Call Her ‘Anima’

SC26 Welcomes Dr. Animashree Anandkumar as This Year’s Keynote Speaker

Animashree Anandkumar, Bren Professor at Caltech and co-founder of Accelerated Understanding, increasingly finds herself at the intersection of science and culture. Her trailblazing work developing artificial intelligence methods that can both simulate and understand complex physical system behaviors earned her a seat on the United Nations Secretary-General’s Scientific Advisory Board, as well as a spot on this year’s “TIME100 Most Influential People in AI.” On Tuesday, Nov. 17, 2026, at McCormick Place in Chicago, Anandkumar will take center stage as the SC26 Keynote speaker.

Anandkumar’s present focus on driving discovery using methods that commingle AI, high-performance computing, and physical sciences has been fueled in part by taking foundational skills and theory and expanding on them. Often, these “dives” coincided with optimal timing as technologies also were making notable gains. Her work on Neural Operators, which can learn mappings between physical fields and model phenomena across multiple scales and different science domains, first changed the speed and scope of weather forecasting. Beyond physics, her work in AI+Science spans into chemistry and biology, where she contributed to “GenSLMs: Genome-Scale Language Models Reveal SARS-CoV-2 Evolutionary Dynamics,” which earned the 2022 Association for Computing Machinery Gordon Bell Special Prize for HPC-Based COVID-19 Research awarded during SC22 in Dallas. 

From academia to industry, Anandkumar has taken her ideas out of the research lab to the TED stage and TIME100 red carpet. She has advised the White House and the U.N., earned multiple named fellowships, and accrued many accolades. Now, ahead of her SC26 Keynote, Anandkumar, who is well known as “Anima,” spoke with SC26 about the rapidly evolving connections among AI, HPC, and physics and where this convergence could take scientific discovery. She also shared some of the inspirations that have paved her way in the field.

Q&A with Anima

You have a decidedly diverse and global perspective that spans research, industry, and academia. What drew you to computing and artificial intelligence, and what keeps you excited about the field?

ANIMA: As someone who has been working in AI for more than two decades, my career has taken many unexpected turns. I started out in statistical signal processing that gave me the theoretical foundations of how to represent data, compress them, and make inferences. But I found the classical theory limiting when it comes to big data, which was beginning to come together thanks to the Internet revolution. This led me to dive into probabilistic graphical models to discover hidden patterns in large datasets in a structured manner. To design large-scale learning algorithms that could be run in a parallel manner, I ended up designing linear and tensor algebraic computations. That got me to go deeper into BLAS and CUDA libraries and design our own primitives for parallel tensor operations, even before deep learning and popular libraries like TensorFlow or PyTorch were constructed. Of course, once deep learning started going mainstream, the foundations we had built paid off, and we could rapidly innovate both on the algorithms and applications. The lesson I learned along the way was that we cannot put AI into silos or have strict boundaries. I had to constantly reinvent myself and challenge myself to learn new skills and collaborate with researchers across disciplines. That’s why I view AI as a great unifier.  

Your research has contributed to major advances ranging from weather forecasting to drug discovery and so much more. What is it about tackling diverse and challenging real-world scientific problems that continues to inspire you?

ANIMA: During my childhood, I was lucky to have exposure both to math and theory, through my grandfather who was a high school math teacher, and also hands-on engineering through my parents who ran a factory with computerized machines for the automotive industry. I was deeply inspired by Feynman’s lectures in physics during high school. I was always hoping that I could go back to science and engineering, even as I pursued AI research. About a decade ago, as deep learning started taking off and I arrived at Caltech, I thought that I had to give this all I had. Back then, there was well-founded skepticism if AI could have any impact on scientific domains since deep learning was only beginning to show results on computer vision and required enormous datasets, which were not available in most scientific areas. I decided to approach the problem by building the theoretical foundations, just as I had done earlier with tensor algorithms. What are the assumptions we can make about many scientific domains? It turns out that a wide range of physical phenomena, that may look very different, still follow common mathematical models, such as partial differential equations [PDEs]. If we could invent efficient AI methods that could solve PDEs, we could unlock progress in lots of areas. When we tried to apply existing AI algorithms, such as convolution neural networks or transformers, they were not sufficient since many physical phenomena are inherently multiscale and not limited to one resolution like standard AI assumes. This led us to come up with Neural Operators, a deep learning framework that is resolution agnostic and learns mappings between function spaces, including solution operators of PDEs.

