Words like critical and essential and mandatory carry a lot of weight. Do you agree that is the case, or is that something we need to take a higher altitude look at where we’re going next?

I would say when it comes to the use of AI for climate action, there are places where it is critical, there are places where it’s helpful, and there are places where it is inapplicable or inappropriate. And so, we really do have to distinguish, making sure that we are leveraging these tools in places where they are the right fit.

When it comes to power grids specifically though, I will say “critical” in that there are a couple of things that are happening as we move to next-generation power grids with large amounts of renewables. We have to manage the power grid at greater speed to deal with the real-time variation in how much power is being supplied into the system because solar power is varying based on the weather, wind power is varying based on the weather, and our existing optimization and control techniques where we write down the equations of the power grid and explicitly solve over them are not able to solve fast enough for that to happen.

We also need to manage the grid at greater scale because we have more devices coming online, not just distributed solar and wind power, but also electric vehicles, batteries, and similarly doing this in a top-down, physics optimization way isn’t working.

On the power grid, a large amount of data is being collected through sensors. That provides an opportunity to manage the grid in a much more dynamic and data-driven way. And so, when we think about analyzing patterns automatically in large amounts of data, well, that’s exactly what machine learning is.

What in your field are you most excited about right now?

I am really excited about the directions for machine learning to contribute to optimization and control of power grids. So, machine learning on power grids has been used for a very long time. The process of, for example, forecasting electricity demand to figure out how to balance the grid used to be done using rule-based systems. It’s been common that you would use machine learning to do this by just learning from past data how to predict future electricity demand.

But where there has been a harder time making inroads are those places where you’re actively optimizing the system. Because when it comes to a forecast, you can have multiple methods. You can look at all of them. You can figure out which one you like. You can figure out which one you trust. When you’re optimizing a system, you are deciding about what to do on the system and it changes the state of the system going forward. You might break the system in the process. So, there’s a lot of care that needs to be taken in that decision. I’m really excited to see that there actually is interest in thinking about this in power grids and that there has been a bigger mobilization of not just researchers, but also power grid operators and software providers who are coming together to try to address this.

What are the ways that AI is critical in solving the climate change problem?

AI is being used in a variety of ways, from helping us to take large streams of data like satellite imagery and turn it into actionable information that can be used for climate change-related decision making. For example, a coalition of entities called Climate Trace is using satellite imagery and on-the-ground sensor data to provide an independent third party estimate of where the greenhouse gas emissions in the world are coming from at the facility level as an input to the UN climate change negotiation. Information transparency is the goal.

One thing we haven’t touched on is the role of AI in accelerating certain time intensive simulations like climate models, energy models or city planning models that model how wind is flowing through a city to figure out the overall efficiency of the city. These physical models are expensive to run. There are many uses of machine learning to better speed those things up to downscale the output so that they can give you more relevant information at the city level rather than just the country or region level. There are lots of ways, from providing information to automatically optimizing systems to facilitating science and engineering, that AI is playing a role.

There are some places where I do think AI is critical. For example, a decarbonized power grid without using large-scale data-driven analysis. And similarly, colleagues have argued things like large-scale biodiversity monitoring. Those are two examples where taking large amounts of sensors and turning that into biodiversity information or optimizing and controlling a system at the scale of the power grid in a way that facilitates renewables. That’s where — from a technical perspective — I’d say AI and machine learning are critical.

In places where AI is speeding things up, that could have otherwise happened at a larger time scale, that is also another form of criticality. Some of these things like getting that better battery by accelerating experimentation or getting a better climate prediction at a particular city, for example, maybe there are other ways to have done that that could have taken more time, but maybe AI speeds up the ability to do that in a way that deals with the criticality of moving quickly.

It sounds like the line between where AI is helpful and critical is thinner than the line between where AI is critical and absolutely inappropriate.

One way in which AI can be not wholly inappropriate, but maybe not the right first step can be in situations where there exists another step that you should take before leveraging AI. For example, optimizing the heating and cooling systems in buildings. You can put all sorts of fancy automation on your heating and cooling system. But if you haven’t insulated the building in the first place, you’ve missed a critical, less high-tech way to do something very good. But then also, your automated system might be miscalibrated because then when you put the insulation in, it has to relearn how to optimize the building in that context. So, there are places where AI is helpful, but it’s not necessarily the first step.

There are other places where I would say that AI is a Band-aid solution — and it doesn’t mean it’s a bad solution — but maybe the real solution lies elsewhere. For example, there’s a lot of really cool work that’s looking at how to take corporate financial disclosures and mine climate-relevant information from those to help to better figure out what corporate action on this should look like, and to shape regulation. And that’s a great application because it basically says this is the information we have, and we have to leverage it in some way. But if the longer-term idea is we want to use corporate disclosures to be able to analyze information at scale, maybe asking corporations to give their information in a structured format rather than text is the longer-term solution. So, figuring out where AI is being used to compensate for the fact that we have a system that is imperfect, and doing that, but then realizing what it means to make that system better rather than, without thinking about it, digging into the AI side as the actual perfect solution.

A place I would say AI is truly inappropriate is where it is being used to pretend that a decision is objective when really a decision is value-laden. When it comes to climate related policymaking, there are lots of hard decisions we’re going to have to make in terms of how we spend our money, with who are the winners and losers of how we spend that money. And putting a bunch of data into an AI system and expecting it to spit out the optimal policy is something that people have pitched to me before but is a way to basically shunt a lot of the decisions to actually the biases that exist in the underlying data rather than facing them head on. AI can maybe provide an assist to that. I talked about information transparency, for example, but you can’t expect an AI system end-to-end to make a big policy decision for you. ■