AI in climate

- This video covers the use of AI for climate resilience, from modeling and forecasting to inform policy and decision making to optimizing the deployment of clean energy infrastructure and assisting the development of greener technology. We will also consider how AI contributes to climate change itself due to the energy consumed to train and run data intensive models. To explain how AI is used to study climate change I'd like to introduce you to Dr. Solomon Hsiang, Chancellor's Professor in the Goldman School of Public Policy and Director of the Global Policy Laboratory at UC Berkeley. He co-founded the Climate Impact Lab and is a National Geographic Explorer and his team is integrating econometrics, spatial data science and machine learning to answer questions that are central to rationally managing planetary resources. - [Solomon] When you think about, you know like natural disasters affecting societies, and those are going to get more frequent with climate change, there's a lot of ways that we're using, for example satellite imagery to monitor what happens around the world, to monitor climate change, to monitor natural disasters and how they affect populations. And so AI's having a big impact on how we take huge amounts of data from satellites and digest it and turn it into information that could be used to respond to disasters. So that's like the kind of work we do here at Berkeley and in the lab. We're like trying to develop the techniques that turn this vast amount of unstructured data into actionable information for policy makers for decision making, to think about how you could like respond to climate change in real time. - In the United States, we have been doing little to measure economic impact on our nation's natural resources. But now satellite imaging, sensors, drones and other new tools are making it possible for environmental economists to monitor and manage a vast inventory of natural resources. On a global scale, Dr. Hsiang's data show that wealthy populations can protect themselves from the impacts of warming, but poor populations cannot. According to his Climate Impact Lab research, as temperatures rise, countries near the equator and in the southern hemisphere including most of the world's least developed countries, will experience the greatest increase in energy consumption. Many of the world's locations that are most heavily impacted by climate change are not as well connected to the internet or global media and therefore are data poor. Nowadays, satellites are changing this, but to provide access to this data and to make it useful to people without special technical skills, Dr. Hsiang and other researchers must distill the many layers of data embedded in satellite images and make the algorithms simpler so that the data and insights are widely understandable and usable by anyone. Let's take a look at weather modeling which can help communities better prepare for and respond to irregular weather patterns. The weather models we use were designed decades ago and are computationally bulky. While newer AI models are still fairly nascent they are more accurate and faster at forecasting. Current AI design uses about 7,000 times less computing power to create forecasts for the same number of points on the globe. Less computational work means faster results, according to University of Washington researchers. It also allows forecasters to run many models with slightly different starting conditions, a technique called ensemble forecasting that delivers a range of predictions for possible expected outcomes such as where a hurricane might strike. A big part of solving climate change is related to deploying and maintaining large scale infrastructure intelligently. For example, AI is being used to determine the optimal locations of windmills and solar farms to better meet energy demands. According to the World Economic Forum, AI can balance electricity supply and demand needs in real time, optimize energy usage and storage to reduce rates and better manage the complexity of decentralized renewable sources like microgrids, wind farms and batteries. Also, AI is used to automate inspection of solar farms using unmanned aerial vehicles called drones and algorithms to automatically detect solar panel defects from images, reducing the need for manual, time consuming and costly maintenance. The UN Intergovernmental Panel on Climate Change estimates that at least 500 billion metric tons of carbon dioxide will need to be pulled back out of the air this century to avoid the worst of global warming. AI accelerates discovery to develop solutions for this hard problem. From the chemistry of new materials for the next generation of batteries, to improved semiconductor chip design to optimize carbon capture and storage systems, AI is reducing the time it takes to invent and deploy new technologies. Even small changes made by a large number of people can have measurable impact. Take a moment and pause here. Can you think of any ways that you've seen AI influencing smart energy use in your daily life? Smart infrastructure, like Nest Learning Thermostats can collectively result in significant energy savings. Today, Google Maps offers users the option of selecting the most fuel efficient route even if it takes a bit longer. Google estimates that eco-friendly routing has the potential to prevent over 1 million tons of carbon emissions per year. That's the equivalent of removing over 200,000 cars from the road. This is a good example of introducing an additional goal to an AI objective function. AI can efficiently optimize for multiple goals. Finally, whenever you set up an AI system, consider its carbon emissions impact. In other words, make reduced energy consumption one of multiple goals for your AI system to optimize. Ideally, your AI system will optimize for your primary goal while also minimizing incremental carbon emissions at the lowest possible cost. This is important because training an advanced AI model takes a lot of energy, not only in the electricity consumed to train it, but also in the building of super computers and the collection and storage of data. To help developers see and reduce the emissions generated from training their AI models, several top AI experts created a tool called CodeCarbon which estimates the amount of carbon dioxide produced by the computing resources used to execute the code. And this is to incentivize developers to optimize their code efficiency. It also advises developers on selecting cloud infrastructure in regions that use lower carbon emitting energy sources. - [Solomon] What's the key ingredient of most AI systems? Like on a daily basis, it's electrons, those are the inputs. That's the fuel that makes the whole thing run. So if I'm a chef and I'm trying to make like really great food, one thing I might do is ask like, "Where did I get my ingredients?" Nowadays, there's a lot of options in how you obtain electricity. Like in California, for example, you can choose to actually source it from clean sources or not. If people were branded like, this AI was brought to you by coal, you know, it's like grass fed beef, like coal fed AI versus, you know, or is it all built on wind and geothermal or something? You know, so like, you know, you can always ask that question, "What am I building this based on?" - Getting to net zero emissions is an ambitious goal. For companies to prioritize these efforts, we need policies to be aligned with this goal. As Dr. Hsiang suggested, we should encourage companies to build sustainability into the objective functions they embed in their AI systems. Now that we're aware of the benefits and risks of AI in key application domains, we'll explore strategies for developing responsible AI systems in the next chapter.