1:10pm - 1:30pmID: 121
/ Oral Session 2: 1
Abstract submission for on-site presentation
Topics: Inverse problem, ApplicationKeywords: Multi-objective optimization, neural network surrogate model
Multi-objective Optimization of Inductors Based on Neural Network
Xiaohan Kong, Hajime Igarashi
Graduate School of Information Science and Technology, Hokkaido University, Japan
This work proposes a method to perform multi-objective optimization of inductors using a surrogate model based on neural network (NN) instead of finite element method (FEM). Traditional design methods require repetitive FEM calculations, resulting in a very long overall design time. In contrast, this work utilizes a well-trained neural network to predict magnetic core loss, the volume and saturation current, avoiding repetitive FEM evaluations and significantly reducing the total time for inductor optimization.
1:30pm - 1:50pmID: 125
/ Oral Session 2: 2
Abstract submission for online presentation
Topics: Theoretical aspects and fundamentalsKeywords: Electromagnetic Computation, Neural Networks, Surrogate Models
Learning to solve Electromagnetic Problems: a Comparison among Different Machine Learning Approaches
Alessandro Formisano1, Mauro Tucci2
1Università della Campania "Luigi Vanvitelli", Italy; 2Università di Pisa, Italy
The possibility of adopting data-driven procedure to create a model of electromagnetic problems has been investigated since long. Recently, the availability of high-performance computing systems even at desktop level has provided different model of “artificial intelligence” processors able to deal with such a problem, examples being Physically-Informed Neural Networks, Generative Adversarial Networks. Etc. In this study, using a simple yet representative benchmark problem, some of the most common solutions are compared, with the aim of highlighting respective advantages and drawbacks.
1:50pm - 2:10pmID: 102
/ Oral Session 2: 3
Abstract submission for on-site presentation
Topics: Application, AlgorithmsKeywords: Eddy currents, Gain tuning, Impulse response, Levitation devices, Model order reduction
Optimal Parameter Estimation for Conductor Movement Using Cauer Ladder Network Representation and Virtual Time-response Based Iterative Gain Evaluation and Redesign
Naoto Tanimoto1, Atsuki Fujita1, Kengo Sugahara1, Manabu Kosaka1, Yasuhito Takahashi2, Tetsuji Matsuo3
1Kindai University, Japan; 2Doshisya University, Japan; 3Kyoto University, Japan
This article propose an optimal design of control gain parameters using a Cauer ladder network with constant basis functions and an Impulse-response model based V-Tiger. We model the TEAM Workshop Problem 28 by the Cauer ladder network method, which is a model order reduction technique, and enable the analysis by a circuit simulator such as Simulink. We determine the control gain parameters by applying the Impulse-response model based V-Tiger to the data obtained from the analysis. This method is useful for efficiently determining control parameters on simulation.
2:10pm - 2:30pmID: 101
/ Oral Session 2: 4
Abstract submission for on-site presentation
Topics: AlgorithmsKeywords: EM-driven design, parameter tuning, global optimization, inverse modelling, simulation models
Globalized High-Frequency Optimization Using Inverse Models
Slawomir Koziel1,2, Anna Pietrenko-Dabrowska2
1Reykjavik University; 2Gdansk University of Technology
This paper discusses a novel technique for quasi-global parameter tuning of high-frequency structures using response features and inverse surrogate models. Our approach enables a low-cost identification of the most promising parameter space regions, followed by a fine tuning by means of local routines. The presented method is validated using several high-frequency structures. Its global search capability and computational efficiency are demonstrated by extensive comparisons with multiple-start local search as well as nature-inspired optimizers.
|