Software platform for high-fidelity-data-based artificial neural network modeling and process optimization in chemical engineering

• Published in: Computers & chemical engineering 2022-02, Vol.158, p.107637, Article 107637
• Main Authors: Moon, Jiyoung, Gbadago, Dela Quarme, Hwang, Gyuyeong, Lee, Dongjun, Hwang, Sungwon
• Format: Article
• Language: English
• Subjects: Artificial neural network (ANN); Chemical engineering; Data prediction; Genetic algorithm (GA); Software
• ISSN: 0098-1354; 1873-4375
• DOI: 10.1016/j.compchemeng.2021.107637

•Developed a high-accuracy python-based software platform for developing neural networks.•Implemented both prediction and process parameter optimization through genetic algorithms.•Higher performances obtained when benchmarked against several published literature.•Graphical plots in 2D and 3D are also possible with the software package.•All validation case studies employed yielded low training time and neural network accuracies >99%. Artificial neural networks are revolutionizing the field of engineering because of their ability to model complex non-linear problems without explicit programming. Their applications in different areas, such as manufacturing and healthcare, have provided a new path away from traditional modeling techniques. Nonetheless, users of this technology, especially those without prior knowledge of neural networks, spend a considerable amount of time gaining a basic understanding of the use of technology. Traditional trial-and-error approaches are often employed in training these neural networks, which further increases the time spent in developing them. Owing to the laborious nature of the trial-and-error method, the optimal or best hyperparameters of a particular neural network may not be determined, thereby affecting the accuracy of the developed model. Hence, in this study, a software platform is presented to aid in the training and development of neural networks by using genetic algorithms (GAs) for optimizing the model's hyperparameters, such as the number of neurons, learning rate, and activation function. As an essential aspect of chemical engineering processes, design or operating-parameter optimization is also included in this software package, wherein the best (optimized) weights and biases from the neural network are saved and employed in another GA to optimize key process variables, such as temperature, and velocity, as required by the user. This dual-purpose provides a complete application of neural networks that are primarily encountered in many engineering disciplines. The software platform can also plot 3D contours, heat maps (correlation plots), and other line graphs. For the validation and generalization of the software, it was benchmarked against five cases presented by different authors across various chemical engineering fields. The prediction results obtained using the software package were higher than those presented in the published literature, demonstrating the superior performance of the software package. [Dis