Performance analysis of Levenberg-Marquardt and Steepest Descent algorithms based ANN to predict compressive strength of SIFCON using manufactured sand

Abstract

The advent of rapid industrialization and urbanization all over the world has led to the depletion of sand, the key structural material in the concrete structures and there is an urgent need of an alternative to the natural sand. In this paper, investigations are carried out for assessing the strength of Slurry Infiltrated Fibrous Concrete (SIFCON) with different percentage fibre fractions using manufactured sand in place of natural sand by experimental method and prediction models. SIFCON is a special type of new steel fibre reinforced concrete structure and is considered as a relatively high performance concrete. An important mechanical property of the concrete is the compressive strength, which can be measured after curing the concrete specimens for a standard curing period of 7, 28 and 56 days. The application of artificial neural network (ANN) models to predict the highly desirable compressive strength of SIFCON has great potential in arriving at solution to inverse problems that are non linear in nature and proves advantageous over the conventional methods. In this study, Levenberg-Marquardt (LM) Algorithm and Steepest Descent (SD) Algorithm based ANN models are used in predicting the compressive strength of SIFCON concrete that contains cement, manufactured sand and various percentage fibre fractions, at different curing times. A total of 15 ANN architectures are implemented and 2-4-14-1 architecture is proved to be the best possible architecture in the case of LM and 2-2-15-1 architecture in the case of SD algorithm. The optimized configuration provides a lower Root Mean Squares (RMS) Error. The LM requires less number of iterations compared to SD, apart from high accuracy (95%), and fast convergence. The proposed ANN methods demonstrate that they are practical and beneficial to predict strength of concrete. The results reveal that although SD algorithm was accurate in predicting the compressive strength of SIFCON, yet LM algorithm was more accurate than SD algorithm.

Keywords

• Compressive strength
• SIFCON
• Manufactured sand
• Artificial neural networks
• Levenberg-Marquardt
• Steepest Descent method