Anima Anandkumar

Anima Anandkumar, also known as Animashree Anandkumar, is Caltech’s Bren Professor of Computing and Mathematical Sciences. Her research includes neural operators: AI models that learn mappings between function spaces, allowing scientific modeling to describe continuous systems rather than only fixed collections of data points. She coauthored the Fourier Neural Operator and FourCastNet papers, connecting this approach to partial differential equations and global weather forecasting.[2][5][6]

Anandkumar argues that scientific AI must represent what happens between measurement locations. Weather, for example, exists throughout the atmosphere, not just where satellites or instruments observe it. In an August 2026 explanation, she described neural architectures that approximate continuous space as more faithful to the physical world.[4]

Her collaborative work on extending neural architectures to function spaces offers a systematic way to adapt existing models for this purpose. As she put it: “You can take what researchers have already done with computer vision and language when they weren't thinking of the physical world and adapt it to make it work in the physical world in a very principled way.” Julius Berner and Miguel Liu-Schiaffini were the paper’s lead authors; the collaboration also included Valentin Duruisseaux, Jean Kossaifi, Boris Bonev, and Kamyar Azizzadenesheli.[4]

Caltech reported that, in fluid-motion and porous-flow tests, operators trained on 128 × 128 grids retained accuracy on 64 × 64 and 256 × 256 grids, while conventional neural networks generally became less accurate when resolution changed. These experiments illustrate the framework’s ability to work across resolutions in the tested systems.[4]

An earlier architecture supplies context for this work: the Fourier Neural Operator paper, first submitted in October 2020. Anandkumar coauthored it with Zongyi Li, Nikola Kovachki, Kamyar Azizzadenesheli, Burigede Liu, Kaushik Bhattacharya, and Andrew Stuart. It parameterizes an integral kernel directly in Fourier space. For partial differential equations, the model learns a mapping from functional parameters to solutions across a family of equations, rather than solving only one instance. The authors tested it on Burgers’ equation, Darcy flow, and the Navier–Stokes equation.[5]

Anandkumar helped develop FourCastNet, a global weather model using adaptive Fourier neural operators, with a team whose coauthors included Jaideep Pathak and Shashank Subramanian. In March 2026, she recalled training the model in 2021; the collaborative paper was submitted on February 22, 2022.[3][6]

The FourCastNet authors described short- to medium-range forecasts at 0.25-degree resolution, including surface wind speed, precipitation, and atmospheric water vapor. They reported matching the ECMWF Integrated Forecasting System’s accuracy at short lead times for large-scale variables and outperforming it for variables with complex fine-scale structure, including precipitation. They presented FourCastNet as a tool to aid and augment numerical weather prediction.[6]

Anandkumar leads Caltech’s AI+Science initiative. In a March 2026 interview, she said she co-founded it as AI4Science in 2018 and that its subsequent reframing reflected a relationship working in both directions: AI advances science, while scientific understanding—including physical laws—must inform AI. She sees that integration as necessary for trustworthy models and new discoveries.[3]

Caltech announced her appointment to the UN Secretary-General’s Scientific Advisory Board on March 16, 2026. Discussing the role, Anandkumar advocated scientifically grounded AI regulation that avoids favoring only large companies or obstructing open-source innovation. She also called for local-language models that reach people beyond already educated populations, framing linguistic access as part of an inclusive global approach to AI.[3]

Before her neural-operator work, Anandkumar studied efficient learning methods for latent-variable models. Her 2014 paper, “Tensor Decompositions for Learning Latent Variable Models”, with Rong Ge, Daniel Hsu, Sham M. Kakade, and Matus Telgarsky, developed parameter-estimation methods for Gaussian mixtures, hidden Markov models, and latent Dirichlet allocation.[7]

The method exploits special tensor structure in low-order observable moments, reducing estimation to an orthogonal decomposition of a symmetric tensor—a generalization of matrix singular value decomposition. The collaborators analyzed a robust tensor power method. Its computational tractability depends on that special structure; arbitrary tensor decompositions are generally intractable.[7]

Anandkumar earned her electrical-engineering bachelor’s degree at IIT Madras in 2004 and her PhD at Cornell in 2009. She conducted postdoctoral research at MIT in 2009–2010, was an assistant professor at UC Irvine in 2010–2016, and visited Microsoft Research New England in 2012 and 2014. She became a principal scientist at Amazon Web Services in 2016 and joined Caltech as Bren Professor in 2017, continuing her AWS work at the time.[2][8]

In her October 2017 interview, she explained: “Bridging the gap between industry and academia is really important. It is a big part of what brought me to Caltech.” Those partnerships, she said, helped connect mathematical theory with practical deployment. Her research judgment also emphasized understanding when simple algorithms succeed: “That is, stop thinking about worst-case scenarios and instead turn it around and ask, under what conditions do we expect simple algorithms to succeed?”[8]