66°

Machine Learning for the Physical Sciences with Dr. Jacob Bortnik

 

Dr. Jacob Bortnik is a pioneering space physics professor in the Department of Atmospheric and Oceanic Sciences at UCLA. His research focuses on a variety of topics related to space weather in the Earth's near-space environment including satellite data analysis, numerical/computational modeling, laboratory plasma studies, nonlinear wave-particle interactions, and the application of machine learning in space weather prediction/specification. His latest article, "Ten Ways to Apply Machine Learning in Earth and Space Sciences," became Eos's lead story on Friday, July 9, 2021.

In the fall of 2020, Dr. Jacob Bortnik taught AOS C111/C204: Introduction to Machine Learning for Physical Sciences for the first time. We spoke with him to learn about the development of the course, its results, and machine learning’s importance and potential for the physical sciences. 

Pictured: An example of "image transfer" of a coronal mass ejection from the sun, in the style of Van Gogh. Courtesy of Dr. Enrico Camporeale.


 

What is machine learning (ML)? How does it relate to the physical sciences?

Machine learning was a term first used by Arthur Samuel in 1959 and refers to the "field of study that gives computers the ability to learn without being explicitly programmed.” It relates to the physical sciences in the sense that computers can run through very large quantities of data and discover hidden patterns in the data without being explicitly told what they are, or what to look for by a human programmer. Such patterns are often too subtle for humans to detect, and the quantities of data we're getting back from experiments these days are often too large for humans to go through manually.

 

Why was it important for you to teach ML and create AOS C111/C204?

The class itself was taught in the fall of 2020 for the first time, but the idea for this class began several years prior, when I noticed a lot of our graduating students moving into data science careers. I was curious about how they made such a seemingly dramatic change in their careers, and it turned out that they were essentially learning data science independently, reading books, or taking short courses or bootcamps online to acquire the necessary tools. It seemed to me that a top-ranked university like UCLA should have a single, comprehensive class that would quickly outfit students with these essential skills, and I resolved to do something about it. 

 

What research influenced you to develop this class?

Back in 2012 or so, machine learning was exploding with autonomous vehicles and machine vision becoming a reality for the first time. It seemed like magic to me, but what really sparked my excitement was seeing a student I collaborated with turn a large amount of data (that I’d worked with previously) into an animation that revealed some features I had been looking for quite clearly. I didn’t know how he did it, but I was determined to find out!

 

How will ML benefit AOS, the physical sciences, and the world at large?

The amount of data in the Earth and Space sciences (i.e., the types of data we at the AOS department work with) is growing very fast, much faster than Moore's law. As a result, the traditional way of analyzing data will not work for much longer. There need to be new tools that fully exploit the capabilities of these gigantic new datasets, looking for subtle patterns in the data that could not be found by a human looking through images one by one, and machine learning is that tool.  

So, what can it do? There are a range of applications from identifying subtle features in data (think about gravity waves in LIGO data, or galaxy clusters in high-res nighttime images), to using data to rebuild and predict the evolution of entire 3D environments like in space or in the ocean. A new area that we're interested in is a range of algorithms that can take in a large volume of data from measurements and discover the equations that describe that data. Imagine that! In the future you might be able to send probes into remote and barely accessible places with algorithms on board to describe the natural physical laws that govern those environments.

 

What are the results thus far from teaching the course at UCLA?

The results have been wonderful! As the instructor, I have gained a tremendous amount of depth into the material by teaching and thinking about the interesting questions that students raise. 

But for the students, the class worked almost exactly as I had hoped: everyone walked away with a set of practical skills, background theory, and hands on experience that allowed them to seek attractive jobs, become more competitive for graduate school, or apply new tools to their own datasets. One of my undergrads has even received the 2021 UCLA Library prize for her project! The course has been such a success that it will be offered each year in the fall and summer!