ERI: A Foundational Framework for Sustainable Generative AI Agents

This Engineering Research Initiation (ERI) project aims to establish a foundational framework for modeling and reducing the environmental impact of generative Artificial Intelligence (AI) systems. By mitigating the massive energy demands of advanced computing, this project serves the national interest by ensuring energy infrastructure resilience and sustaining United States leadership in AI innovation. The project will bring transformative changes by enabling engineers to predict and minimize AI carbon emissions before deployment, pivoting from reactive energy measurement. This will be achieved by developing novel predictive models to evaluate energy use at the fundamental hardware level and during the continuous self-learning phases of autonomous generative AI agents. The intellectual merit of the project includes transitioning the field from coarse-grained, hardware-dependent measurements to a predictive, architecture-aware modeling approach capable of estimating emissions even for physically inaccessible cloud computing environments. The broader impacts of the project include establishing a community-engaged undergraduate fellowship program focused on rural communities and Primarily Undergraduate Institutions, ensuring all outreach is open to all Americans. Additionally, the project empowers the technology industry with open-source sustainability tools and trains future engineers through a newly developed course module. This research project addresses the challenge of estimating and optimizing carbon footprints on partly inaccessible Graphics Processing Units (GPUs) and continuously adapting AI agents. The first technical challenge is the black-box nature of whole-chip energy estimation. To resolve this, this project introduces a "Carbon-per-Tile" methodology. This approach maps complex Deep Neural Network (DNN) operations to fundamental GPU execution units, creating predictive carbon models that explicitly account for device-specific complexities, varying software implementations, and specialized hardware like Tensor Cores. The second technical challenge is the overlooked carbon emissions of the continuous self-learning phase in autonomous AI systems, where real-time tracking can introduce severe computational overhead. The project addresses this by developing a low-overhead "Carbon-per-Iteration" metric, applying amortized, dynamic Life Cycle Assessment (LCA) principles to the iterative reinforcement learning loops where AI agents update their parameters. By combining these methods, the project will generate an open-source optimization platform and a prototype generative AI agent capable of efficiently estimating its lifecycle impact. This contribution provides data center operators and developers with actionable guidance to optimize AI workloads, effectively bridging the gap between algorithmic design and sustainable AI infrastructure. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria. NSF Award ID: 2553490 | Program: 01002627DB NSF RESEARCH & RELATED ACTIVIT | Principal Investigator: Wenkai Guan | Institution: University of Minnesota Morris, MORRIS, MN | Award Amount: $200,000 View on NSF Award Search: https://www.nsf.gov/awardsearch/show-award/?AWD_ID=2553490 View on Research.gov: https://www.research.gov/awardapi-service/v1/awards/2553490.html