en.Wedoany.com Reported - A research team at the University of Washington has developed an artificial intelligence system capable of automatically estimating the environmental impact of different electronic devices during the production phase. The system utilizes multiple AI agents (programs that autonomously perform tasks) to extract information from publicly available data and conduct Life Cycle Assessments (LCA), achieving an average error rate between 5% and 19%, which is comparable to the accuracy of LCAs completed by human experts. The research findings were published on June 12 in the academic journal Nature Electronics.
Research indicates that consumers have a higher willingness to pay for more sustainable devices, but obtaining detailed environmental impact data for electronic devices is currently very difficult. Taking a mobile phone as an example, it consists of hundreds of chips and other components, each with varying production emissions, and this data is often not publicly available or even unmeasured. Manually collecting the information required for an LCA by human experts can take days or even months. The multiple AI agents designed by the research team can work collaboratively, automatically extracting electronic component information from product descriptions, images, and documents, providing comparable estimates in about one minute.
The system simulates different roles in the LCA process through two AI agents. One acts as an analyst, defining the required information and how to integrate it, and reviewing the accuracy of the results; the other acts as an engineer, gathering information on electronic device components from publicly available data, including filtering spreadsheets and searching for internal device images to obtain chip information. Data sources even include unconventional channels such as the FCC database and iFixit posts. The two agents work in a loop until the first agent confirms the information is complete, then references an LCA database to convert the component list into a carbon emission estimate.
The team also developed a "nearest neighbor" method to directly estimate carbon footprint without detailed data collection. For common devices like laptops and smartphones that already have publicly available carbon footprint reports, they found that products with similar specifications, such as screen size and processor, have similar carbon values. Therefore, the carbon footprint of an unknown device can be expressed as a weighted average of similar products. This method is also applicable for estimating materials not included in LCA databases, such as new sustainable plastics, which can be estimated based on plastics with similar properties and chemical structures. In tests, this method had an average error of 23%, compared to 143% for human experts.
The research team emphasized that the system aims to help reduce overall carbon emissions, but running the AI model itself requires energy. To mitigate this impact, they have taken several measures, including using smaller AI models that consume less energy than general-purpose models, and avoiding redundant calculations by first querying existing emission estimates. If the system does need to call an AI model, the carbon emission from estimating the carbon footprint of one device is currently equivalent to the emissions from brewing a cup of tea. The team plans to collaborate with companies in the future to help automate their workflows.
Vikram Iyer, senior author of the study and assistant professor at the Paul G. Allen School of Computer Science & Engineering at the University of Washington, stated that people are willing to pay a higher price for more sustainable devices, but products like mobile phones consist of hundreds of chips and other components, and the production emission data for each component is not publicly available and difficult to obtain. Manually collecting this data by human experts can take days or even months, whereas the multiple AI agents they designed can work collaboratively and complete the estimate within one minute. Zhihan Zhang, first author and a doctoral student at the Allen School, explained that they interviewed LCA experts and built the system to simulate the interaction process. Other co-authors include: Alexander Metzger, Felix Hähnlein, Zachary Englhardt, and Shwetak Patel from the Allen School at the University of Washington; Yuxuan Mei from Wellesley College (who was a doctoral student at the Allen School at the University of Washington when completing this research); Tingyu Cheng from the University of Notre Dame; Gregory D. Abowd from Northeastern University; and Adriana Schulz from Brown University (who was an assistant professor at the Allen School at the University of Washington when completing this research). This study was funded by the Amazon Research Awards and the National Science Foundation, and Zhang was supported by the Google PhD Fellowship.
This article is compiled by Wedoany. All AI citations must indicate the source as "Wedoany". If there is any infringement or other issues, please notify us promptly, and we will modify or delete it accordingly. Email: news@wedoany.com
