Why AI isn't (yet) a true enabler on the road to sustainability
#AI22 is a series of articles highlighting what we believe to be 10 developments that will be impacting AI this year.
This series is co-written by Dr. Johannes Otterbach, Dr. Rasmus Rothe and Henry Schröder.
Environmental tech has quickly become perhaps the most sought-after equity in the VC world - with a record-breaking volume of venture capital invested last year. It is no surprise that at the intersection of environmental tech and AI extensive opportunities present themselves. PwC and Microsoft predict that through the use of AI, greenhouse gas emissions could be cut by 4% until 2030 - an amount equivalent to the predicted 2030 annual emissions of Australia, Canada and Japan combined.
The possibilities of application are significant: from improving power storage, by controlling when certain battery systems are charged to optimize lifetime, to analyzing enormous amounts of climate data and understanding the dynamics around sea-level rise or ice sheet shifts or tracking deforestation in rainforests, by analyzing satellite footage with computer vision models. The possibilities of AI will primarily advance in the measurement, reduction and substitution of emissions as well as in agricultural tech. A key use case in emissions measurement, is the analysis of extensive datasets to determine ESG scores. These systems will experience high traction - especially with the introduction of the EU taxonomy of sustainable activities, going into full effect by 2023. The use of AI to reduce emissions can range from less energy consumption through smarter street lights, to route optimization in daily traffic or by anticipating electricity supply and demand better and integrating renewable resources more efficiently while also reducing waste. Additionally, companies that harness the possibilities of AI for the development of more sustainable synthetic materials for now emission heavy materials, will advance significantly. In agricultural tech, AI will increase resource efficiency (e.g. less water use through weather data-based irrigation or optimization of pesticide use through drone data) or optimize the level of nutrition per output. This progress is further incentivized through the EU common agricultural policy (CAP) going into effect in 2023.
While AI models will act as a beneficial tool to assist industries in becoming more environmentally friendly, they will also have to become more sustainable. As the size and compute of models increase exponentially, so do their CO2 emissions. Google’s AlphaGo Zero generated 96 tonnes of CO2 whilst training, which amounts to c. 1000h of air travel. The industry's focus on computing efficiency and thereby also compute sustainability is not solely promoted due to a lower CO2 footprint, but also further implications. There have been strong efforts in the compression of larger models in order to embed applications more economically and speed up computing inferences. Moreover, algorithmic advances such as training a model in 2020 on Imagenet with 44x less compute compared to AlexNet in 2012, should not be ignored. Large tech companies (MAMAA) in collaboration with research institutes will further explore strategies such as reducing computational costs through sharing infrastructure or using AI to reduce cooling energy consumption.
Undoubtedly, the benefit of AI in the fight against climate change is evident, however, the degree to which AI will be helpful is ultimately dependent on its ability to self-optimize.