Abstract
As machine learning models grow increasingly powerful driven by big-data and the scaling law, the environmental sustainability of artificial intelligence research becomes an alarming issue. My Capstone Project leverages deep learning techniques to improve rainfall-runoff modeling in hydrology, addressing the challenge of accurately predicting floods and managing water resources. Additionally, my research generalizes a data-driven approach to tackle large-scale scientific challenges where neural networks outperform traditional methods. My STS research investigates how governmental funding decisions in AI research have profound environmental impacts. Despite the enormous energy consumption of computationally intensive models, it is often overlooked in the grant approval process. My Capstone Project provides a basis for quantifying the energy usage of machine learning research, as well as exploring how government funding directly impacts environmental research projects like mine. This study emphasizes the importance of integrating sustainability into AI research at a high level, ensuring that technological advancements are pursued responsibly and ethically.
My Capstone Project, Deep Learning in Hydrology, provides a computational solution to rainfall-runoff modeling by employing the Long Short-Term Memory (LSTM) neural network. Traditional hydrological models, though effective, are computationally expensive and limited by data availability. In this work, we analyzed hydrology time series using the CAMELS and Caravan global datasets. These datasets include up to 6 time series variables and 209 environmental features collected from around 8,000 locations worldwide.
We found that including environmental data in training significantly boosts model accuracy, reducing the error by 40% when tested on the largest dataset. Additionally, including encoding techniques that captures the relationship between catchments and some periodic hydrological patterns further improves model performance. When compared to state-of-the-art time series foundation models that require huge computational resources to train, our domain-specific LSTM model outperforms all of them in a benchmark experiment. These results advocate for the use of domain-specific knowledge in large-scale scientific time series research to improve efficiency and reduce environmental footprint.
My STS research paper examines how government funding impacts AI development and environmental sustainability. My research question explores the effects government funding has on academic research topics and the extent to which agencies like National Science Foundation (NSF) and the Department of Energy (DOE) incorporate environmental considerations into AI project evaluations. Utilizing a systematic literature review and ethical frameworks, this study highlights the critical oversight of energy efficiency and environmental impact within governmental AI funding strategies.
The analysis reveals a significant gap: governmental funding processes prioritize immediate economic and societal benefits, often neglecting long-term environmental consequences. Case studies of large-scale AI projects illustrate substantial environmental costs due to high GPU usage and energy consumption. The conclusion calls for integrating sustainability metrics into government funding criteria, advocating a balanced approach guided by both utilitarian and environmental ethics to incentivize energy-efficient AI innovations.
