Training Neural Networks to Predict the Energy Efficiency of Screw Rotor Profiles

Artificial Neural Networks (ANN) are emerging as promising tools for advancements in state-of-the-art design and optimization techniques. Twin screw compressor technology is matured in all aspects of design and manufacturing. But the potential application of Artificial Intelligence (AI) or Machine Learning (ML) has not yet been explored in this domain. This paper attempts at training an ANN, which is a class of AI/ML techniques, to predict the energy efficiencies of twin screw compressors for different rotor profile shapes. The N profile by N. R. Stosic (1997) is chosen as the profile system to generate multiple retrofitted profiles by varying only five curve parameters. Generally, the energy efficiency of a screw compressor with a certain rotor profile is calculated by solving the chamber model which is a thermodynamic simulation of the compression process. The objective of this study is to check if an ANN can be successfully trained with large enough data of different profiles and their respective energy efficiencies to capture the physics of the compression process and predict the energy efficiencies for given profiles with reasonable accuracy. It has been found that the ANN is able to learn the pattern associated with profile shapes and their energy efficiency to a fair degree. But increasingly large data sets are required for training the ANN to achieve a higher accuracy of prediction. This work stands as a pilot study to explore further possibilities for use of these techniques in rotor profiling and/or screw compressor design and optimization.