Data-Driven Cooling Architecture Decisions for Data Centers: A Geolocation-Informed Machine Learning Framework
Name
gao-bobgao-sm-sdm-2026-thesis.pdf
Size
6.34 MB
Format
Adobe PDF
Checksum (MD5)
d9aa68ec0ab6c3fd1a5d6ad2cf855825
Author(s)
Gao, Bosheng
Advisor(s)
Rhodes, Donna
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
The growth of data-intensive applications, including Artificial Intelligence and cloud computing, is driving surging demand for data center capacity, leading to the equipment generating more heat. Data center cooling is a critical component that manages heat and impacts the data center’s efficiency and performance in terms of energy, water, costs, and sustainability. As resource constraints become tighter, technologies rapidly evolve, and the climate constantly changes, a quick, transparent, cost-effective approach to cooling architecture selection is needed to support early-stage planning and decision-making.
This thesis proposes a geolocation-informed, data-driven framework for selecting cooling architecture. The framework learns from thousands of thoroughly analyzed real-world cooling architecture designs implemented at various data centers and trains a supervised machine learning model on public records of those data centers’ geolocation, climate, and available natural resources at each site. When the geographic coordinates are provided, the framework can retrieve the location’s attributes, predict cooling architectures, and provide the likelihood of fitness for each architectural option.
A Gradient Boosting Classifier is employed for the machine learning model. During model training and validation, an overall AUC of approximately 0.8 is achieved, demonstrating meaningful predictive performance. SHAP is used to increase the model’s explainability. This thesis uses a systematic approach to reveal and associate geolocation with data center cooling architecture decisions.
MIT Department
System Design and Management Program.
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
