Ki Yong Lee, YoonJae Shin, YeonJeong Choe, SeonJeong Kim, Young-Kyoon Suh, Jeong Hwan Sa, JongSuk Luth Lee, Kum Won Cho

KIPS Transactions on Software and Data Engineering, Vol. 5, No. 11, pp. 593-606, Nov. 2016
10.3745/KTSDE.2016.5.11.593,   PDF Download:
Keywords: Simulation, Simulation Result Prediction, Statistical Machine Learning
Abstract

Computer simulation is widely used in a variety of computational science and engineering fields, including computational fluid dynamics, nano physics, computational chemistry, structural dynamics, and computer-aided optimal design, to simulate the behavior of a system. As the demand for the accuracy and complexity of the simulation grows, however, the cost of executing the simulation is rapidly increasing. It, therefore, is very important to lower the total execution time of the simulation especially when that simulation makes a huge number of repetitions with varying values of input parameters. In this paper we develop a simulation service system that provides the ability to predict the result of the requested simulation without actual execution for that simulation: by recording and then returning previously obtained or predicted results of that simulation. To achieve the goal of avoiding repetitive simulation, the system provides two main functionalities: (1) storing simulation-result records into database and (2) predicting from the database the result of a requested simulation using statistical machine learning techniques. In our experiments we evaluate the prediction performance of the system using real airfoil simulation result data. Our system on average showed a very low error rate at a minimum of 0.9% for a certain output variable. Using the system any user can receive the predicted outcome of her simulation promptly without actually running it, which would otherwise impose a heavy burden on computing and storage resources.

Statistics
Cite this article