How to Make AI Faster and Smarter—With a Little Help from Physics
Rose Yu has applied principles from fluid dynamics to enhance deep learning systems that predict traffic patterns, model the climate, and stabilize drones in flight.
At age 10, Rose Yu received a computer from her uncle—a rare gift in China 25 years ago. This sparked her early interest in technology, leading her to win awards in web design during middle school. She continued to excel in computer science at Zhejiang University, then pursued graduate studies at the University of Southern California (USC), where she earned her doctorate with an award-winning dissertation in 2017. Recently, she received the Presidential Early Career Award from President Joe Biden.
Now an associate professor at the University of California, San Diego (UCSD), Yu is a pioneer in “physics-guided deep learning,” blending physics knowledge with artificial neural networks. This approach has led to advances in traffic forecasting, faster simulations of turbulence for climate modeling, and tools to predict the spread of Covid-19. Her broader vision is an AI Scientist—an ensemble of AI tools that work alongside human researchers, grounded in physics, to accelerate scientific discovery.
Starting with Traffic: Combining Physics and Deep Learning
Yu’s interest in physics-guided AI began during her graduate studies at USC, inspired by the traffic congestion near the I-10 and I-110 freeway intersection in Los Angeles. She wondered if deep learning could tackle problems involving constantly changing data, unlike static images.
Her team treated traffic flow like fluid diffusion, governed by fluid dynamics. They represented the road network as a graph where sensors monitoring traffic are nodes, and roads between them are edges. This graph captured the average velocity of cars over time, enabling the prediction of traffic evolution.
With access to extensive LA traffic data, their model extended reliable traffic forecasts from 15 minutes to one hour. This breakthrough led to their code being integrated into Google Maps in 2018, and Yu later joined Google as a visiting researcher.
Applying Physics-Guided AI to Climate Modeling
In 2018, Yu expanded her work to climate science after collaborating with Lawrence Berkeley National Laboratory scientists. They focused on predicting turbulent flows—complex swirling fluid patterns critical to climate models but difficult to simulate quickly due to the heavy computations involved in solving Navier-Stokes equations.
Deep neural networks trained on high-quality simulations learned to emulate these calculations, speeding up predictions by 20 times in 2D and 1,000 times in 3D settings. This acceleration could improve models forecasting hurricanes and tropical cyclones.
The Many Faces of Turbulence
Turbulence appears in various contexts beyond climate. For instance, unstable blood flow can cause strokes or heart attacks. While at Caltech, Yu helped model turbulence affecting drone stability during takeoff and landing, improving drone control.
Currently, she collaborates with UCSD and General Atomics on fusion power research. Controlling plasma turbulence at extremely high temperatures is essential for fusion reactors. Physics-based models are slow, so Yu’s team is developing deep learning models to predict plasma behavior almost instantly, though this work is ongoing.
The AI Scientist Concept
Yu’s group has created AI algorithms that can automatically find symmetry principles from raw data—discovering known symmetries like Lorentz and rotational symmetry without prior training. These tools could potentially uncover new physics symmetries, a major scientific breakthrough.
This capability inspired the AI Scientist idea: an ensemble of AI programs assisting researchers by generating hypotheses, analyzing data, and aiding weather forecasting or causal inference, such as studying vaccination effects on disease spread.
The goal is a versatile foundation model capable of handling diverse data types—numbers, text, images, and video. While an early prototype exists, further development is underway to make it more comprehensive and intelligent, with a release expected within a couple of years.
How AI Scientist Could Assist Researchers
The AI Scientist acts as a scientific assistant, streamlining steps like literature review by rapidly summarizing vast amounts of information. However, AI cannot yet judge scientific validity or replace the nuanced creativity of human researchers.
Yu emphasizes that AI should relieve scientists of repetitive tasks, allowing them to focus on creative problem-solving. The vision is collaboration, not replacement.
For those interested in advancing skills in AI and machine learning, exploring specialized courses can provide practical knowledge to contribute to fields like physics-guided AI. Comprehensive AI training resources are available at Complete AI Training.
Summary
- Rose Yu combines physics and deep learning to improve predictions in traffic, climate, and plasma physics.
- Her traffic models extended forecasting accuracy and were adopted by Google Maps.
- She accelerates turbulent flow simulations, a key challenge in climate modeling and fusion research.
- The AI Scientist is a future AI assistant designed to support researchers by discovering patterns and generating hypotheses.
- AI aids in data processing and hypothesis formation but complements rather than replaces human creativity.
This approach shows how integrating physics knowledge with AI can lead to faster, smarter systems with practical benefits across multiple scientific domains.
Your membership also unlocks:
