Generative AI in Energy: 201

The transcription of the video is below. 

Introduction and Overview

Jon: Hi, welcome back. This is Jon Bonanno from Fractal, and we have Kerim Baran as always from Solar Academy and we have Priya Donti, who is gracing us. She is the executive director of Climate Change AI and also a professor at MIT. We are going to reengage on this topic of learning more about artificial intelligence and how we can apply it or die from it, which is AI in climate solutions. So welcome back Priya, and thanks for making your time for us. 

Priya: Thank you. 

Jon: Excellent. So we had a wonderful first session where you were so clear on. Breaking out artificial intelligence, moving into machine learning. And then from there, we discussed a supervised, unsupervised, and reinforcement learning model. And if anyone hasn’t seen that, the link is below and the slides are below. And I think it’s really important to get that base. Before you come and watch this, but this is gonna be excellent cos Priya’s really gonna break apart how we can take those tools and apply them to climate solutions. So, before we get started, in the week that has spanned between last week, First recording, and now, artificial intelligence is moving so quickly, it’s like really hard to keep up. And I lived through the late 90 telecom stuff. I lived through the early search stuff. I lived through climate tech 1, 2, 3.0, et cetera. And the pace of change right now in artificial intelligence is shocking. Tell me, I’m gonna push this to Kerim first. In the last two, or three weeks, let’s say recently, what’s the thing that surprised you most about artificial intelligence?

Kerim: The frequency which I am going to chatGTP’s website. I would say 

Jon: That’s a good point, 

Kerim: is exponentially increasing.

Jon: That’s a good point. That’s a good point.

Kerim: And yeah I guess for me, that’s it. And I and how much I’m hearing it in every, every podcast and every piece of content that I’m consuming, I haven’t yet. Found a way to apply it to my day-to-day operation, probably because I’m not that into the weeds of coding and integrating technology, at this point in my life. But, yeah that is hat’s probably what I would say. And just to note for the viewer’s content, because it’s gonna take a little bit of time to produce this, we are in late May, early June of 2023. Right now,

Jon: Wow timestamping it as if someone’s gonna watch this capsule in 100 years and say, what were they doing?

Kerim: Yeah. So.

The Intersection of AI, Climate Change, and Policy Considerations

Kerim: Well, yeah, but Priya on to you.

Jon: Lemme injects here because Priya’s got a lot of work to do at the end here and we’re gonna be like the peanut gallery as she moves through more sophisticated content. The thing that really surprised me, cos I am applying it to our work at Fractal, and what I’ve found is the quality of the input is so important because we’re trying to, we’re trying to apply these tools to specific data sets and if the data has. messiness or noisiness in it, or inaccuracies, it really throws off the outcomes, which I guess is not surprising, but at some point it’s a point you made Priya on our last session where you said, hey, garbage in, garbage out, and now behold, we’re having to go back and fix the data so that we can, and that’s like, wow. We discussed it before too, so right on. You’re already in the last week I’ve experienced what you taught me, so thank you. And now today we’re onto you again. Thank you, Priya.

Priya: Yeah, absolutely. And I guess interestingly, something I’m gonna talk about that surprised me over the last week is not on the technology front as much as the policy front. So, as the EU AI act is sort of moving closer and closer to becoming, law, there actually now are some draft amendments in there that are explicitly talking about the interaction between AI and environment and how that affects how we should think about AI. So actually, you know, was it last year or anyway, I think around last year climate change, AI had put in a comment basically saying that when thinking about the risks of AI and what you wanna regulate and what you wanna, you know, ask for transparency about, thinking not just about the, you know, emissions of an AI model itself, which is really important to think, think about, also, it’s important to think about what the effects of the application are gonna be for climate, be that good or bad. And so, the draft amendments address that latter point when they have, a whole provision around what does it mean for an AI system to be classified as high risk? And there are various things about, you know, surveillance or ethical considerations of other sorts. And now if an, if an application has, a negative impact on climate and environment, Then in the draft amendments, that’s proposed to also be kind of flagged as a high-risk application that requires additional transparency and oversight. So that was, that was really cool to see.

Jon: That’s super insightful. I can remember the early days of blockchain when this was gonna be sort of the liberation of monetary exchange and, you know, In effect, I mean, what I see is just a massively increased load profile from all these data centers grinding away and burning up coal and natural gas electrons to fuel their pursuit of Bitcoin. So I, that’s very insightful and I hope that we do get some legislation, at least in the EU region, followed hopefully by some work that I know Vice President, Harris is working on with a bunch of stakeholders. So yeah, that’s a great sighting. 

Priya: Yeah. And the other I mean, thinking about the US policy side, the Biden administration sort of put out a, request for information on the AI strategy that they’re trying to develop at the national level in the US. And within that, they’re enumerated a bunch of different questions that they want, input on. And one of those is what are the implications of climate for the way we should be thinking about our AI strategies? So for those who, you know, have opinions on that, there’s actually, you know, I think some really interesting policy windows opening up, in, in multiple jurisdictions to actually affect that.

Jon: Right. And Climate Change AI as a nonprofit is also working with stakeholders to support strong policy, but one that allows us to have a robust ecosystem around solutions. Is that right?

Priya: Absolutely. I mean, I think in some sense, the whole goal is how do we accelerate climate action, use all the tools we have in the right places and responsibly in order to do that. And one of those tools is undoubtedly AI. So really just thinking about how do you kind of get the policy and research and entrepreneurial and industry landscapes to all really work together in service of that goal.

Jon: Well, I mean, I think that it’s a very noble nonprofit group that is worthy of foundation dollars and grants and donations from individuals. It’s really great work you and your team are doing at Climate Change AI, so thank you. I’m so glad that that’s happening. 

Machine Learning Role in Distilling Raw Data

Jon: But let’s get to your content today because I know that our viewers are probably interested in learning more.

Priya: Absolutely. So today, last time the overview I gave of AI and machine learning, I did, we did talk about some, you know, climate and energy-related examples, but it was in some sense a more general overview. So today I actually wanted to dive in and give a sense of what are actually some of the kinds of ways that AI and machine learning are used and can be used across the climate action space. And also give some concrete examples of where that’s happening. So I’m gonna go through, I believe, six themes, today. So first one. So, we’ve started to see, right? There are huge dreams of raw information available that kind of are hard to analyze manually. So things like satellite imagery, aerial imagery, and text documents, which invariably contain insights that are useful and actionable for how we act on climate change, but are hard to just kind of scrape and glean at scale.

So machine learning is one use case is to actually distill this raw data into actionable insights. So for example, one kind of coalition of organizations called Climate Trace is using a combination of satellite imagery on the ground sensor data and machine learning to try to generate an independent emissions inventory of greenhouse gas emissions. And the idea is that by generating something that is independent and auditable and outside of he political process, it can provide kind of an impartial input to macro-level efforts and to actually, think about how we should be, you know, allocating greenhouse gas emissions between countries or really trying to,figure out which facilities are emitting. This actually gets emissions down to the kind of facility level and entity level in many cases.

Jon:  Now can I inter… can I interject just one second. I wanna ask a question about Climate Trace. Now, that’s a for-profit company, or is that a nonprofit organization?

Priya: So it’s a coalition of organizations, some of which are nonprofit and some of which are for-profit.

Jon: Okay. Okay. 

Priya: And they’re all working together towards this goal with different organizations, dealing with different sectors within this, this, to try to get kind some quality, quality methods of getting estimates in different sectors. Given that this looks different in different. 

Jon: Great. Thank, thank you.

Priya: Yeah, absolutely. similarly, there are entities using, satellite imagery and machine learning to detect deforestation. For example, the MAP project, which is trying to detect deforestation in the sort of Andean Amazon region. And that’s sort of run by a nonprofit initiative. There, similarly academic efforts that are trying to detect. The locations and energy efficiency characteristics of buildings using satellite imagery. This was actually the theme of the women in data science data-ton, last year, which brought together like hundreds of participants to really try to just make progress on this, this initiative. Mapping.

Jon: So break that down a little bit. So the public use, there is a database or somewhere that there’s, every building is being tracked and that information is available. Is that right?

Priya: So what’s happening is right now we have, kind of granular information about the energy use of a small number of buildings. But there are many buildings where, first we don’t even know the building is there. And then second of all, we don’t know much about its energy use characteristics. So it’s an area of research at the moment or of kind of early practice to say, can we learn something about the correlation between what we’re seeing in our satellite imagery and the energy use of a building? And try to actually now extrapolate that to other buildings where we haven’t had that opportunity to get targeted and more granular data.

Jon: I see. So we get sort of a proxy, a appropriate. I see, I see. That’s helpful. 

Priya: Yeah. And proxy is a great word to be using here in the sense that for all of these applications I’m talking about, about, you know, turning satellite imagery into insight. The insight you get is never going to be as good as granular on-the-ground sensing. But the insight you get is potentially much more scalable. Yeah. And so you are, you know, getting proxies and you wanna treat them as proxies and really understand what the limitations there are. But it is sort of that ability to get information in places where you didn’t have that capability to do, to do fine-grain sensing. 

Jon: Excellent. 

Priya: Yeah. and then, yeah, on the you know, text side, there are, initiatives like climate policy radar that are really trying to aggregate, you know, what are the climate policies that different jurisdictions have made? Can we like, you know, OCR them and label them and various things like that. And then use that to actually, you know, analyze these policies and understand what worked, what were the outcomes for different jurisdictions, and as a result, what should we do?

Jon: That’s super interesting. 

Kerim: This is actually something that came up in my last two weeks. One of my portfolio companies is starting to use AI. I’m their first application is this to analyze policy across all 50 states and see, you know, which states are friendly to their approach and which are further along or further away. And, you know,… 

Jon: But like down to the city level. 

Kerim: Yeah. But again, the CEO’s comment to me was, it’s all about the data that you input. So it’s all about being able to find that tax that code that, you know, defines it and how you input it, I guess is probably another, challenge. But,…

Priya: Yeah, I mean, and also in addition to the data, the actual approach you take, so we talked last time about supervised and unsupervised learning approaches. Right? And, you know generative models like, like GPT are unsupervised. and as we talked about, you know, there are issues potentially with content quality and things like that, whereas there are actually a lot of supervised techniques that say, Hey, can I actually. You know I have a policy I know what’s in it. Can I label it in some way? Can I, and then I can, I can, I generalize from there. So actually there are different approaches there as well. And, some of the kind of existing supervised techniques actually work quite well for this purpose when applied properly. Yeah.

Jon: Wow. That’s great. and so climate policy radar is using supervised or unsupervised?

Priya: So I believe so. Their first step was actually on the data side. So they actually took this, created this large, you know, database of policies, and really got it, you know, into a place where it is ready for machine learning-based analysis. 

Jon: Ah, okay. 

Priya: And I think I, believe they are starting with supervised approaches, but they may be incorrect. I think they’re, 

Jon: So they first had to clean the data up.

Priya: They first had to clean the data up, that was a huge task. 

Jon: Yes. It’s a huge task.

Priya: Yep. Yeah, absolutely. So, yeah, this is kind of one theme. Basically, take your large raw streams of data and try to distill them into actionable insights, that allow you to make some more targeted decisions.

Jon: Excellent. So we’re, we’re looking at some images here. I’m not sure where that is physically in the world, but it looks like there’s some sort of crop issue or some learning something about crop stuff. Is this the type of satellite where you’re looking at, you know, deforestation or crop selection or something like this.

Priya: Exactly. So this is where you’re, this is a crop cover mapping image, which is trying to figure out kind of, you know, what are the, what kinds of crops are in a particular place and sort of leads to things like crop yield estimation and things like that, which, can also be helpful for adapting to climate change, obviously, as we start to see extreme events and they’re, their actually, you know, their, government programs, so like, NASA Harvest and the Land Monitoring Service are doing this kind of work. NGOs like Geogram and companies like Indigo, Atlas and sites. So this is actually an area that has gotten, you know, government and nonprofit and for-profit attention.

Jon: Excellent.

Priya: Yeah. 

Jon: Food’s important. 

Priya: Food’s very important.

The Role of Machine Learning in Forecasting

Priya: So, right. So, The next theme, maybe unsurprisingly, is forecasting. So there are lots of scenarios where having some kind of foresight can enable you to make a, you know, a better decision. and what machine learning is really useful here four is basically that its, allows you to ingest very different types of information and leverage those together to create a forecast. So I think I gave this example last time of, open climate fix, which is a nonprofit working with entities like National Grid, the UK power system operator to improve their forecasts of, renewable energy generation like solar power generation, and electricity demand.

And, what they’re doing is that they’re combining different data sources, so they’re combining historical time series data that’s telling you, you know, what did the production of your solar panel look like in the past along with, weather data and the outputs of kind of weather prediction, like physical weather prediction, modules, alongside things like imagery like video imagery or aerial imagery of clouds moving overhead of where solar panel sites might be, so, they’re able to take all of that and kind of feed it all into a machine-learning model and. Let the machine learning model figure out what is the relationship between these different data streams in order to produce a prediction of solar power, like a forecast in the sense of a very short-term prediction.

And, you know, using kind of these advanced, data source aggregation techniques and also by changing up some of the models that are used Open Climate Fix was able to, I believe, half the error of predictions that National Grid so was, was making. So, 

Jon: wow. 

Priya: This is something that, yeah, it kind of, they got really drastic improvements by leveraging the right data and the right modeling techniques.

Jon: That is enormous for them. That’s huge. And they also have load. Load information so they can,… 

Kerim: So, are the improvements mostly on the time or on the accuracy as well? Because I’m assuming this kind of forecasting was being done for the last decade or so. Like, so what is the, what’s the generative or you know? Well, 

Jon: Kerim, Kerim, let me pause you for a second. The one of the largest T&D operators in the United States captures video from helicopters and from drones of their remote assets, their lines, their poles, all this stuff, and they were doing less than 3% of human analysis on that. That was the total analysis that was done on their complete, you know, world of images captured. That is like nothing, right?  It’s like they didn’t, they, they shouldn’t have even looked. I mean, it was, it’s remarkably small. But now when you start to be able to churn through this stuff, it, you can get through maybe half of the data. That’s such an incredible delta, like this half the amount of errors. 

Kerim: Yeah. 

Jon: Is it must have been a game changer for this particular group and it was National Grid, right?

Priya: Yeah, it was National Grid and, yeah, so Kerim the improvement was on the accuracy side, as opposed to the speed side in that case. And, yeah, there’s been some evolution in how forecasting has been done. It used to be done using like rule-based technique. So just, you know, what time of day is it, what day is it, what part of the year? And then it moved to using, you know, simpler machine learning models and maybe kind of more standard sets of data. And now there’s just been kind of additional innovation both on the modeling side and on the, just which data sources you’re pulling inside. 

Kerim: Yeah, makes sense. 

Priya: Yeah. 

Kerim: Great. Thanks. 

Priya: Yeah. And this theme comes up again in a lot of different places of forecasting. So entities that are trying to do things like forecast demand for transportation in order to inform, build outs of transportation systems or extreme events. And when we say extreme events, I mean we often think, of course, like hurricanes and floods and these kinds of things. But for example, locust outbreaks are being, exacerbated by climate change, and kind of getting foresight on where those are happening and when is really important. So there’s a, company in Kenya, I believe, I hope I’m getting that right, a Kenya-based company called, Selenawamucii that developed an app called Kuzi, that basically provides more targeted forecasts of when locust outbreaks are going to occur in order to provide foresight to farmers.

Jon: Wow. And so what, just out of curiosity, in this case with Kuzi, now that you know it’s coming, what do they do?

Priya: Yeah, it’s a great question. So Kuzi itself, as far as I understand, they’re basically providing that phone app and foresight to farmers, and then they can kind of plan,  their planting season or also when they’re spraying. Cause one challenge is like if you spray too much, you’re kind of damaging your crop health, and if you spray at the wrong time. And so I think it’s, I think it’s largely like application timing, for example. 

Jon: Oh, okay. Okay. Okay. 

Priya: Yeah.

Jon: Thank you!

Priya: Yeah.

How Machine Learning Improves Operational Efficiency

Priya: All right. So both of the themes so far of kind of distilling raw data in a forecasting are both about how do you give someone insight in order to make a decision? But there are also various places where machine learning is used for automated decision-making. For example, control of some kind of automated system. So, for example, I mean there are, you know, many machine learning algorithms that are being used for optimization of heating and cooling systems. So, in residences, but also in kind of industrial facilities in data centers, in food storage. And so, for example, you have, you know this application got a lot of press, but you know, DeepMind was using machine learning to reduce the heating and cooling use in Google data centers. But then you also see, you know, myriads of companies, you know, for example, like Arup is doing some building optimization work in Hong Kong to try to figure out just like how do you reduce the energy use of buildings by controlling the heating and cooling systems more effectively. And also, you, of course on the residential side, for example, have entities like Nest. So this idea of just optimizing your heating and cooling system to make it more efficient is one use of ML. And how this works is basically, in your building, you should or likely have some, you know, sensors and these sensors are telling you something about what the current temperature of the building is, what the occupancy of the building looks like. And if you’re able to combine that with information about the building characteristics and what the weather looks like outside, you can really create a much more kind of nuanced strategy to keep your building or your industrial center within the correct temperature parameters but do that in a way that is reducing the overall energy use needed to do that.

Jon: Yeah, that’s Elexity is a company that I’m an investor in, and it’s, they’re doing this exact thing. They’ve figured out that if you can control the HVAC in most industrial buildings, it is an enormous percentage of their capacity charges and their usage. And if you turn it on at the right time, you turn it off at the right time, you monitor carefully that thermal float. So you use the building as like a thermal battery, but that intelligence is, is really hard to do. And so they’ve had their challenges like you’re saying that the data’s gotta be good inputs. You’ve gotta do the right things around supervised and nonsupervised, et cetera. But when you get it right, it really is, wow. It’s astonishing. So 50% or half the errors for National Grid sounds like a great outcome and, they’re seeing similar outcomes, too. It’s, it’s really impressive when you get this right.

Priya: Yeah, absolutely, absolutely. And, I think this idea of, yeah, like improving operational efficiency, it comes up in power grids. I think right now we’re at the kind of, research stage, but the French system operator RTE has basically put together this, challenge called learning to run a power network, which is basically validating the use of reinforcement learning to help us more efficiently run our power grids. And another one, and I think this is still more in the research realm, is often when freight transportation is allocated. This is cleared through freight auctions. So how do you actually just like improve the clearing process for freight auctions and like sort of clear them more quickly or more effectively? These are really complicated optimization problems. If you can figure out how to, like, speed up the optimization problem, how do you, you can just basically improve your freight clearing process. So there are lots of applications that come up there and just how do you actually just improve the efficiency of some kind of operational system?

Jon: Well you mentioned the French example, I believe on our first session. I looked into that, and it’s not only about efficiency, but there are two other motivations they’re trying to avoid substation upgrading and they’re also trying to put more clean electrons into the network. 

Priya: Yes. 

Jon: So it’s, you know it’s got us sort of so many good benefits to doing this type of analysis and creating these, you know, real-life testing environments for this, these types of things is fantastic.

Priya: Absolutely. And in their case, also trying to, build robustness to climate extremes or other kinds of outages that could occur. So, exactly, you want the ability to really take in and specify all these different objectives, and, you know, use machine learning in this case potentially with, I think as I mentioned last time, like a human in the loop in certain cases to really understand how you can kind of leverage your data in order to make nuanced decisions on your power grid that kind of, you know, yield outcomes that are good from, from the perspectives of these different objectives.

Machine Learning in Predictive Maintenance

Priya: Yeah. All right. Right. So, Another theme is predictive maintenance. So the idea that if you have some kind of operational system and it, you know, breaks or is about to break, this can often reduce the efficiency with which it operates or reduce the sort of robustness against some unexpected outcome. And so machine learning is often used to analyze data, about your system in order to either quickly identify a fault or an outage, you know, right after it occurs, or even preemptively try to identify that. So, For example, you know, methane leaks are a huge issue, as natural gas is transported to where it’s from, where it’s extracted to where it’s used. And machine learning is already being used to try to again, analyze satellite imagery in order to understand where kind of methane might be emitted, or aerial imagery as well, to just understand where methane might be emitted in places where we didn’t detect it. But there are also some more researchy applications that are looking at. Well, if you have some sensors along your methane pipeline or in your compressor station, can you detect anomalies even before that leads to a full leak or a full breakage? So you can detect that ahead of time and try to stop it before it occurs. 

Jon: Wow. So, I mean, this is like predictive O&M.

Priya: Yeah, exactly. Yeah. 

Jon: That’s really would be huge. 

Priya: Yeah. And those are still in research as far as I understand. The sort of like truly predictive maintenance part for methane. But it’s, it’s something that, you know, is very viable and I think could be quite impactful.

Jon: It also assumes an efficient marketplace or penalty for leaking methane? Right. so,…

Priya: And that’s actually, and that’s exactly the issue. It’s basically like what are the incentives of the existing players to install the relevant instrumentation and exactly. So that’s the that’s what I was sort of skirting around, but that’s exactly right.  So sort of, I think there’ll be a lot of progress that the right incentives were in place.

Jon: Right. I think a lot of these, I think a lot of the parties that are in the energy business now, they are saying, you know, we’d like to do this if there was a market for it. And okay if that’s possible, then great.

Priya: Yeah, absolutely. Yeah. And the other example and this is the example on the slide, and I realized the text is kind of tiny. but this is a figure from Deutsche Bahn, the German rail operator, who uses machine learning to, for example, detect issues in the switches on their rail infrastructure or just other things that may go wrong that might cause the trains to run, you know, not on time. And this is a place where you want your kind of transportation infrastructure to be more resilient to faults and outages that could occur, but also you wanna increase their competitive advantage against carbon-intensive forms of transportation. So they basically try to analyze all the sensor data around the system in order to understand where there maybe breakages, ahead of time or in real-time, and to fix them as quickly as possible.

Accelerating Scientific Experimentation Using Machine Learning

Priya: So, all of the applications I’ve talked about so far, they’re really in this idea, you have an existing operational system where you need to, you know, generate some insight to make a decision or just have the decision made in a semi-automated manner. But it’s also the case that there’s a lot of, you know, science and engineering that is not yet kind of deployed in an operational system, that, it can still benefit from machine learning. So, for example, machine learning is being used, to, help accelerate the creation of clean technologies by doing things like analyzing the outcomes of past experiments to suggest which experiments are the most fruitful to try next. So, the image here on the slide is from an academic paper that came out of Stanford back in 2020, and the author is now the, you know, co-founder and CEO of a startup called Aionics that uses the technologies that were developed there to accelerate the process of, battery creation or battery experimentation. What they basically do is they work with different battery manufacturers to accelerate, you know, their work by saying, basically, we’re gonna take your past data, we’re gonna combine that, with physical knowledge. They use physics-informed machine learning explicitly. Because in some sense, the amount of data you have about battery experiments is not super big data.

So you wanna leverage that data, but you wanna leverage the physical information that you have. And they use that to make predictions about what battery should, should that entity synthesize next. And in some cases, you know, they’ve cut down the number of design cycles by like a factor of 10 for some of the companies that they’re working with. And so this can really be a game changer when it comes to just speeding up the process of getting to your better battery.

Jon: Yeah. Massive iterations. Yeah, yeah, yeah. There is a company called H2U, no relationship to them at all, but Tom Werner, is on the board there, former CEO of SunPower, and what they’ve figured out, it’s technology out of Caltech, so similar to, professor Sandeck, that’s now spun out, but, Nate Lewis at Caltech has, is using an AI engine to do molecule discovery around replacing the exotic materials in electrolysis machines. And because those electrolysis machines have traditionally used palladium and, you know, some extremely exotic and pretty delicate materials that wear out, they’re creating these synthetic possibilities really is what it is. And it gives them a better chance of discovering something sooner. 

Priya: Yeah, exactly. So great. 

Jon: That’s, this is incredible. Like I love, I love seeing this.

Priya: Yeah. Yeah. I think this is a really like cool group of applications. I’m really excited about. And this, yeah, we talked about, you know, batteries and hydrogen and various other things, but this comes up in the discovery of electro fuels in next generation, you know, CO2 sorbents. so this comes up in a lot of different places. 

Jon: Yeah. Plastic inputs, safs, there’s so many applications for this. 

Priya: Yeah, exactly.

Machine Learning for Approximating Time-Intensive Simulations

Priya: Exactly. Yeah. And the last theme I wanted to talk about is, approximating time-intensive simulations. So there are lots of kinds of workflows in the climate and energy space where you’re running some kind of large physical simulation to gain some kind of insight. So climate modeling is a great example of this, right? What’s gonna happen with the climate in the future? And, and where? But also large-scale energy models that we run to optimize the power grid and also city planning models that really try to understand for a particular city, where are the buildings, how does fluid flow throughout the city, and how does that affect the kind of energy efficiency characteristics of the overall city? 

Jon: Right. Flooding?

Priya: Yeah, there’s two. And so, What machine learning can do? Machine learning is not gonna completely replace these physical simulations by any means. Like the physics, there are really important and, hard to capture unless you have literally unlimited data, which we don’t tend to. But what machine learning can do is actually approximate sub-parts of these models that are particularly computationally expensive. And then you can, you can plug your kind of machine learning-based approximator back into your larger model to enable the overall thing to run a bit faster. So, the image on this slide, for example, is from a paper that, was trying to provide a data set to enable the, you know, classification of the type of cloud that you might be seeing in an image. So map from, you know an image to a, to a cloud type classification. And, you know, clouds are some of the biggest sources of uncertainty in climate models. Cause the physics of each individual cloud is really, really complicated, and some clouds, you know, cool and some clouds form. And so, if you have the ability to use machine learning to just create an approximation of your individual cloud model and then have that runway faster and then plug that into your bigger climate model, then your overall climate model is less computationally intensive to run.

And you can start to do things like figure out, well, how do I, run the climate model at higher spatial granularity in order to get more targeted insights? Or, kind of a separate line of work is, people are doing things like you take your, you know, very, your more expensive, kind of less granular climate model, but then you take those outputs and use machine learning to downscale them to be more, more spatially granular. So basically there are ways that people are trying to make the model itself faster or to take your existing model and its outputs and kind of generate some more targeted insight from them.

Overview of Machine Learning Themes

Jon: Well I was just at the MIT energy event in Boston and there was a group looking at how climate will affect immigration patterns. And I mean, it’s just like such an enormous dataset and it has this, what did you say? Multiple sources of data that can sort of push out some new intelligence and insights. This is exactly what they were doing and their findings were not, particularly nice. there’s, there’s gonna be, literally billions of people on the move because of climate, extreme weather. And, it’s gonna cause a lot of friction between the haves or the current residents of those places and the new arrivals. So, this is, this, this approximating time-intensive simulation. I think they might have been using something like this or something similar. To indicate how people are gonna move and the volumes was just shocking.

Priya: Yeah, that definitely makes sense. Yeah, and I guess one thing I wanted to show, I’m not sure if you’re seeing this like overview slide now of like all the different applications. 

Jon: Yeah.This is great. 

Priya: Yeah. So those themes that I talked about, they all kind of, you know, as you saw, they really spanned different sectors like, you know, electricity systems and agriculture and transport, and climate prediction and industry supply chains, all of that. And so this is just a very, I know, like overwhelming figure, but, one that just really tries to show, you know, there’s so many different applications of the, she learning across these climate change-related areas. And the figures on the slide, they’re all from this paper tackling climate change with machine learning that a group of co-authors and I put out, about four years ago, at this point just trying to detail like what are the different ways that, AI machine learning can be used across the space. And, the space has progressed since then. I think we’ve seen a lot of the things that we talked about there as, you know, early-stage or more kind of, you know, more early-stage applications now you see startups and, you know, nonprofits and governmental entities, that have really made a lot of progress on there, but, for people trying to just get a mental model, or, or more detail beyond what we were able to talk about today about, you know, where is it that AI and machine learning can play a role? I would definitely encourage you to check out this paper.

Jon: Oh, that’s fantastic. And we’ll put the reference for the paper link in the notes as well below. What a good group of people look at these. Some of these names I actually recognized from our previous conversations. What a treat. Wow. This is like an information treasure trove for anyone that wants to dive in, we’re gonna have a heap of link links below to more information and so, are you good with your information that we wanted to share today? Okay, perfect. 

Priya: That was everything. 

Jon: Awesome, wow.

Wrapping Up and Final Thoughts 

Jon: Thank you. I feel like I’ve taken, a giant tub of information that I have to digest now, and hopefully, I’m like a sponge and I get some portion of it. We are, I think going forward it’s unclear how we’re gonna use this, but, I think we’d like to do maybe, some organization snapshots going forward. It’s unclear, but man, I am so grateful, Priya, for you sharing all of this information with us and the people in our community. And, we hope to get it out to people and they should really dive in. These tools, I believe, and I’m seeing day-to-day they’re very important and they can really create incredible leverage on for humans. This isn’t like an us against them scenario. This is really a potential partnership and it’s so beautiful. So thank you for being on Solar Academy and sharing like this. It’s really, really amazing.

Priya: Absolutely. And yeah, I mean, maybe just to add to that, I agree. I think basically like a lot of the most compelling applications of AI are ones where we really are kind of. Supporting or accelerating an existing climate change-related strategy. It’s not about kind of, you know, automating away the innovation that’s already going. It’s, it’s trying to make it more effective. But it’s also worth noting, I mean, we’ve talked today about the ways in which AI is an accelerant for climate action, but maybe even like bringing it full circle to, the EU AI act discussion and such before, it is again, worth noting that AI is a, is a general purpose tool. it’s being used in all sorts of ways across society to accelerate oil and gas exploration and extraction. I think as we talked about last time and, you know, accelerate consumption and advertising in itself has a carbon footprint. So I think really the best thing we can do here is leverage AI for climate action, where, we feel like it can accelerate, our work, but also really try to align the broader macro trends in AI, including business as-usual applications with climate action. Because we really need to do that whole thing. We can’t sort of let ourselves, use AI for climate action as a way to greenwash the sort of other trends that are going on in AI and just forget about their climate impacts. We need to think about the whole package and align the whole thing with our climate control.

Jon: Yeah, policy, policy, like you said, EU seems to be leading here. United States I’d hope would get on board. We should, we should hope that China and India would be interested in this type of thing. Everyone faces the same problems of clean air, clean water, clean food, you know, living in a built environment that is temperature controlled, moving people around medicine stakehold, there’s there, these challenges are universal to the human existence today. So,

Priya: Absolutely, and policy’s gonna be super important, but I think when people think of policy, they just think of public policy in the sense of governmental policy. But like organizational policy is also super important here, right? For example, does your organization have an internal carbon price? You don’t need to wait for there to be a governmental one to do that? Does that carbon price include not just your like sort of onsite and scope one and two emissions, but are you also, especially if you’re a tech company, vending solutions, are you really thinking about your scope three emissions and what the kind of net effect of the solutions you’re vending is. So, I think there’s a lot that can be done at the organizational level. Policy is super important, but it’s not an excuse to wait at the organizational level either.

