Anima Anandkumar bets on physical AI to solve science's hardest problems

Caltech professor and Accelerated Understanding co-founder Anima Anandkumar is developing AI models that simulate physical processes, helping researchers speed up scientific processes

For a young Anima Anandkumar, preferring neem over jaggery during Ugadi – the new year day for Kannadigas – was an obvious choice, much to the bewilderment of her mother.

 

“As a kid, I would always ask for more neem, which is the bitter part and didn’t want the jiggery as I was not interested. I want more challenges in my life. That's what you have to eat, more neem, to get more challenges in the coming year,” she says in an interview.

 

Anandkumar is a Professor at California Institute of Technology - better known as simply Caltech - of computing and mathematical sciences and co-founder of Accelerated Understanding, a San Francisco-based physical artificial intelligence (AI) start up. The company has developed a different type of AI model that, ‌in tests, handled 5 trillion pieces of data in a single prompt. And it is not a language model like the ones from OpenAI and Anthropic, but built specifically to meet the needs of the world of physics.

 

While language models are probabilistic, Anandkumar says she and her husband have built an AI model that replaces expensive experiments which need to be done to determine if a scientific idea works.

 

“It provides directional feedback to see how to improve upon ideas to make our inventions and discoveries even better. That's the key component that we feel is missing in current just language-based AI. Those models lack that physical grounding and understanding and that ability to simulate innately what's going to happen in those experiments in a physically valid way,” Anandkumar says.

 

For example, her team worked to come up with a better medical catheter that could reduce the backflow of fluids to prevent bacterial infection. Fluid flow and wave propagation are complex branch of physics which makes it hard to determine how the it changes if different types of triangular shapes are placed inside the catheter.      

 

“But we didn't have to 3D print those because we could simulate what different designs provide and come up with the best design that we believe will work really well in the physical world. And then we had to go 3D print and do the experiment only at the end which led to a hundredfold reduction in bacterial contamination.”

 

At the heart of this model is a technology pioneered by her years ago, called neural operators, that predicts physical phenomena in space and time. It is a deep learning architecture that learns mappings between infinite-dimensional function spaces rather than traditional finite-dimensional metrices.

 

Her AI model has already been embraced by weather agencies, who are able to predict natural disasters faster through better modelling. Other examples of her work are in chip design and coming up with cleaner sources of energy. The common problem currently in the semiconductor space is heating of the chips as they get smaller and more features are packaged into a miniscule space.

 

Accelerated Understanding, Anandkumar says, is already working with chip makers to conceive new kinds of chips. She declined to name the companies but said the work is primarily around new designs and ways to solve some of the hard physics that companies encounter when designing newer chips.

 

“Because we try to pack more and more transistors in smaller and smaller spaces, the question is how to lay that out to make sure the physics does not come in the way.”

 

Her work has already placed her among the 100 most influential AI personalities in 2026, according to TIME Magazine, alongside Vivek Raghavan of Sarvam, Sanjay Mehrotra of Micron, Peter Steinberger of Openclaw and Utkarsh Saxena of Adalat AI. After graduating from IIT Madras, she got her PhD in electrical engineering at Cornell University and worked with Amazon Web Services (AWS) and Nvidia before starting her company along with her husband Benedikt Jenik in early 2024.

 

“Her algorithms have since been used to forecast weather phenomena thousands of times faster than traditional methods, design novel medical devices and components for quantum computers, optimize a key step in computer chip manufacturing, conduct sustainable nuclear fusion simulations to one day safely harness their energy source, and much more,” TIME magazine wrote.

 

Anandkumar also declined to discuss her backers or how much the company has raised. She also has partnerships with computing providers, especially hyperscalers, who gave hardware clusters to develop and run the company’s model. She declined to name them too.

 

Having worked with language models at Nvidia, and now with physical AI models, she feels the world of the latter is getting bigger. “I've personally made a bet to work more on these less-explored greenfield area because there's so much more runway to go further.”