHub) Interview (10k views) We propose to encode continuous dynamical symmetries as self-supervised conservation losses and meta-learn them directly from trajectories, discovering invariant conservation laws that regularize sequence prediction.
	A Large-Scale Benchmark for Few-Shot Program Induction and Synthesis (ProgRES) Ferran Alet, Javier Lopez-Contreras, James Koppel, Maxwell Nye, Armando Solar-Lezama, Tomás Lozano-Pérez, Leslie Pack Kaelbling, Joshua B. Tenenbaum ICML, 2021 • Spotlight Paper & Benchmark Website A few-shot program induction benchmark capturing over 200,000 subprograms obtained by running real-world code instruction-by-instruction to generate rich input/output execution specifications.
	Neural Relational Inference with Fast Modular Meta-Learning Ferran Alet, Erica Weng, Tomás Lozano-Pérez, Leslie Pack Kaelbling NeurIPS, 2019 Paper Code (GitHub) We formalize neural relational inference as modular meta-learning and accelerate modular optimization by two orders of magnitude, making combinatorial module routing practical.
	Modular Meta-Learning Ferran Alet, Tomás Lozano-Pérez, Leslie Pack Kaelbling CoRL, 2018 Paper (arXiv) Video Code (GitHub) We train neural network sub-modules to be mutually composable, adapting to unseen tasks by recombining primitive modules in novel configurations akin to linguistic compositionality.
	Finding Frequent Entities in Continuous Data Ferran Alet, Rohan Chitnis, Tomás Lozano-Pérez, Leslie Pack Kaelbling IJCAI, 2018 Paper (arXiv) Video We propose a stream clustering algorithm defining recurring entities as dense modes in continuous space with provable PAC performance guarantees.
	Robotic Pick-and-Place of Novel Objects in Clutter with Multi-Affordance Grasping and Cross-Domain Image Matching Andy Zeng, Shuran Song, Kuan-Ting Yu, Elliott Donlon, François R. Hogan, Maria Bauza, Da-Sheng Ma, Orion Taylor, Melody Liu, Emi Romo, Nima Fazeli, Ferran Alet, Nikhil Chavan-Dafle, Rachel Holladay, Ian Morona, Prem Nair, Dan Green, Ian Taylor, Weber Liu, Thomas Funkhouser, Alberto Rodriguez ICRA, 2018 🏆 Best System Paper Award by Amazon Robotics 🥇 1st Place in Stowing (ARC '17) Paper (arXiv) Talk Project Website MIT News The vision and manipulation architecture for the MIT-Princeton robot team winning the stowing task at the Amazon Robotics Challenge in Nagoya, Japan.

