Aligning artificial intelligence with climate change mitigation

Artificial intelligence and machine learning (ML) is a growing technology that has the potential to both mitigate and exacerbate climate change. It plays a role in data mining and remote sensing, accelerated experimentation, simulations and forecasting, system optimization and control, and predictive maintenance, among other applications. The authors present a framework for categorizing and understanding the varying impacts of ML on global greenhouse gas (GHG) emissions. Their framework traces (1) computing-related GHG emissions, (2) the immediate GHG emissions of ML applications, and (3) structural or system-level GHG effects of ML applications.

The authors explain that in order to measure the computing-related GHG impacts of an ML model one must examine the full life-cycle of the model itself. An ML model passes through the following life-cycle stages: model training, model development and tuning, and model inference. Model training is when the model “learns” the underlying function that governs how it works. Model development and tuning is when the researcher trains different variants of the model on varying datasets in order to refine the model. Lastly, model inference is when the model is actually deployed in the world. These three stages of an ML model’s life-cycle require varying degrees of energy consumption, and therefore GHG emissions. To accurately assess the GHG emissions associated with ML, one must also account for the embodied carbon emissions of the IT infrastructure that ML models rely upon. The embodied carbon includes emissions associated with material extraction, transportation, and the end-of-life phase of equipment.

The immediate GHG impacts of an ML model are defined as those impacts that are associated with the short-term outcomes of ML applications. Some examples of short-term outcomes of ML applications include data mining and remote sensing, tracking deforestation, evaluating coastal flooding risks, designing next generation batteries and materials, forecasting renewable power production, improving the operational efficiency of complex systems, and advancing climate simulations and modeling. While these are all positive uses of ML, negative examples exist as well.

ML applications can have system-level impacts that reach beyond their more immediate impacts. These impacts can have profound implications for the climate and global GHG emissions. The authors highlight several ways that system-level impacts manifest. One way is through rebound effects, whereby increased efficiency leads to greater consumption. ML models can also produce ‘lock-in’ effects, which enable inferior technologies to maintain their market dominance and stymie innovation. The last pathway cited is behavioral changes and changes to consumer patterns due to ML-enhanced advertising and marketing. These are all examples of system-level impacts.

The authors conclude their work by outlining several ways that the climate impacts and GHG emissions associated with ML could be better understood and explored. They highlight the need for well-developed metrics and reporting standards regarding the GHG emissions that result from ML model development, training and fine-tuning, and deployment/inference. Some key metrics might include model type and size, training requirements for model development and the type of pre-trained model used, the type and location of the computing infrastructure that’s used, and the frequency of training/retraining/fine-tuning and inference.

The researchers also call for benchmarking frameworks that help measure and evaluate training and inference efficiency. They acknowledge the need for more data collection, novel measurement methodologies, and new approaches for developing forecasts and scenarios. They argue for interoperability standards to prevent technological ‘lock-in’ and for best-practices that would indicate when certain ML models should be used over others for the sake of reducing GHG emissions. Many of these ideas must contend with the feasibility of data collection and reporting by actors with varying capacities.