Graduation Semester and Year




Document Type


Degree Name

Master of Science in Aerospace Engineering


Mechanical and Aerospace Engineering

First Advisor

Alan P Bowling


Modern data centers consume an astonishing 1.3\% power all around the world. As the number of data centers continue to grow, there is an increasing need and demand to develop new ways to reduce the power footprint. Several approaches are being made to achieve this. One of such several approaches is to develop control systems that would keep the data centers running energy efficiently. Various control theories have been developed throughout the world to achieve the optimal energy efficient state. However, during the synthesis of such control schemes, the CFD simulations take up excessive time for plotting the thermal map of such complex, dynamic and highly nonlinear data center systems. In this paper, we aim to develop and train artificial neural networks for a typical scaled setup of modern data center like a Black Box Model which would predict the temperature at points in the state space throughout the room as a function of the dissipating heat at those points and CRAC fan speeds at the time. Due to significantly low analysis time than computational fluid dynamics, the Black Box is able to predict the temperatures in real time at different points in the setup thereby enabling faster optimization analysis. The Neural model is trained on a huge set of data generated by CFD simulations from hypothetical arrangements in a data center. Discussion about neural network training functions, its training parameters and comparisons of accuracy and computational time and the reason for the same is also done in this paper. Various suggestions to train such highly non linear and dynamic systems are summarized. To prove the accuracy of the neural network, the data generated is compared to the output of the CFD model. The robustness of the Black Box within the training data limits has been verified for changing CRAC fan speed and server heat. The Black Box tool developed is not only accurate but also very fast which enables its use in a feed forward adaptive control setup or the dynamic learning setup or both. Such a Black Box tool mimicking the CFD proves very useful in development of control systems for data centers.


Aerospace Engineering | Engineering | Mechanical Engineering


Degree granted by The University of Texas at Arlington