Extreme learning machine as a new framework in predicting material properties: Methodology and comparison

Material properties are very important in most mechanical engineering computations. Numerous approaches have been proposed to estimate these material properties such as State of Equations, Statistical Regression, and Neural Networks modeling schemes. Unfortunately, accuracy of some of these earlier approaches is often limited. Recently, extreme learning machine has been proposed as a new computational intelligence paradigm in building decision based on both hidden and unhidden information. This paper proposes the extreme learning machine in identify and predict the material properties, such as, the strength of material (gray cast iron). A comparative study on system performance is conducted between extreme learning machine and the most common existing data mining and machine learning modeling techniques, namely, multiple regression, backpropagation neural networks, and radial basis function. Results show that extreme learning machines needs up to four orders of magnitude less training time compared to neural network. The forecasting/classification accuracy of extreme learning machine is also higher than those of existing modeling schemes. For given network architecture, the new computational intelligence scheme does not have any control parameters, such as, stopping criteria, learning rate, learning epoches, etc) to be manually tuned and can be implemented easily.