Abstract

Electronic cooling is essential for maintaining optimal performance during complex computation tasks. Computational simulations have become a key tool to assess performance under various operating conditions. However, computational modeling involves a lengthy process to reach the final solution. The present work introduces a point-based neural network model to generate instant predictions of temperature for various heat sink configurations. A design of experiments methodology was utilized to define different heatsink designs, which were simulated using OpenFOAM. The training data included fin length, fin width, thermal conductivity, and Cartesian coordinates of point clouds. PyTorch was enabled with a fully connected neural network for learning the temperature distributions. The point-based neural network model took about 2 hours to train and about 0.04 seconds to predict the test data set with a per-sample prediction time of 0.07 microseconds. The model achieved a root mean squared error of 0.056 K and a mean absolute error of 0.030 K when compared to the simulation. Analysis with different noise levels revealed the machine learning model being robust for up to 70 mK noise. Results indicate that 27 geometries are required for the neural network to learn the temperature distributions in the analyzed range of parameters. Temperature contours from the simulation and the machine learning model were compared as well, revealing that the machine learning model accurately and instantly reproduces the thermal behavior given by the simulation.

Publication Date

8-7-2026

Document Type

Thesis

Student Type

Graduate

Degree Name

Mechanical Engineering (MS)

Department, Program, or Center

Mechanical Engineering

College

Kate Gleason College of Engineering

Advisor

Isaac Perez-Raya

Advisor/Committee Member

Michael Richards

Advisor/Committee Member

Michael Schertzer

Comments

This thesis has been embargoed. The full-text will be available on or around 8/13/2027.

Campus

RIT – Main Campus

Download

Share

COinS