When I set out to design a new course at Columbia that explored how machine learning (ML) could be applied to climate applications, I wanted to truly survey the landscape. I tried to define climate as broadly as possible, ranging from the science of Earth’s climate to how industries affect it, from climate policy analysis to ML’s own carbon footprint.

I had the privilege of working with amazing students who not only brought a diverse set of viewpoints to each week’s session, but also provided valuable input on how the course could be improved. Here are some takeaways.

Prior expectations and a predictable surprise

Prior to taking the class, most students were interested in three themes from the class:

1. climate policy,
2. geological climate science, and
3. the carbon footprint of machine learning itself.

The last point is clear—as practitioners of machine learning we want to quantify and minimize the impact of our own actions. The middle point is also sensible since much of this kind of work is done in academic institutions. The interest in climate policy surprised me, perhaps because of my own bias of knowing how hard this field is. Studying climate policy involves grasping complex political science concepts, working with challenging data, and embracing advanced machine learning topics such as causal inference. This is a topic I am considering covering over two sessions next time.

When polled at the end of the semester, students remained enthusiastic about the top three subjects. But the following two topics shot up in popularity:

1. manufacturing, and
2. sustainable buildings.

This is important. Manufacturing and how we heat and cool buildings accounts for over a third of global emissions (Gates, 2021); and yet, they are not common ML application areas typically taught in school. I’m glad to see students discovering new areas that they may choose to focus on in their studies and careers.

Learning objectives

My goal was to expose my students to a broad segment of how machine learning interfaces with climate applications. To this end, we read two papers per week. One about an application of machine learning to an aspect of the climate; the other, typically, a survey paper about the machine learning technology being used. I would argue that we covered almost the entire gamut of machine learning techniques, starting off from the basics of regression and exploratory data visualization, and ending up tackling advanced topics like reinforcement learning and causal inference.

While the class is a graduate-level course, I failed to recognize that all my students are in various points of their learning journeys. Some students were experts in specific topics, but none were familiar with all. Any insights that came up during discussions, especially practical tidbits, were well received. A snippet from some feedback:

Certain tips you gave/that were discussed along the class were really interesting, especially for CS majors (such as OneHotEncoding not suiting decision trees, and GridSearch being less desirable [and more carbon intensive] than RandomSearch, etc.). Students could make the best with more of these tips during the class!

This is something I intend to lean in on next time around; perhaps with a segment of each week’s session focused on practicalities of the ML method we’re studying.

What we read

Finding representative readings of how machine learning is being applied to such a broad segment of climate applications was difficult. The ML community is doing a fantastic job drawing attention to many climate applications. For example, Climate Change AI hosted a great workshop at NeurIPS 2021, encouraging ML practitioners around the world to explore the challenges we face. Workshop papers are meant to be exploratory and lower resolution than typical ML conference manuscripts; while they capture recent ideas and initiatives, they have some shortcomings. A student summarized it well:

I think I would have rather read more “solid” or descriptive papers rather than the small workshop papers that led to a lot of discussion and criticism, although it’s interesting to see ongoing research.

This is a delicate balance. When we read journal-length papers in other sessions, I received feedback that some students struggled to internalize the entire breadth of those studies on a weekly cadence. Going forward, I intend to complement workshop papers with some additional readings; as for journal papers, I’m thinking of pairing each with a reading guide to help students focus on salient sections.