There are two issues to consider in the relationship between the environment and AI. The first is how AI can help in conserving the environment and tackling climate change. The second is what is the environmental impact of AI.
Let’s go with the positive part.
Weather forecasting
Data, meteorology has an enormous amount of data. Analysing that data and comparing the results with previous models to see if they are similar or if they deviate is very much the basis of meteorology. It is actually more complex than what I have just said, but it is broadly consistent and allows us to understand the importance of AI in this field. If there is one tool capable of handling data at enormous speed and with unprecedented accuracy, it is AI.
Currently, GraphCast, GoogleDeepMind’s AI model, is capable of giving accurate predictions for the next ten days in one minute, as well as extreme weather phenomena that have become more virulent and frequent due to climate change.
In a comprehensive comparison with the standard weather simulation model used in the industry, the High Resolution Forecasting System (HRES), GraphCast managed to outperform HRES in accuracy in more than 90% of the 1,380 variables analysed, including the prediction of run times. Furthermore, in the specific assessment of the troposphere, which ranges from 6 to 20 kilometres in altitude and is crucial for accurate forecasts, the model developed by Google DeepMind proved superior in 99.7% of the variables examined for future weather forecasts.
The traditional method relies on numerical weather prediction (NWP) models, which use physical and mathematical equations to simulate the atmosphere and other aspects of the climate. These equations describe how weather variables such as temperature, pressure, humidity and wind evolve, and are solved using supercomputers due to their complexity and the volume of data involved.
On the other hand, artificial intelligence, especially deep learning, offers a complementary approach. Instead of relying solely on physical equations, AI models are trained on large amounts of historical data to identify patterns and make predictions about the future state of the weather. It is also an open source system, so scientists and institutions have access to it and can use it.
Both methods have their strengths and are often used in a complementary way to improve the overall accuracy of weather forecasts. The combination of traditional modelling and AI techniques is opening up new possibilities in the field of meteorology.
Let us list, without having to think too hard, who can benefit from this: agriculture, livestock farming, logistics or tourism. It is even more obvious that lives can be saved, especially in developing countries where weather can be even more damaging.
Energy efficiency
Today’s electricity grid is not an example of a distributive grid and self-consumption, but it has long since ceased to be a centralised grid at the point of energy creation and distribution. Solar panels, wind farms, pellet boilers and many other systems have reconfigured the grid and maintaining a stable supply is no longer so easy. The increase of extreme metrological phenomena also affects it and many of the renewable energy generation systems as well.
Lots of real time data, a long history of models and new situations and the need to make quick and effective decisions. By now we know that this is one of the typical problems where AI has a lot to contribute. And so it is.
The increase in electric cars leads to increases in consumption due to recharging. Knowing the new habits of the users of these vehicles is important to keep the grid operational. Winds that cause damage to electricity grids and that until now could have caught technicians unawares, are now predicted more in advance. In the same way, the hours and intensity of sunlight, as well as the wind, make it possible to calculate more accurately the energy to be generated.
This same approach not only improves energy generation and consumption, but also prevents failures and extends the life of installations and power plants by avoiding peaks and troughs.
Networks of sensors and alarms have been running through power grids for some time now. The leap that AI allows is to create patterns and understand in less time and with more precision the enormous amount of data that such a complex network generates on a daily basis.
And if this were not enough
Mapping rubbish in the ocean and establishing models to calculate its displacement, creating deforestation maps and proposals to tackle it, predicting forest fires, optimising recycling processes and avoiding food waste. In short: any complex system that requires complex calculations benefits from this tool. In a context of climate change in which uncertainty pervades everything, the use of AI to address the environmental challenge is a huge advantage.
Let’s get to the badside.
The energy consumption of AI
To give you an idea, training ChatGPT generates about 300 tonnes of carbon dioxide emissions, which is the equivalent of a European citizen in 60 years. A single ChatGPT query is equivalent to five or six queries to a normal search engine in terms of environmental impact.
There are geographical variables that are not technology-specific but country-specific. If the server to be trained is in the United States, we will need 700,000 litres of clean, fresh water to train it, but if it is an Asian server, we will need three times as much. Chinese data centres are 73% coal-fired, while Canadian data centres are mostly hydroelectric.
These figures are from 2019 and 2021 so we can assume that they are much higher today given the growth of AI.
There are several points where the energy consumption of AI becomes important.
It demands a large amount of specialised hardware (GPUs and ASICs) which also adds raw materials and waste in its production.
Data storage. Business profitability and sustainability go hand in hand here. Improving storage through data compression and the use of low-power data transmission technologies can lead to significant savings.
Data processing centres have a very high level of energy consumption. The use of renewable energy and even looking for geographic locations with favourable environmental conditions has long since become part of the strategy for these huge technology facilities.
But if there is anything better than offsetting the impact, it is avoiding it. Increasing the efficiency of algorithms and software in general can lead to considerable savings. This is especially urgent during AI training, as this is the time of highest consumption and the time when the least impact mitigation measures are in place.
It’s a high price to pay. Should we put AI to work to find solutions?
The photo accompanying this article was created with ChatGPT4. At the end I asked him:
“Before I go, could you tell me the energy and water consumption of this conversation?”
And his response was:
“I don’t have access to specific data on the energy or water consumption used during our conversations. Energy and resource consumption depends on many factors, including the servers and network infrastructure used to run and access services like this.
In general, running artificial intelligence models on servers can consume a significant amount of electrical power, but quantifying this for a single session or question would be complex and dependent on the hardware and energy efficiency of the specific infrastructure in use. OpenAI, the organisation behind ChatGPT, is continuously working to improve the energy efficiency of its operations and mitigate environmental impact through various sustainability initiatives.
As for water consumption, this is generally indirect, mainly associated with cooling and other data centre operations, but these specific details are also handled by the infrastructure providers and are not directly available through this service.
If you have further questions or there is anything else I can help you with, let me know!”
