A Scientist’s Life: Duncan Watson-Parris

Duncan Watson-Parris is an atmospheric physicist with appointments at UC San Diego’s Scripps Institution of Oceanography and Halıcıoğlu Data Science Institute. He received his bachelor’s degree in physics from Cardiff University in 2007 and his doctorate in theoretical physics from the University of Manchester in the United Kingdom in 2011. He convenes the annual “Machine Learning for Climate Science” session at the European Geosciences Union general assembly and co-convenes “AI and Climate Science,” a discovery series that is part of the United Nations’ AI for Good program. He joined UC San Diego in 2023.

explorations now (en): What do you do for a living?

Duncan Watson-Parris (DWP): I have the privilege of working here at Scripps as a climate scientist trying to understand the role of clouds in the climate system.

en: What are some of the main questions in your field?

DWP: Some of the big questions consider what the earth is going to look like in 20, 50, 100 years. How do we make projections about what the world will look like given certain assumptions about emissions and how we get there? 

One of the driving questions in our lab is around the details of how microscopic particles, specifically aerosol particles which are emitted from burning fossil fuels, interact with the climate system. Clouds are such an important part of the climate system. They cover something like two-thirds of the earth's surface at any given time and reflect about a third of all incoming sunlight straight back to space.

One of the things aerosols do is scatter sunlight back to space. They impart a little bit of cooling that way, but they also interact with clouds. All of the cloud droplets in a cloud are formed on little aerosol particles. If you have more aerosol, you end up with smaller cloud droplets. That makes the cloud slightly brighter. 

When we make them a little bit brighter, that imparts a large cooling on the system, which we need to understand and quantify. As we transition to renewable fuels in the future, that cooling, which has been masking some of the warming from greenhouse gases, is going to go away and we're going to end up with more warming in total. Being able to quantify and really understand this somewhat esoteric process enables us to make better projections about what our future will hold and how much warming we will experience.

We as humans have the most say over where we end up. We want to be able to be as precise as possible about what those choices will mean for our kids.

Climate modeling as a whole is an international affair. The U.S. has three or four climate models of its own, and there are other models from around the world that each also bring their own characteristics and have their own strengths. The data Scripps brings is somewhat unique, though, and so that comes to an initiative that I've been spinning up here at Scripps and more broadly at UC San Diego called the GAIA Initiative, where we're trying to harness the power of machine learning to unlock all this data that we've been collecting from around the world. We have this fantastic, unique set of data from Scripps ships and from the Argo floats that we produce. Artificial intelligence (AI) gives us an opportunity to learn more from that data and apply it more broadly. I'm pretty excited about where that will go, where that will lead us.

The scale is vast, right? We're going from these particles that you can't even see with your naked eye — micrometers or even nanometers in size — and we’re trying to understand the sum effect of that on the whole Earth. How do we model that? How do we ask questions about a system that is changing day to day, year to year on scales that we can't directly model? And how do we take the amazing observations that we do have of these processes, and scale them up to make inferences about the larger question? That complexity I find fascinating.

en: What tools do you use in your research? 

DWP: We use a combination of traditional climate models that we run on supercomputers that pretty accurately capture the large-scale features of the climate system. We combine those with remote sensing data from satellites and in situ measurements from the ocean, from the lab or from ships or aircraft. We do this in such a way to make the models as good as they can be to answer the questions that we want to answer. 

I started off working purely with traditional models that have hundreds of thousands or millions of lines of Fortran code, a programming language, which we run on supercomputers. Now we're actually working on a whole new class of climate models, which aren't written in Fortran, but rather written in modern programming languages that we can run on graphics processing units and tensor processing units that are super efficient for running machine learning algorithms. They give us new opportunities for making them faster, but more accurate as well, which is just pretty exciting. 

For these big, complex models, we need a lot of computational power. We use the Expanse machine at the San Diego Supercomputer Center to run this model across hundreds of processors to be able to  solve all the equations in parallel, which  allows us to run hundreds of years of simulations in only only a few days, which is fantastic.

In traditional computer models, we write out the equations we come up with that we think capture the essential physics and we write down the program that solves those equations. But when we come to trying to model problems of these scales, there isn't a set of equations that we can write that will exactly solve that. We have to come up with statistical relationships that we observe on scales of tens of kilometers or hundreds of kilometers of how clouds respond to different drivers. 

In machine learning, it’s about you telling the machine: Here's the data, here's what we observe, you tell me what that relationship looks like without providing quite so much of your prior knowledge into the algorithm. The machine is going to tell me what it thinks it can be. 

We have started leaning more on machine learning tools to do that, to make that easier and to actually build surrogate models or emulators of these climate models to enable us to do what we call the inverse problem of learning how to make a model look like real-world data.

en: Do you think development of the models will keep pace with climate changes themselves?

DWP: The questions I'm trying to answer are detailed to the second decimal. But for the first decimal point, we know enough and we have known enough for 10 or 20 years. So the question I'm asking is in the details, right? We know what we have to do: We have to reduce emissions at some point. We have to get to net zero. If we don't want to be warming any more, we have to not be net emitting CO₂. Hopefully by talking about the research, by making it clear how we understand the climate system and the amazing complexity of it, it will encourage people to take more ownership of it and to make those changes.

en: Why did you want to come to Scripps?

DWP: We're from the U.K. originally, but my family came out to California on a road trip to get the lay of the land and just get the kids excited about California — and San Diego's way up there. It's an amazing place, but more than that, Scripps itself is special. Scripps is one of the oldest oceanographic institutions in the world. It’s a privilege and an honor to be here. I'm super excited to be spinning up the lab, and I'm working with some fantastic people.