Climate change and machine learning — the good, bad, and unknown

Machine learning and climate change have a complicated relationship: Machine learning can enable climate-friendly actions, but it can also hurt sustainability goals, given its large demand on energy resources and its role in climate-adverse business models and trends.

Organizations need to continuously push the boundaries of diverse machine learning technologies to meet climate change challenges while considering their energy costs, according to MIT professor Priya Donti. Speaking at the 2024 MIT Sustainability Conference, Donti said that organizations must also remain pragmatic about how machine learning can upset efforts or create uncertainty around meeting broader societal sustainability goals. 

“There are lots of subtle but transformative effects machine learning has that we should be paying attention to in the context of climate,” said Donti, a co-founder and chair of Climate Change AI, a global nonprofit focused on the intersection of climate change and machine learning. 

The upside: machine learning advances sustainability goals

Machine learning can help with climate solutions across several fronts, including improving the efficiency of electricity systems and smart buildings and accelerating climate science research. Donti and her coauthors highlighted these innovations in a 2022 paper that details how machine learning applications can be applied to climate change in several broad categories:  

Turning raw data into actionable insights. It’s not always possible to gather on-the-ground data at the scale necessary to understand greenhouse gas emissions — for example, when collecting information from areas where deforestation is happening or capturing the energy efficiency characteristics of smart buildings throughout a city. Combining satellite and aerial imagery with machine learning can provide insights that can be extrapolated to a broader scale. For example, a coalition of nonprofits called Climate Trace uses a combination of satellite imagery and on-the-ground data to gather independent emissions inventories for different sectors. 

Improved forecasting. Machine learning systems can process data to uncover relationships among variables for improved forecasting. For example, historical data about weather and solar power production can be used to forecast what solar power production might look like in the near future. This could facilitate better power grid management.

Automated decision-making. Machine learning programs can analyze real-time information to automatically calibrate the temperature of buildings, data centers, or refrigeration environments efficiently.

Predictive maintenance. Asset downtime is detrimental to business operations, so the ability to identify and address potential problems before they occur is a huge efficiency advantage. For example, German rail company Deutsche Bahn is using machine learning to identify faults in the railway switching infrastructure, which enables it to make proactive fixes that keep the trains running on time, Donti said. Another example is natural gas providers that are using sensor data and aerial and satellite imagery to detect anomalies that predict methane leaks before they occur. 

Science and engineering discovery. Machine learning can analyze the outcomes of experiments to accelerate discovery in areas like synthesizing molecules or carbon dioxide sorbents. Machine learning can also help approximate time-intensive simulations, which enables faster overall model performance and higher-resolution outputs. 

Data management for climate change workflows. Data is key to the modeling process. Machine learning can help organizations match and merge datasets from diverse sources. 

Given the heterogeneity of climate change challenges, these diverse approaches are needed. “We need to make sure we’re fostering a diverse ecosystem that can meet this set of challenges, rather than conflating one particular set of AI techniques with one particular AI paradigm,” Donti said. 

The bad: machine learning uses a lot of energy 

Machine learning has a significant impact on energy and water resources, given the computational horsepower required for processing and training large models and the water needed to cool data centers, Donti said. Hardware use consumes energy, and producing, transporting, and disposing of hardware also creates carbon emissions. Research shows that data centers and information and communications technology accounted for 1% to 2% of greenhouse gas emissions in 2020. Though it isn’t known how much AI contributes to those emissions, energy use related to developing, training, and running machine learning models is undoubtedly rising, Donti said.

Moreover, research shows that newer technologies are inherently more carbon-intensive. For example, the carbon footprint for model training and model execution has historically been split 50/50, Donti said. Using large language models has been found to be more carbon-intensive than training them. Similarly, task-specific models have given way to multipurpose models that are larger and more compute-intensive. Choosing the wrong approach or model can significantly increase energy use. 

Creating a greener power grid with renewables and energy workload management is essential but still insufficient to fully address these issues. “Understanding these dynamics and trends is key to understanding what to do about it,” Donti said. “This is where increased transparency and data collection becomes exceedingly important.”

The unknown: whether machine learning facilitates climate-adverse technologies

Widespread machine learning use also has nuanced climate implications. For instance, machine learning is being used to accelerate production levels for the oil and gas industry, a hugely carbon-intensive sector, Donti said. Similarly, machine learning paired with technologies such as internet-of-things devices can help farmers manage larger groups of livestock. While such innovations can increase profits and productivity for their users, their potential for boosting carbon emissions is significant. 

Autonomous vehicles, another promising innovation that relies on machine learning, could also have a negative impact on climate action goals, Donti said. While individual self-driving vehicles are more energy efficient than most vehicles on the road today, they may entice people to drive more, entrenching privatized transportation and slowing a transition to a more multimodal transportation model. 

Another area where machine learning innovations are in potential conflict with climate action is personalized advertising that encourages emissions-intense consumption and amplifies polarized views.

Donti made the following recommendations for countering those effects:

• Invest in heterogeneous implementations of AI and machine learning so the organization isn’t limited to one approach or constrained by a single vendor’s strategies.
• Commit to purposeful work on applications that have proved to do good.
• Adopt practices that reduce Scope 1, 2, and 3 emissions.
• Communicate the benefits and risks of AI in an appropriately nuanced and grounded way as opposed to engaging in hype and overpromising results.

“As we move AI forward, we must actively account for both AI’s and machine learning’s direct impact and the implications for different applications,” Donti said. “We must use it for good applications and for business as usual.”