Author ORCID Identifier:
Date of Graduation
5-2026
Document Type
Thesis
Degree Name
Master of Science in Mechanical Engineering (MSME)
Degree Level
Graduate
Department
Mechanical Engineering
Advisor/Mentor
Huitink, David
Committee Member
Liao, Haitao
Second Committee Member
Millett, Paul
Keywords
Data center cooling; Direct liquid cooling; Mean time to failure; Physics of failure; Processor reliability; Thermal management
Abstract
Data center cooling strategy selection is typically evaluated on thermal and energy metrics, without a comparable metric for processor operational lifetime. This thesis develops the Thermal Stack Reliability Module (TSRM), a physics-based framework that maps workload-driven thermal behavior to component-level reliability outcomes for server-class processor packages. TSRM couples a compact thermal stack model, evaluated by the open-source ParaPower solver, with rainflow cycle counting, five physics-of-failure acceleration models spanning eleven failure mechanisms, coverage-weighted damage accumulation, and series reliability combination, returning a mean time to failure (MTTF) and a dominant failure mechanism for each cooling configuration. The framework is exercised across seven cooling strategies spanning air-cooled and direct-liquid-cooled operation. A 200-point CT ×Rth cooling grid over coolant temperature and cold plate thermal resistance is evaluated at four simulation horizons (one day, one week, one month, and one year), together with one-at-a-time perturbations of six load-schedule and two reliability parameters. Outputs include a continuous MTTF performance envelope with the seven strategies embedded, a five- to ten-year desired lifetime region anchored to the typical server refresh horizon, a tornado sensitivity ranking, and a self-consistency check of short-horizon extrapolation against the native one-year result. Predicted lifetimes span 28.9 years for chilled-water direct liquid cooling to 1.8 years for elevated-temperature air cooling, with a consistent strategy ranking across horizons within this load envelope. The peak utilization parameter dominates the tornado at every horizon, and the Weibull shape parameter is the only reliability input competitive with the top load parameters. Inside the desired lifetime region, a 0.01 °C/W reduction in thermal resistance buys approximately the same lifetime as a three- to five-degree reduction in coolant inlet temperature. The one-month horizon is the shortest window at which framework outputs stabilize, within roughly 20 percent of a native one-year simulation. TSRM enables reliability-based comparison of cooling strategies and supports design-stage selection for data center deployments.
Citation
Stubbs, N. I. (2026). Development of an Integrated Decision-Support Tool for Thermal-Reliability Prediction of Processor Lifetimes in Data Centers. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/6294
