Nature-Inspired Optimization Algorithm-Tuned Feed-Forward and Recurrent Neural Networks Using CFD-Based Phenomenological Model-Generated Data to Model the EBW Process

TL;DR: In this article, both feed-forward and recurrent neural networks are developed for the electron beam welding process, which have been trained using the welding data collected from an existing computational fluid dynamics (CFD)-based phenomenological model with the help of some natured-inspired optimization tools like cuckoo search, firefly, flower pollination, crow search algorithms, particle swarm optimization, covariance adaptation evolution strategy and spider monkey optimization, separately.

Abstract: To automate the electron beam welding process, the identification of its contributing parameters is a must, for which it is required to establish the input–output correlations in both forward and reverse directions as accurately as possible. In the present investigation, both feed-forward and recurrent neural networks are developed for the said purposes, which have been trained using the welding data collected from an existing computational fluid dynamics (CFD)-based phenomenological model with the help of some natured-inspired optimization tools like cuckoo search, firefly, flower pollination, crow search algorithms, particle swarm optimization, covariance adaptation evolution strategy and spider monkey optimization, separately. The results of the trained networks have been validated using some real experimental data. The novelty of this study lies with the applications of these newly developed nature-inspired optimization algorithms to tune the neural networks using the CFD-based phenomenological model-generated welding data. In addition, the performances of these neural networks tuned using the said nature-inspired optimization algorithms have been compared through some statistical tests. In general, flower pollination-tuned recurrent neural network is found to provide the best predictions.