Machine Learning Models for the Design and Optimization of Stack-Structured Permanent Magnet Thrust Bearings

This paper presents artificial neural network (ANN) models for the design and optimization of multi-ring permanent magnet thrust bearings (MRPMTB) based on results from three-dimensional (3D) mathematical models. To begin, a mathematical model is presented to calculate the axial force and stiffness of MRPMTB. Then, to understand the influence of bearing geometrical dimensions on its features, a single-ring thrust permanent magnet bearing (PMB) was optimized within a control volume to maximize axial force and stiffness, using modified MATLAB codes to solve the force and stiffness equations. Furthermore, based on the optimization results, comprehensive datasets of the optimal geometric dimensions of bearings for maximum axial force and stiffness within a control volume are generated using MATLAB codes. Five ANN models are developed to predict optimal geometric parameters and corresponding maximum force and stiffness values using generated datasets. Out of the applied ANN models, the Bayesian Multi-layer Perceptron gave higher values of R² and lower mean absolute error (MAE). Finally, the ANN model results are validated using a finite element simulation tool (COMSOL Multiphysics).