AI That Connects the Digital and Physical Worlds | Anima Anandkumar

AI That Connects the Digital and Physical Worlds | Anima Anandkumar | TED

Growing up with parents who were engineers and pioneers in computerized manufacturing, I was fascinated early on by computer programs that extended beyond screens to create precise metal parts. This childhood curiosity deeply influenced my journey in AI research, inspiring me to bridge the gap between the physical and digital worlds. My current work focuses on transforming scientific research and engineering through AI, minimizing trial and error and enhancing efficiency in laboratories.

Bridging Science and Engineering with AI

Scientific and engineering advancements traditionally involve extensive trial and error. Great ideas must be validated through painstaking experiments, which often require long hours. While language models like ChatGPT can generate designs for complex objects such as aircraft wings or drones, they lack the physical grounding to assess these designs' viability. These models hallucinate—not understanding the physics, making text alone insufficient for achieving a universal physical understanding through AI.

Neural Operators: A Revolution in Scientific Simulations

To address this, our team developed neural operators. Unlike standard deep learning, which blurs details upon zooming in, neural operators represent data as continuous functions or shapes, allowing indefinite zooming into any resolution or scale. This technology enables us to train AI on multi-scale data and integrate mathematical equations to fill in finer details when data is limited.

With neural operators, we have achieved up to a million times faster simulation of physical phenomena, such as fluid dynamics, than traditional simulations. One remarkable application was in improving medical catheters. Bacterial infections from catheters are common, and our AI developed a ridged design inside the catheter to hinder bacterial movement. This optimized design reduced bacterial contamination by over 100-fold without the need for multiple trial-and-error iterations.

Extending AI’s Reach: From Weather Forecasting to Nuclear Fusion

Our neural operator technology has shown promise in various applications. For example, we developed ForecastNet, an AI-based weather model, which is significantly faster and often more accurate than traditional weather models. It successfully predicted Hurricane Lee’s landfall ten days in advance, whereas traditional models took longer to correct their forecasts.

Another application lies in nuclear fusion, where we use neural operators to predict the behavior of plasma in reactors, aiming to prevent disruptions and make fusion a viable clean energy source.

Towards General Intelligence with Universal Physical Understanding

While current AI models are specialized, our goal is to develop general AI capable of solving a wide range of scientific problems—designing aircraft, rockets, drones, and even medical drugs and devices. Such a universal AI model would greatly benefit humanity by enabling new, innovative designs and solutions across various domains.

Keywords

AI
Neural Operators
Scientific Research
Engineering Design
Fluid Dynamics
Weather Forecasting
Nuclear Fusion
Continuous Functions
Medical Catheter
Physical Phenomena
Universal Intelligence

FAQ

Q: What inspired Anima Anandkumar to work in AI?
A: Her fascination started with her parents, who were engineers, and early exposure to computerized manufacturing that connected the digital and physical worlds.

Q: Why are language models like ChatGPT insufficient for scientific research?
A: They lack physical grounding and cannot simulate necessary physics, which are critical for validating designs and hypotheses.

Q: What are neural operators?
A: A technology that represents data as continuous functions or shapes, allowing AI to zoom into any resolution and simulate phenomena at multiple scales.

Q: How have neural operators improved medical catheters?
A: They optimized the design to include ridges that prevent bacterial contamination, reducing infections by over 100-fold.

Q: What is ForecastNet?
A: An AI-based weather model that can predict weather significantly faster and more accurately than traditional models.

Q: How is AI being used in nuclear fusion?
A: Neural operators predict plasma behavior in reactors to prevent disruptions, aiding the development of viable fusion energy.

Q: What is the future goal of AI research according to Anima Anandkumar?
A: To develop a general AI model capable of addressing a wide array of scientific and engineering challenges, offering universal physical understanding.