Convolutional Neural Network Approach for Impeller Blade Loading Inference

Artificial intelligence (AI) techniques are increasingly being adopted in engineering applications to accelerate design processes and optimize performance. Among these, neural networks have proven particularly effective in capturing complex, non-linear relationships between geometric inputs and overall performance outputs. These correlations are essential for enabling efficient design and optimization. Nevertheless, a significant limitation of such approaches lies in their frequent disregard for the underlying flow physics. While computational fluid dynamics (CFD) simulations provide detailed insights into the flow field, this information is typically unexploited in conventional neural network models. Developing a methodology that can leverage this rich dataset effectively is crucial for combining global performance metrics with detailed local flow characteristics, offering a comprehensive evaluation of the system's behavior without increasing the time required for database generation.This work presents a methodology that leverages advanced neural network architectures to predict flow characteristics in a transonic impeller. Geometric inputs are generated using an in-house parametric tool, and the outputs are categorized into two types: global performance parameters, which are used to train a multilayer perceptron (MLP), and blade loading data represented as pressure fields, which are collected as images. These images are then employed to train a reverse convolutional neural network (CNN). The trained network provides predictions of both overall performance and blade loading for any given input geometry. The use of pressure field images offers designers enhanced tools and insights into the flow characteristics of the machine, enabling a deeper analysis of key features and performance metrics.