Can Power Hungry AI Help Fight Climate Change?

AI can do remarkable things. Just recently researchers at OpenAI in San Francisco reported that they had worked out an algorithm that was capable via trial and error of manipulating the pieces of a Rubik’s cube using a robotic hand. However, it took 1,000 desktop computers plus a dozen machines operating specialised graphics chips crunching intensive calculations for several months.

Evan Sparks, CEO of Determined AI, a start-up that provides software to help companies manage AI projects, estimated that this process may have consumed about 2.8 gigawatt-hours of electricity which is roughly equal to the output of three nuclear power plants for an hour.

The extraordinary advances produced by artificial intelligence such as computers learning to recognise images, converse, beat humans at sophisticated games, and drive vehicles need huge amounts of computing power and electricity to devise and train algorithms. AI experts are increasingly realising that these energy demands are too great in the face of climate change.

Sasha Luccioni, a postdoctoral researcher at Mila, an AI research institute in Canada said:

“The concern is that machine-learning algorithms in general are consuming more and more energy, using more data, training for longer and longer.”

It’s not only a worry for academics but for companies too across many industries that are beginning to use AI. There is growing concern that the technology will worsen the climate crisis. Evan Sparks says that his company is already working with a pharmaceutical company which is using enormous AI models.

He added:

“As an industry, it’s worth thinking about how we want to combat this.”

Some AI researchers are looking at ways to offset their emissions and using tools to track the energy demands of their algorithms. There is a growing appetite for algorithms that burn fewer kilowatts.

A sophisticated code is being developed by Sasha Luccioni that can be added to an AI program to track the energy use of individual computer chips.

Sasha Luccioni and others are keen that companies not only track the performance of code but also include some measure of energy or carbon footprint.

She says:

“Hopefully, this will go toward full transparency. So that people will include in the footnotes ‘we emitted X tons of carbon, which we offset.’”

A team at UMass Amherst produced a research paper that revealed how training a single large NLP (neuro-linguistic programming) model may consume as much energy as a car over its entire lifetime, including the energy needed to build it. Recent advances in natural language processing, which is an AI technique that helps machines parse, interpret, and generate text have proven particularly power-hungry.

Huge banks of computers often have to run for days and sometimes weeks in order to train a powerful machine-learning algorithm.

It has been estimated by the department of energy that data centres account for about 2 percent of total US electricity usage and that worldwide they consume about 200 terawatt hours of power per year which amounts to more than what some countries consume.  Significant growth in computing and communications technology is predicted in the years leading up to 2030. These technologies could consume between 8% and 20% of the world’s electricity with data centres accounting for a third of that.

In the last few years companies offering cloud computing services have been looking at ways to offset their carbon emissions with different degrees of success. Google, for example, claims ‘zero net carbon emissions’ for its data centres due to its extensive renewable energy purchases. Microsoft have declared that they have a plan to become ‘carbon negative’ by 2030 by offsetting all of the carbon produced by the company over its history. OpenAI signed a deal to use Microsoft’s cloud last year.

Many of the AI industry’s top thinkers believe that AI itself can help fight climate change. Success in the development of AI in recent years has led many to question how this technology could help with one of the greatest dangers facing humanity, climate change. A research paper originating from some of the field’s best-known thinkers last year gives a number of examples of how machine learning could help prevent human destruction.

The authors of the paper which include DeepMind CEO Demis Hassabis, Turing award winner Yoshua Bengio, and Google Brain co-founder Andrew Ng say that AI could be “invaluable” in diminishing and avoiding the worse effects of climate change though they do not think it is a “silver bullet” but believe political action is seriously needed too.

The paper suggests a number of fields where machine learning could be put to good use.

Despite the fact that electricity systems are “awash with data” very little is being done to take advantage of this information. Machine learning could help by forecasting electricity generation and demand, allowing suppliers to better integrate renewable resources into national grids and reduce waste. Google’s UK lab DeepMind has demonstrated this sort of work already, using AI to predict the energy output of wind farms.

It’s important to remember that greenhouse gases aren’t just emitted by engines and power plants but come from the destruction of trees, peatland, and other plant life. Plant life has captured carbon through the process of photosynthesis over millions of years with deforestation and unsustainable agriculture leading to this carbon being released back into the atmosphere. AI and satellite imagery can be used to monitor agricultural emissions and deforestation by pinpointing where this is happening in order to protect these natural carbon sinks.

AI can be used to help scientists find new materials by giving them the potential to model the properties and interactions of never-before-seen chemical compounds. At the moment 9% of all global emissions of greenhouse gases comes from the production of concrete and steel.  Machine learning could help reduce this number by creating low carbon alternatives to these materials.

As we see greater effects from climate change in the coming decades, driven by highly complex systems such as changes in cloud cover and ice sheet dynamics, AI could prove to be invaluable. Modelling these changes will undoubtedly help scientists predict extreme weather events, like droughts and hurricanes, which in turn will help governments protect against their worst effects.

The transportation sector currently accounts for a quarter of global energy-related CO2 emissions, with two-thirds of this generated by road users. In the same way that machine learning can help with electricity systems it could make this sector more efficient, by reducing the number of wasted journeys, increasing vehicle efficiency, and shifting freight to low-carbon options like rail.

Though energy consumed in buildings accounts for another quarter of global energy-related CO2 emissions, it presents some of “the lowest-hanging fruit” for climate action. Reducing wasted energy from buildings can be easily tackled by adding just a few smart sensors to monitor air temperature, water temperature, and energy use. Buildings are long-lasting and are very unlikely to have been retrofitted with new technology. Energy usage could be reduced by 20% in a single building and large-scale projects could monitor whole cities to have an even greater impact.

Some scientists are very hopeful that ways could be found to make clouds more reflective or to create artificial clouds using aerosols so that we could reflect more of the Sun’s heat back into space. This example of AI’s use is more extreme and speculative than others and modelling the potential side-effects of any schemes is hugely important. Geoengineering a more reflective earth presents significant “governance challenges” ahead.

According to the authors of the research paper, it’s a “common misconception that individuals cannot take meaningful action on climate change.” It is true however that people do need to know how they can help.  Machine learning could be used to calculate an individual’s carbon footprint and indicate what small changes they could make to reduce it. This could be a very useful too to reduce your carbon footprint. An individual might learn that they need to use public transport more, buy meat less often; or reduce their electricity use in their house. Adding up individual actions can create a big cumulative effect.

Getting to the point when AI can handle all the things, we would like it to do, to help with the climate crisis will require some pretty substantial upgrades to the computing infrastructure that underlies it.

Though many people are not aware of this, AI is a technology that is transforming every walk of life. It is a wide-ranging tool that enables people to rethink how we integrate information, analyse data, and use the resulting insights to improve decision making.

Post by Renewable Energy Hub