Unveiling the Artificial Neural Network Mystery with Special Reference to Applications in Civil Engineering

• Published in: Computational engineering and physical modeling Vol. 7; no. 3; pp. 18 - 44
• Main Authors: Shreenivas Londhe, Preeti Kulkarni, Pradnya Dixit, Shalaka Shah, Shardul Joshi, Pali Sahu Sahu
• Format: Journal Article
• Language: English
• Published: Pouyan Press 01.07.2024
• Subjects: artificial neural network; black box; civil engineering; influence; knowledge extraction; mystery
• ISSN: 2588-6959
• DOI: 10.22115/cepm.2024.461759.1315

Over the past few decades, Artificial Neural Networks (ANN) have carved a niche within the Civil Engineering applications and has been a supplementary to the traditional mathematical models. However, ANN being a ‘Black Box’ concealing the knowledge in weights and biases and thus not reflecting the underlying physics of the process making ANN difficult to accept and implement in practice. An attempt was made by the first and fourth author by developing the Knowledge extraction (KE) process and comprehending the underlying physics of process. The study was confined to modelling evaporation process. The present study endeavors by applying the KE process to a variety of Civil Engineering applications: water resources engineering, concrete technology, earthquake engineering, Soil mechanics and environmental engineering to uncover the mystery of ANN working. The trained weights and biases can further be processed to recognize the magnitude and influence of the respective input parameter on the output. Areas of applications discussed suggest that Evaporation is highly and directly dependent on average temperature, concrete strength has a high direct relation with cement content and indirectly related to w/c ratio, CO2 is major contributor towards carbonation coefficient, Dissolved oxygen shows a direct and high influence in water quality assessment and width of structure in dynamics of structures and is identified through magnitude and direct or indirect influence). The influence of input parameters and its relation with the output is unlocked through the biases and weights of each model and thus attempting to unveil the mystery of ANN.