Jon: Yeah, I mean, that’s a good call out to, you know, the big tech, which is, you know, Amazon, Meta, Google, et cetera, Apple. They have probably collectively, what, 400,000 employees and we can make, evangelists out of every single one of them in some capacity or another. So thank you for mentioning that. I also would be remiss in not saying once again, that, my experience with Climate Change AI, which is www.climatechange.ai, is an excellent nonprofit organization to support with foundation grants with PRI dollars from private offices or just individual giving. And they can go, you can go to that site and get that the tremendous work that Priya, Donti, and I should say Professor Priya Donti is doing, with her colleagues, is really, really ambitious insightful, and really inspiring. So thank you for the work you’re doing there, and it’s really great. So, Kerim, you wanna wrap up here? 

Kerim: Well, thank you both for all the information, you shared. Maybe as we finish, we can give a little preview of what the next session would be about. What do you guys think? I have some ideas if you don’t, but I think you have some things already in mind. Jon and Priya.

Jon: I’d like to dissect some individual cases, you know, have the CEO on and talk with her about, you know, the change that they’re looking to make. Maybe it’s for, we’ll do some smattering of for-profit and not-for-profit organizations. This is just a snapshot, you know, five, 10 minutes just digging in. You know, the normal stuff. I mean, that, would that be of interest to the viewers? Would that be.

Kerim: I think that would be great. Something else that I was thinking, and Priya, please feel free to finish it, finish it up on top of our comments. But I was thinking like, as I was. Listening to you, Priya, today I was thinking, so essentially this is like scaled computing at layman’s terms is what’s going on right now with Machine Learning and AI. Like it’s just another level of scaled computing that is happening. And having been in the tech sector, for a chunk of my career, the questions that come up for me are like, okay, I really wanna understand like the engines behind it, the tech stacks, the computing services, who is providing what, how do they all fit together? Are there like competing engines from Microsoft, Amazon, Google Meta, and then how do they partner with different, maybe software partners or services partners, consulting tech shops, and you know, like how does the whole ecosystem fit together? And then who is the innovative small groups that are like taking little pieces and making connections to create wow solutions for all these cases that you, you talked about. I would love to get a little bit wider understanding of that landscape and ecosystem for the benefit of our view viewers.

Priya: Yeah, that’s super interesting. And yeah, I think there’s getting somebody in with sort of an ML ops background who can really speak to that because I mean, there are, for example, different, you know, frameworks for the actual implementation of your, for example, machine learning or, or deep learning models. So I think, even last time we talked about like. Psyche learns as an open-source library and PyTorch and TensorFlow and all of these things. But then there are different cloud providers that, allow you to run these models. There are different services that help you with hyperparameter tuning so you can figure out kind of what the best configuration of your model is. I’m not necessarily like fully tuned into that full landscape, but I think somebody with an ML ops background definitely would be. So that could be, yeah. Super interesting.

Kerim: Yeah. Thank you very much for all of that. And even, I don’t know if this is a cutoff point or not, but may maybe we could even go into discussing like our aspirations with SolarAcademy to create an AI-powered, learning, platform for the the solar & storage and beyond, sector, because we talk about with especially with, well, Nico and Jon, we talk, we talk about, you know, chunking chunk, creating chunks of small chapters of video, audio content and then relating them to each other and creating different learning paths for people operating at different roles in the industry or who want to get into those roles. So can that be done at scale, is another question that we have.

Jon: I like your suggestion of like dissecting the plumbing a little bit. That would be pretty fun. You know, like if we could, it could be Priya’s guest and we wouldn’t even have to be there. You know, she just, we’ll just do a session where it’s a SolarAcademy session and the ML ops expert comes on and she starts talking about how yeah, we’re using this cluster with that thing and the other thing, and the, you know the tape goes there and then that works. And, you know, like, it seems like there’s a, there’s a lot of like, a kaleidoscope solution set right now, and the various pieces working together are that magic symmetry that happens. But it is magic and there are a few musicians that are getting it right and there aren’t that many of them, honestly. So how can we build the knowledge so that more people go, oh, they talked about that data set and whatever. Like maybe let’s try to educate to have more magicians. Something to think about. 

Kerim: Great. 

Jon: Anyway. This has been a great session. Thank you again to Priya. Thank you as always to Kerim. This is really fun and I look forward to seeing the finished product.

Jon: See you all.

Kerim: Thank you. Thanks. See you guys soon.

Jon: See you soon. Bye.

Priya: Bye