Abstract Preview

Here is the abstract you requested from the Thermal_2018 technical program page. This is the original abstract submitted by the author. Any changes to the technical content of the final manuscript published by IMAPS or the presentation that is given during the event is done by the author, not IMAPS.

Methodology to develop Neural Network models for predictive cooling of data centers
Keywords: Neural Network Models, Predictive Cooling, Data Centers
Various applications of artificial neural networks on datacenters have been coming up recently in the past few years. Artificial Neural Network (ANN) mimics the datacenter’s operational characteristics to develop and adopt control methods for the systems that maintain the datacenters in desirable operating conditions. Finding physical interactions happening inside a Datacenter and parametrically characterizing their interdependency and non-linearity is challenging and difficult to understand based on gained domain knowledge. This research mainly focuses on capturing the important features that are necessary in developing an ANN model that adequately represents the key physical interactions present in an operational datacenter. Dimensionality reduction improves the learning process by extracting the most important data representations, possibly the parameter space having the maximum information of the original data and have better generalization capabilities. It is quite desirable not only in the aspect of the quality and desirable data, but also in terms of data storage and computational complexity. We use Feature Extraction, a statistical dimensionality reduction technique that can leverage the physical relations between the spatial parameters when constructing a datacenter. Python software with certain libraries is useful tools to perform the mathematical analysis of feature extraction process retaining the maximum variance in the parameters. Artificial Neural Networks have the ability to model more accurate physical relations after the dataset is generated. CFD models are created using 6SigmaRoom software to perform steady state analysis on datacenters and for generation of training datasets. This study leads to the formation of a methodology to address important features to consider while modelling an efficient neural network thereby increasing the potential of system control strategies. Funding for this work was provided by an ES2 partner company
Vibin Shalom Simon,
University of Texas at Arlington
Arlington, TX

  • Amkor
  • ASE
  • Canon
  • Corning
  • EMD Performance Materials
  • Honeywell
  • Indium
  • Kester
  • Kyocera America
  • Master Bond
  • Micro Systems Technologies
  • MRSI
  • Palomar
  • Promex
  • Qualcomm
  • Quik-Pak
  • Raytheon
  • Rochester Electronics
  • Specialty Coating Systems
  • Spectrum Semiconductor Materials
  • Technic