Invited Talks & Keynotes

NASA / USAID SERVIR Geo-AI Working Group Lecture, August 2026: WeatherNext: Machine Learning for Operational Weather and Tropical Cyclone Forecasting
NOAA NWS Office of Modeling and Development (OMD) Inaugural Seminar, June 2026: AI Weather and Tropical Cyclone Prediction
Universitat Politècnica de Catalunya (UPC), April 2026: AI for Science and Sustainability
WNI Weather & Climate Forecast Conference (WCFC 2025) Keynote (Tokyo, Japan), December 2025: Advancing the State of the Art in AI Weather Prediction
Oxford Saïd Business School Guest Lecture, March 2025 & June 2023: AI Weather Forecasting: from Scientific Innovation to Real-World Impact
Oxford Intelligent Earth CDT Seminar & Model Uncertainty Workshop, September–October 2024: Skillful Probabilistic Weather and Cyclone Forecasting with Machine Learning
Barcelona Analytics Keynote, April 2024: GraphCast: Machine Learning for Medium-Range Global Weather Forecasting
University of Cambridge Computer Science Seminar, February 2024: Graph Neural Networks for Weather Prediction (GraphCast)
Texas A&M University Colloquium, October 2023: GraphCast: Global ML Weather Forecasting at Scale
OVGU / ECMWF Workshop on AI in Weather Forecasting, September 2023: GraphCast: Learning Skillful Medium-Range Weather Forecasting
ECMWF Headquarters Technical Colloquium (Reading, UK), February 2023: Verification Metrics and Autoregressive Rollouts for ML Weather Models
OpenAI, April 2022: Why Adaptation is Useful Even if Nothing Changes
Princeton University, March 2022: Learning to Encode and Discover Structure
EPFL, March 2022: Learning to Encode and Discover Structure
Google DeepMind Seminar, March 2022: Tailoring: Adaptation is Useful Even When Nothing Changes
CMU Scientific ML Seminar, January 2022: Learning to Encode and Discover Physics-Based Inductive Biases
Caltech, January 2022: Learning to Encode and Discover Physics-Based Inductive Biases
DLBCN, December 2021: Learning to Encode and Discover Inductive Biases (Video)
Meta-learning and Multi-Agent Workshop, June 2020: Building Up Knowledge Through Modularity
ICML Graph Neural Network Workshop, June 2020: Growing from Simple Tasks to Complex Problems with GNNs
INRIA, April 2020: Meta-Learning Curiosity Algorithms
MIT Machine Learning Tea, November 2019: Meta-Learning and Combinatorial Generalization
UC Berkeley, October 2019: Meta-Learning Structure (Slides)
KR2ML @ IBM Workshop, September 2019: Graph Element Networks (Video)

Mentoring & Teaching

I deeply enjoy mentoring and collaborating with students and researchers. I was honored with the 🏅 MIT Outstanding Direct Mentor Award (awarded to 2 PhD students across all of MIT). Here are some of the talented researchers I have had the privilege to mentor:

Research Internships Supervised at Google DeepMind

Maksim Zhdanov: Forecasting impacts of weather extremes
Stratis Markou: End-to-end cyclone impact prediction (converted to Research Scientist FTE at DeepMind)
Amy Li: End-to-end cyclone track prediction as grid probability distributions
Andrew El-Kadi: Efficient operational fine-tuning of weather models (converted to Research Engineer FTE at DeepMind)
Landon Butler: Interpretability of AI weather models (co-mentored with Andrew El-Kadi)

Master's Theses Supervised at MIT

Shreyas Kapur (co-mentored with Josh Tenenbaum): Simulator-based modular few-shot inference and action → moved to UC Berkeley PhD
Dylan Doblar: Meta-learning and Enforcing Useful Conservation Laws in Sequential Prediction Problems → moved to NVIDIA
Martin Schneider: Program synthesis approaches to improving generalization in RL → MIT PhD, Co-founder & CEO of RemNote
Erica Weng: Modular graph-structured models for prediction and control → moved to CMU PhD
Paolo Gentili: Active learning using meta-learned priors → moved to Hudson River Trading

Undergraduate Researchers Mentored at MIT & UPC

Jan Olivetti: Planning in belief space with meta-learned priors → moved to Columbia University MSc
Javier Lopez-Contreras: Program synthesis & learning theory → visiting scholar at UC Berkeley
Adarsh K. Jeewajee: Graph element networks → moved to Stanford PhD
Max Thomsen (with Maria Bauza): GNNs for robotic gripper design → moved to MEng in MechE at MIT
Catherine Wu (with Yilun Du): Energy-based models for trajectory prediction → continued undergrad at MIT
Nurullah Giray Kuru: Tailoring for model-based RL → continued undergrad at MIT
Margaret Wu: Unsupervised approaches to program synthesis → continued undergrad at MIT
Edgar Moreno: Library-learning for program synthesis → continued undergrad at UPC-CFIS
Shengtong Zhang: Tailoring and adversarial examples → continued undergrad at MIT
Patrick John Chia: Compositional neural scene representation learning → MSc at Imperial College London
Catherine Zeng: Modular meta-learning for reinforcement learning → Harvard University
Scott Perry: Energy-based models → continued undergrad at MIT