Climate change is one of the most pressing issues of our time. Despite increasing global consensus about the urgency of reducing emissions since the 1980s, they continue to rise relentlessly. We look to technology to deliver us from climate change, preferably without sacrificing economic growth.
Our optimistic—some would say techno-utopian—visions of the future involve vast arrays of solar panels, machines that suck carbon dioxide back out of the atmosphere, and replacing fossil fuels for transport and heating with electricity generated by renewable means. This is nothing less than rebuilding our civilisation on stable, sustainable foundations.
Meanwhile, society is increasingly being shaped by machine learning algorithms: automating occupations, performing tasks from diagnosing illnesses to serving up adverts, and nudging people into different behaviours. So how can AI help in the fight against climate change?
“In many ways” is the answer. Just as tackling climate change involves practically every sector—agriculture, transport, architecture, energy, industry, logistics to name but a few—so machine learning solutions can find their niche to help solve some of the thousands of problems that arise. This can range from improving our understanding of the problem by making better climate models, helping businesses and industries reduce their emissions, aiding in the design of new technologies, or helping society adapt to the changes that are already on the way.
Now, a team of researchers from multiple institutions—including Coursera founder Andrew Ng, Chief Scientist of Google John Platt, and Turing Prize winner Yoshua Bengio—have published a 100-page research paper outlining some of the areas where machine learning is best-placed to make a difference.
A classic example is in the field of renewable energy. Solar and wind are now, in most regions, the cheapest electricity generation to build, even without a price on carbon. The main barrier is intermittency: how to integrate these power sources, which vary with the weather and seasonally, onto a grid driven by human demands. Doing this efficiently allows us to minimise the amount of fossil fuels we burn, but it requires skill in forecasting both supply and demand.
Machine learning algorithms can process huge amounts of data, from real-time weather conditions to information about pollution to video streams from areas near solar panels, and can rapidly convert these into predictions for the amount of power that will be generated. Beyond just forecasts, though, machine learning algorithms can be in charge of “scheduling and dispatch”—determining which power plants should operate at any given time, and which can be switched off.
In the future, Internet of Things technologies may provide more flexibility for demand-side management: the most power-intensive processes can take place when supply peaks, avoiding wasted energy and overproduction. Electrification of transportation will also add local storage options to this more complex grid: the large batteries of electric cars could be used to power your home, and the first models that can do this are forthcoming.
Controlling such a network of supply, storage, and demand in the presence of uncertainty and streams of data from millions of different sources is a job for machine learning. Algorithms such as those that serve up ads already use mathematical infrastructure like bandit theory to decide which action is likely to maximise a given reward; they could be well adapted to control this new, greener grid if that reward is minimising emissions, or maximising profit for the electricity company.
Another network that might benefit from machine learning control is transportation. Cutting down on unnecessary journeys or alleviating traffic can help to reduce pollution. Uber’s algorithms already excel at matching riders to drivers, and ride sharing is another alternative means of reducing emissions from transport. As autonomous vehicles become increasingly prevalent, machine learning algorithms can optimise with emissions in mind and help cut down on the sector that accounts for a quarter of carbon dioxide emissions.
On the research and development side, machine learning is increasingly combined with physics-based models and experimental data to predict how new materials will behave. This can help us to find materials for flexible, super-efficient solar panels or LEDs by predicting which crystal structures will have the best photovoltaic properties; it can be used to design thermoelectric materials that can turn waste heat back into useful electricity; and it can be used to help find absorbent materials for those negative emissions CO2-scrubbers. One could even imagine, someday, the entire process of choosing, designing, fabricating, and testing a new crystal could be entirely automated and subject to machine learning control.
The Paris Agreement is much-vaunted as the main international agreement to reduce emissions. However, it is based on voluntary targets and self-reporting of emissions. Not only are there as many ways of carbon accounting as there are accountants, but there is also the potential for fraud and deception: after all, Volkswagen systematically cheated on emissions tests for years. More trust might arise if emissions could be monitored remotely.
Satellite data, including a new fleet of CO2-monitoring satellites due to be launched by the EU in the 2020s, could allow for independent measurements of CO2 to take place, helping nations take stock of their individual and collective efforts and identify key areas to work on. Churning through satellite data, particularly where it requires feature recognition, is a job that machine learning algorithms already excel at. The drive for natural gas production through fracking and other techniques has led to leaks from methane pipes, driving up concentrations of a potent greenhouse gas. But these can also be spotted with satellites.
This is not all satellite data can be used for. A large part of our uncertainty in how the climate has responded to human influence is due to clouds, which can be influenced by pollution in complex ways. ML algorithms that scan through satellite cloud data, correlating it with sources of pollution on the ground, can help us narrow down this uncertainty and hence better constrain forecasts of global temperature.
Neural networks are very good at encoding subtle, statistical relationships between multiple variables. This means they can potentially be used to represent physical processes in a more computationally efficient way, allowing us to improve climate and weather models, potentially allowing us to integrate more real-world data and better representations of processes that take place on small scales into these models. This is crucial, as we rely on climate models to understand which impacts are likely to affect which regions in the future, and even to determine whether geoengineering schemes might do more harm than good. Improving these models means better decision-making.
Meanwhile, those most vulnerable to climate change live in the poorest nations, where governments are least able to adapt and extreme heatwaves, droughts, or floods are deadly. Machine learning can be used to map informal settlements from satellite data: the first step in disaster response is knowing where people actually live. When crisis hits, machine learning algorithms can trawl through aerial photography, satellite data, and even social media posts in real time, providing information to rescuers about where help is most needed. Automated monitoring of social media combined with natural language processing can tell rescuers where supplies of water and food are low, even when conventional means of communication are unreliable.
There are aims to use machine learning to help in the social side of climate change as well. Tools that allow you to optimise your own energy use, or keep track of your carbon footprint, can be improved by machine learning algorithms. Yoshua Bengio’s project aims to galvanise people into action by visualising possible future impacts of climate change with neural networks that generate imagery of flooded homes.
Machine learning can even be used to try to reduce the carbon footprint of… machine learning. The energy consumption from GPUs can be huge, particularly when you’re running them to do work that is useless or redundant by design. Training advanced neural networks comes with a carbon footprint of its own. But, of course, saving energy saves money as well as benefiting the environment: this is why Google seeks to use machine learning to reduce the energy footprint of its datacenters by changing operation strategy and cooling techniques.
In short, the possibilities for machine learning to help with climate change are all around us. The machine learning revolution is based on the idea that the more data we collect and process, the more statistical relationships we understand, the better decisions we can make. Climate science is heavily driven by climate data: adaptation will require policies that are tailored to the individual changes expected in each region; mitigation will require improvements in efficiency and changes in energy use in virtually every sector of society. The time is ripe to deploy some of our most advanced and exciting computational tools to help solve the outstanding challenge of our age.
Originally posted here.