← Back to Projects

COMSOL • THERMAL FEA • AI DATA CENTERS

FEA Modeling for Extreme Cooling on NVIDIA Blackwell GB200

Thermal-fluid modeling of a single-phase immersion cooling system for high-density AI computing, combining heat transfer, fluid flow, and data-center-scale cooling architecture.

Institution San Diego State University
Course ME 610
Project Type Thermal FEA
Software COMSOL Multiphysics
AI data center immersion cooling concept

Cooling extreme heat loads in next-generation AI hardware.

High-density AI hardware creates intense thermal loads that traditional air cooling struggles to manage.

This project investigated a single-phase immersion cooling architecture for NVIDIA Blackwell GB200 server systems using coupled thermal and fluid-flow simulation in COMSOL Multiphysics.

AI server specifications and cooling requirements

Server specifications used to establish thermal and cooling requirements

Traditional cooling does not scale well to extreme AI power density.

Modern AI server clusters generate exceptionally high heat flux in compact spaces, creating hot spots, efficiency losses, and reliability concerns.

The cooling concept therefore focused on direct immersion of server hardware in a dielectric fluid, allowing heat to be removed through forced liquid convection rather than relying primarily on air cooling.

AI server cooling engineering calculations

Engineering calculations used to size and evaluate the cooling system

Coupled heat transfer and nonisothermal flow.

The COMSOL model used two-way thermal-fluid coupling, allowing temperature to influence fluid behavior while moving coolant transported heat away from the servers.

01

Heat Source

Server heat generation was represented as volumetric thermal energy input.

02

Heat Transfer

Heat transfer in solids and fluid regions modeled thermal transport through the system.

03

Laminar Flow

Coolant motion was modeled to capture convection and fluid circulation.

04

Nonisothermal Flow

Temperature and velocity fields were coupled to capture the interaction between flow and heat transport.

05

Coolant

FC-40 dielectric liquid was selected for the immersion-cooling model.

06

Boundary Conditions

Inlet temperature and flow conditions were imposed to study cooling performance.

Large-scale immersion tank simulation.

The modeled tank measured approximately 24.99 m × 3.92 m × 3 m and contained a large array of server-rack elements.

The coolant system used 3M Fluorinert FC-40, with a representative inlet temperature of approximately −30 °C and flow velocity near 0.1 m/s.

The immersion system removed heat faster than it accumulated.

28.16 kW

Average thermal energy removal rate reported in the cooling validation.

3.50%

Reduction in internal energy rate over the evaluated four-hour period.

112.6 kWh

Total thermal energy removed over four hours based on the average rate.

Temperature-distribution simulations showed heat being transported away from the server racks through the circulating coolant.

The system exhibited a net decline in internal thermal energy, indicating that thermal energy was being removed faster than it accumulated.

The project concluded that single-phase immersion cooling could provide an effective and scalable approach for future high-density AI data centers.

COMSOL temperature distribution simulation of server cooling system

COMSOL temperature distribution across the modeled cooling system

COMSOL temperature comparison and convergence analysis

Temperature comparison and model convergence analysis

Cooling the server is only part of the problem.

The project also expanded beyond the local FEA model into a larger cooling-system concept including server tanks, pumps, a heat-pump loop, compression, condensation, and heat rejection.

Immersion Tank

Server hardware transfers heat directly into the dielectric fluid.

Circulation Pump

The coolant is circulated through the system to transport thermal energy away from the servers.

Heat Rejection

A reversible heat-pump architecture was considered for removing heat from the circulating coolant.

From one server tank to an entire AI data center.

The design study extended the cooling concept to server-farm layouts, pump buildings, maintenance spacing, and large-scale facility planning.

This helped connect the component-level thermal analysis to the system-level engineering required for data-center deployment.

Server immersion cooling tank schematic

Server tank and cooling-system schematic

Building server tank and pump cooling system schematic

Integrated server, tank, pump, and facility cooling-system architecture

Better flow distribution and thermal control.

Additional spacing between servers could reduce stagnant thermal zones and improve coolant convection.

Side-inlet and side-outlet configurations could produce more symmetric flow and reduce hot pockets near tank corners.

Low-temperature-compatible pumps, reservoirs, and insulated piping would also be required for practical implementation.

Thermal analysis at both component and system scale.

This project combined finite-element modeling, heat transfer, fluid mechanics, thermodynamics, and systems engineering to evaluate a scalable immersion-cooling strategy for high-density AI hardware.

FULL PROJECT DOCUMENTATION

AI Server Cooling — Final Engineering Presentation

View the complete ME 610 project presentation for additional details on server requirements, cooling-system architecture, engineering calculations, thermal analysis, COMSOL modeling, and system-level design.

View Full Project Presentation ↗
COMSOL Multiphysics Finite Element Analysis Heat Transfer Fluid Mechanics Nonisothermal Flow Thermal Management Immersion Cooling Thermodynamics System Design Data Center Engineering

More Engineering Work

View All Projects →