Connecting AI with the Physical World: Revolutionizing Science and Engineering

Background and Inspiration

• Speaker's parents were engineers who introduced computerized manufacturing in their hometown in India.
• Childhood fascination with computer programs impacting the physical world.
• Inspiration for AI research and connecting physical and digital realms.

Current Scientific and Engineering Challenges

• Science and engineering involve extensive trial and error and long hours in labs.
• Importance of experiments in validating findings and generating new ideas.
• Limitations of current language models: lack physical grounding and struggle with validating designs.

Bridging the Gap with AI: Neural Operators

• Need to train AI on observable world data and its hidden details for universal physical understanding.
• Importance of capturing phenomena from quantum chemistry to large-scale planetary processes.
• Neural Operators: A breakthrough AI technology that represents data as continuous functions or shapes.
• Allows zooming in indefinitely to any resolution or scale.
• Supports learning at multiple scales and incorporating mathematical equations to fill in details.
• Enabled simulations of physical phenomena like fluid dynamics much faster than traditional methods.

Medical Application: Catheter Design

• Medical catheters often lead to infections due to bacterial contamination.
• Neural operators used to redesign catheter interiors from smooth to ridged surfaces to create vortices.
• Result: Significant reduction in bacterial contamination verified through 3D-printed prototypes.

Weather Prediction Breakthrough: FourCastNet

• Traditional belief: Years or decades needed for AI to match traditional weather models.
• FourCastNet: First AI-based weather model that's high resolution and much faster.
• Successfully predicted Hurricane Lee's landfall 10 days in advance, outperforming traditional models.

Addressing Climate Change and Nuclear Fusion

• Application of AI in predicting and preventing reactor disruptions in nuclear fusion.
• Importance of these applications in tackling broader scientific challenges such as climate change.

Future Prospects: Generalist AI Model

• Current AI models are domain-specific.
• Vision: An AI model capable of solving any scientific problem, designing better technology, and benefiting humanity.
• Long-term goal: Scale up neural operators for general intelligence with universal physical understanding.

Conclusion

• Neural operators and AI are enabling transformative advancements in science and engineering.
• Ongoing work towards creating a generalist AI to solve a wide array of scientific challenges.

Thank you! (Applause)