We started with weather forecasting as a challenging real-world use case, after advancing from proof-of-concept datasets. Many weather scientists believed deep learning was, at least, a decade away from coming anywhere close to traditional forecasts when we got started. Thankfully, an open dataset with weather data for the last 40 years was available, and we decided to give it a try. Our first attempt ended up being competitive with traditional forecasts while being tens of thousands of times faster. What would earlier take a supercomputer could now be run locally on a gaming PC. This speedup also enabled us to do better risk assessment of extreme weather events, like hurricanes and heat waves, since we could now afford much larger statistical ensembles. Coincidentally, I first presented these results almost five years ago at my invited talk at SC21. We were the first to build such a high-resolution weather model and the first to open source it permissively. It became the foundation for NVIDIA’s Earth-2 effort that Jensen Huang personally conceived. That jumpstarted a revolution in AI weather and climate modeling. Today, we have not only weather agencies using AI models, but people around the world, including farmers in India for monsoon planning. Seeing this success in weather forecasting, we gained confidence to attempt many other challenging problems using Neural Operators, including modeling plasma in nuclear fusion reactors to prevent disruptions, designing a medical catheter that cuts down bacterial infections hundredfold, designing better chips and quantum devices, among many others.

You have described the need for AI systems that can understand and model the physical world. Why do you see that capability as an important next frontier for AI? What should the HPC community, and society for that matter, be focused on in terms of AI?  

ANIMA: The HPC community has played a pivotal role in making large-scale AI possible. From early development in linear algebra libraries, like BLAS and LAPACK by visionaries such as Jack Dongarra, to the birth of CUDA and using GPU accelerations for scientific simulations in national lab supercomputers, the HPC community built so many ingredients for the AI revolution. Most of this happened while AI itself was going through an “AI winter,” and many didn’t believe AI would ever pan out. That’s why it is so important to build the right foundations without being overfocused on one particular application. In the current era of large-scale supercomputing that AI needs, the role of the HPC community is even more important. I believe that AI requirements are not set in stone. They are ever evolving, especially since energy becomes such a critical bottleneck. We need agile large-scale computing that can adapt to these changing needs. New frameworks, hardware-software co-design, and self-improvement for better resource usage are all problems ideally suited for the HPC community to contribute towards.

More broadly, as a society, we need to think of large-scale AI in more diverse terms rather than just large language models [LLMs]. While there are many claims being made about LLMs solving cancer or discovering a room-temperature superconductor, these cannot happen in a vacuum. LLMs can only go so far as to provide new ideas, but we still need to go to the lab or do trials to test them out. That is the critical bottleneck, and that’s why we need AI that understands the physical world to remove this requirement. We have already seen this be immensely successful in areas – many I have already mentioned. So far, all of these areas have required separate models trained from scratch. LLMs have shown us that it doesn’t have to be that way, and one model trained on all tasks exceeds narrow models. Universality, meaning putting a range of physical phenomena in the same AI model, is the next frontier for physical understanding and simulation.

Much of the current conversation around AI focuses on increasingly large models and enormous computing requirements, as well as the resources they require. What do you believe is currently overhyped about AI? Is there an area that is not, but should be, receiving attention?

ANIMA: Conflating all of AI with LLMs is a big concern. I believe physical AI will be far more impactful than what we have already seen with LLMs. Moreover, physical AI is not just robotics or visual world models, like it is usually implied, but bigger than that. It encompasses not just the world we see, but the vast universe invisible to us, from atomic to cosmological scales. We need to put physics back in physical AI. 

Aligning AI with the laws of physics is essential for scientific discovery. By definition, discovery is about pushing the frontier, so that means it will not be present in the training data. So, we need AI that can extrapolate to regimes beyond training distributions, and the laws of physics enable it to do so. I believe, more broadly, such AI that understands and simulates physics is complementary to human intuition and perception, instead of competing directly with humans like LLMs do.

SC26 will include many students and early-career researchers. What advice would you give someone hoping to build a career at the intersection of AI, HPC, and scientific discovery?

ANIMA: Always be curious and creative. This is something I remind myself, and this doesn’t change because of AI. Removing boundaries across disciplines and continuing to pick up new skills are important in this age of AI. The HPC community has a lot to offer for further development and scaling of AI.

Can you provide a “quick tease” of your SC26 Keynote, perhaps one idea that you intend to bring into focus?

ANIMA: My Keynote will explore combining the physical simulation use cases we all know in HPC with the scaling and universality ideas from AI and the impact on many scientific areas. I will also cover some of our results at Accelerated Understanding, scaling these models to trillion context length in 4D across space and time.

After your Keynote in Chicago, what is the one question, possibility, or challenge you hope the SC26 community will continue to think about?

ANIMA: I want everyone to continue to think about AI that can simulate and understand physics, the opportunities it will unlock, and the scale required to make it happen.