AI Startup Speeds Up Scientific Discovery With Physics Know-How

Scientists are harnessing artificial intelligence to tackle some of humanity's biggest challenges, and a new startup is leading the charge in a field where discoveries typically take decades.

Liam Fedus, former Vice President at OpenAI and co-founder of Periodic Labs, is building AI foundation models specifically designed to understand atoms. His company focuses on accelerating discoveries in material science, chemistry, and physics by teaching AI to recognize patterns in atomic behavior that human researchers might miss.

Fedus brings a unique blend of experience to the mission. He studied dark matter physics as an undergraduate before joining Google Brain and later OpenAI, where he helped develop ChatGPT. Now he's applying that expertise to solve scientific problems that could transform everything from drug development to climate solutions.

The shift reflects a broader trend of physicists moving into AI research. Fedus explained that the analytical thinking and problem-solving skills developed in physics translate perfectly to AI challenges. Many of his former colleagues at Google Brain and OpenAI share similar backgrounds, drawn by AI's potential to crack grand scientific puzzles.

The challenge differs dramatically from training language models. While ChatGPT learned from vast amounts of internet text, scientific AI models need highly specialized data from costly, time-consuming experiments. Periodic Labs tackles this by combining real-world experimental data with simulation data, helping AI learn patterns even when information is scarce.

The Ripple Effect

The implications extend far beyond the laboratory. AI that can predict how atoms behave could revolutionize drug discovery by identifying promising compounds faster. It could help engineers design better batteries, stronger materials, or more efficient solar panels. Climate researchers could use these tools to model new carbon-capture materials or develop cleaner energy solutions.

What makes this approach powerful is AI's ability to navigate vast amounts of scientific knowledge and spot connections humans might overlook. Instead of replacing scientists, the technology acts as a collaborator, suggesting efficient experiments and identifying promising research directions. This could compress discovery timelines from years to months in critical fields.

The movement of top AI talent toward scientific applications signals growing confidence that machine learning can handle complex real-world problems. As more physicists and researchers join this effort, the pace of innovation accelerates.

Scientific breakthroughs that once seemed generations away might arrive sooner than we think.