Machine learning in Industrial turbomachinery: development of new framework for design, analysis and optimization

In the age of Industry4.0, turbomachinery community is facing a significant effort in rethinking all manufacturing activities. A new industrial paradigm is developing, which encourages the introduction of Big-Data, Internet of Things (IoT) and Artificial Intelligence (AI) into the production and manufacturing environment to follow the market trend of innovation and strong customization of product. This thesis aims to improve the classical approach to turbomachinery design, analysis and optimization, which relays to empiric procedures or computational intensive CFD based optimizations, implementing new framework and strategies based on the use of Machine Learning and Deep Learning, the successful sub-fields of AI, recognized as crucial technologies to stay competitive in this dynamiclandscape. Theintroductionoflearningalgorithmschangestheglobalstructureof classic approach determining modern design support tools. The introduction of regression models facilitates the exploration of design space providing easy-to-run analytical models. Regression models can also be used as cost functions in multi-objective optimization problems avoiding the use of time consuming simulations. The application of unsupervised learning methods to turbomachine performance data extracts new insights to improve the designer understanding of blade-flow interaction and allows a better formulation of specific problem. The design space exploration artificial neural network based allowed the creation of multi-dimensional Balje charts, enriched using correlations between the main interesting geometric parameters and the rotor performance, which guide designer in preliminary blade geometry selection. Surrogate model based optimization allowed reducing thecomputationalcostrequiredbycanonicalapproachproducingreliablesetofbettersolutions in comparison with the standard optimization. Application of unsupervised learning techniques allows developing a new parametrization strategy based on statistical shape representation of complex three dimensional surfaces. This work can be ideally divided into two parts. The first reviews the state-of-the-art of artificial intelligence techniques highlighting the role of machine learning in supporting design, analysis and optimization for turbomachinery community. Second part reports the development of a set of strategies to exploit metamodeling algorithms improving the classical design approach.