AI systems trained to predict the next word in text may excel at mimicking human conversation, but they struggle when the task is understanding how the physical world actually changes. That's the insight driving a new wave of research at UC Santa Barbara, where scientists are rebuilding AI from the ground up to tackle materials science and medical imaging challenges that text-based models simply weren't designed for.

What's Wrong With Using Text-Prediction AI for Physical Science?

Large language models, or LLMs, have become household names for their ability to chat, write, and reason through problems. But their core strength, predicting the next word in a sequence, becomes a liability when scientists need to understand how atoms rearrange in a material or how a pregnant woman's brain physically changes over nine months. "Reality is not made of text," explained Nina Miolane, assistant professor of electrical and computer engineering at UC Santa Barbara and co-director of REAL AI (Reliable, Efficient, and Aligned AI) for Science.

The challenge is fundamental. Text-based AI systems ingest enormous quantities of written content and perform statistical pattern matching. "They're predicting the most probable next word, given the words that came before," noted Simon Billinge, director of the California NanoSystems Institute at UC Santa Barbara. When applied to physical systems, this approach misses the underlying physics entirely.

How Are Scientists Redesigning AI for Materials and Brain Science?

Researchers at UC Santa Barbara are pursuing several strategies to make AI more effective at understanding and predicting physical processes:

• Multimodal AI: Instead of relying solely on images, systems combine brain scans with medical records, patient notes, and other contextual data to provide richer information about what's changing and why.
• Synthetic Data Generation: For well-understood phenomena like brain atrophy in Alzheimer's disease, researchers use generative AI to create synthetic training data, reducing the need for millions of real images.
• Geometric Intelligence: Redesigning AI's fundamental building blocks to take advantage of the mathematical structure of data, not just the quantity of it.

Miolane's team is already applying these techniques to study how pregnancy and motherhood reshape the brain. "In the brain sciences, we have at best thousands of brain images for a particular condition," she explained, "so most of our work at REAL AI for Science is to rethink AI to make it more efficient".

Why Does This Matter for Materials Discovery?

In materials science, the stakes are equally high. Billinge and his colleagues at the California NanoSystems Institute study how atoms are arranged in materials, since those arrangements determine what properties the material will have and what it can be used for. "As materials scientists, we can play around with rearranging atoms to get all kinds of different materials," Billinge said.

Traditionally, scientists face two interconnected problems. The forward problem asks: given a material's atomic structure, what properties will it have? The inverse problem works backward: given a desired set of properties, what atomic arrangement would produce them? "Companies such as Meta, Google, and Microsoft are using AI to accelerate these efforts, which can lead to discovering new materials more quickly. AI is powering a revolution in materials science," Billinge noted.

The goal at CNSI is to capture the entire discovery process in large databases and develop autonomous machine-learning systems that can identify the atomic recipes for new materials without requiring researchers to manually test thousands of combinations.

What's the Real-World Application? A Pregnancy Brain App Launching in 2027

Miolane's research has already produced a tangible application. She and her team have developed digital twins, or 3D computer models that replicate the human brain. Using AI trained on these models, they can predict how a woman's brain changes throughout pregnancy and after birth. The work received a $1 million grant from the Chan Zuckerberg Foundation.

"I was in my pregnancy, waiting for this tsunami of hormones, and also waiting for the hormonal crash that follows birth. And it's not only a question for me, but for the approximately 200 million women who become pregnant each year, one in five of whom suffer from a post-partum disorder," said Miolane.

Nina Miolane, Assistant Professor of Electrical and Computer Engineering at UC Santa Barbara

When the app launches in 2027, users will be able to enter their gestational week and see the anatomical changes happening in an average pregnancy brain at that stage. A user who is 14 weeks pregnant and experiencing brain fog, for example, could see that the hippocampus, a brain region critical for memory, is losing some volume. Over time, as more data is collected, the app could show personalized predictions for how an individual's brain is changing, not just population averages.

The implications extend beyond pregnancy. Miolane noted that people routinely run financial scenarios when making major decisions like buying a house, testing different outcomes before committing. "But when it comes to health care, you might get some prescription options from your doctor, try a particular medication," she observed. Digital twins could eventually allow patients to see predicted health outcomes under different treatment scenarios, making medical decisions more informed and personalized.

Miolane

Is This a Fundamental Shift in How AI Tackles Science?

The work at UC Santa Barbara represents a broader recognition that one-size-fits-all AI architectures are insufficient for science. Text-based models excel at language tasks because language is fundamentally sequential and statistical. But the physical world operates according to different rules, governed by geometry, physics, and causality. Building AI systems that respect those rules, rather than trying to force physical problems into a text-prediction framework, may unlock discoveries that current approaches cannot reach.