Physics-Guided AI: Faster, Smarter Deep Learning

Rose Yu and Physics-Guided Deep Learning

Rose Yu, an associate professor at the University of California, San Diego (UCSD), is a leader in physics-guided deep learning. She has spent years incorporating physics knowledge into artificial neural networks. This work has introduced new techniques for building and training these systems and has allowed her to progress on real-world applications.

Yu has used fluid dynamics principles to improve traffic predictions, sped up turbulence simulations to enhance hurricane understanding, and created tools to predict Covid-19's spread. This has brought her closer to deploying AI Scientist, a suite of digital lab assistants. She envisions a partnership between researchers and AI tools, based on physics tenets and capable of yielding scientific insights. She believes combining inputs from these assistants may be the best way to boost discovery.

Combining Physics with Deep Learning

Yu's work began with traffic. As a grad student at USC, she wondered if deep learning could help with traffic problems. Deep learning uses multilayered neural networks to elicit patterns from data. While there was excitement about image classification applications, Yu wondered if deep learning could help with constantly changing problems. She and her colleagues found a new way of framing the problem.

A Novel Approach to Traffic Prediction

They thought of traffic in terms of diffusion, a physical process. In their model, traffic flow over a road network is like fluid flow over a surface, governed by fluid dynamics laws. Their innovation was to think of traffic as a graph. Sensors monitoring traffic on highways serve as graph nodes, and the graph's edges represent the roads between those sensors.

A graph provides a snapshot of the road network, showing the average car velocity at every point. Putting together snapshots spaced five minutes apart gives a picture of how traffic evolves, allowing for future predictions. The challenge in deep learning is needing much data to train the neural network. Fortunately, Cyrus Shahabi, one of Yu's advisors, had accumulated a lot of LA traffic data.

Prior to their work, traffic forecasts were reliable for about 15 minutes. Their forecasts were valid for one hour, a big improvement. Their code was deployed by Google Maps. Later, Google invited Yu to be a visiting researcher.

Climate Modeling and Turbulence

Yu began working on climate modeling after giving a talk at the Lawrence Berkeley National Laboratory. She and scientists there looked for a problem that would be a good testbed for physics-guided deep learning and settled on predicting turbulent flow evolution, which is a key factor in climate models and an area of major uncertainty.

Examples of turbulence include swirling patterns when pouring milk into coffee. In oceans, these swirls can span thousands of miles. Predictions of turbulent behavior based on solving the Navier-Stokes equation are considered the gold standard, but the calculations are slow, which is why there are no good models for predicting hurricanes and tropical cyclones.

Deep Learning for Turbulence Prediction

Deep neural networks trained on numerical simulations can learn to emulate those simulations by recognizing properties and patterns in the data. They don't have to go through time-consuming calculations to find approximate solutions. Their models sped up predictions by a factor of 20 in two-dimensional settings and by a factor of 1,000 in three-dimensional settings. Their turbulence prediction module might someday be inserted into bigger climate models that can do better at predicting things like hurricanes.

Turbulence also shows up in blood flow, which can lead to strokes or heart attacks. Yu coauthored a paper at Caltech that looked into stabilizing drones, as propellor-generated airflows interact with the ground to create turbulence that can cause the drone to wobble. A neural network was used to model the turbulence, leading to better drone control during takeoffs and landings. Yu is working with scientists at UCSD and General Atomics on fusion power. One of the keys to success is learning how to control the plasma. At temperatures of about 100 million degrees, different kinds of turbulence arise within the plasma, and physics-based numerical models that characterize that behavior are very slow. They’re developing a deep learning model that should be able to predict the plasma’s behavior in a split second, but this is still a work in progress.

AI Scientist: Assisting Scientific Discovery

Yu's group has developed AI algorithms that can automatically discover symmetry principles from data. For example, their algorithm identified the Lorentz symmetry and rotational symmetry. While these are well-known properties, their tools can also discover new symmetries unknown to physics. These tools could also generate research ideas or new hypotheses in science, which was the genesis of AI Scientist.

AI Scientist is an ensemble of computer programs that can help scientists make new discoveries. Yu's group has developed algorithms that can help with tasks like weather forecasting, identifying global temperature rise drivers, or discovering causal relationships like the effects of vaccination policies on disease transmission. They are building a broader foundation model versatile enough to handle multiple tasks. Scientists gather data from instruments, and they want their model to include various data types. They have an early prototype and want to make their model more comprehensive, intelligent, and better trained before releasing it. That could happen within a couple of years.

AI can assist in practically every step of the scientific discovery process, serving as an AI scientific assistant. A large language model can read and summarize thousands of books during a lunch break, assisting in the literature survey stage of an experiment. While AI could help with hypothesis generation, experiment design, and data analysis, it cannot carry out sophisticated experiments. The goal is not to replace human scientists but to relieve researchers of some of the drudgery while letting people handle the creative aspects of science.