65169 - Artificial Neural Networks: From Basics to Turbomachinery Applications 

Machine learning in the past decade has found a key role in many engineering applications [1]. In particular, artificial neural networks, from the first formulations to the most complex architectures, have shown excellent capability in handling complex problems. The aim of the tutorial is to provide an insight on how neural network work, including in the dissertation several relevant experiences from the authors. The tutorial will be structured in three parts.

The first one will be dedicated to the basic neuron structure and the underlying mathematical formulation of simple neural networks. Additional topics, as algorithm optimization, batch feeding and activation functions will be touched (see also [2]). More complex structures will be gradually added, ending with the analysis of multi-layer neural networks for classification   and regression tasks. Advanced algorithms of significant interest for scientific and engineering applications, i.e. Convolutional, Recurrent and Adversarial neural networks, will also be included in the dissertation.

A description of the preprocessing techniques and problematics will follow in the second part. There we will focus on describing the best practices for preliminary data analysis and feature selection. We will describe the usage of statistical tools, such as correlation matrices, frequency plots or ANOVA. The second part will be concluded by presenting the capability of dimensionality reduction algorithms, e.g. POD or PCA, applied to turbomachinery data preprocessing.

In the third part of the tutorial, we will provide a detailed description of some engineering problems that can be solved using neural networks. The shown applications will include:

ü  The modelling of the boundary layer in run-time computation of CFD solvers, with feed forward neural networks [3]

ü  The generation of better physical inflow conditions for numerical simulation with adversarial generative neural networks [4]

ü  Neural network as fitness function in evolutionary optimization algorithms [5].

A summary of the tutorial and a Q&A session will conclude the tutorial.

[1] Rafiq, M. Y., G. Bugmann, and D. J. Easterbrook. "Neural network design for engineering applications." Computers & Structures 79.17 (2001): 1541-1552.

[2] Goodfellow, Ian, et al. “Deep learning”. Vol. 1. No. 2. Cambridge: MIT press, 2016.

[3] Tieghi, Lorenzo, et al., “Assessment of a machine-learnt adaptive wall-function in a compressor cascade with sinusoidal leading edge”, GTP20-1311

[4] Delibra, Giovanni, et al., “Machine learnt synthetic turbulence for LES inflow conditions”, GT2020-15338.

[5] Angelini, Gino, et al. "On Surrogate-Based Optimization of Truly Reversible Blade Profiles for Axial Fans." Designs 2.2 (2018): 19.