This research proposes a data processing pipeline employing Fourier analysis and deep neural networks to replicate the phenomenon of magnetic hysteresis, in particular frequency components derived from experimental data gathered using a newly developed 3D-printed material. The characterisation of hysteresis is essential for enhancing material performance and constructing precise models to anticipate material behaviour under diverse operating circumstances, especially in 3D-printed materials where properties can be meticulously regulated to ensure successful applications. The experimental signals were used for training and testing a neural network, exploiting Fourier coefficients to condense signals into the frequency components. This compression extracts fewer parameters and thus reduces and optimises the resources required by the neural network. It also improves the generalisation performance of the model, allowing it to make more accurate predictions on unseen data. This therefore optimises traditional modelling that requires a complete representation of hysteresis loops in the time domain, which must be addressed with the use of complex neural networks and large datasets. The experimental results show lower computational costs during the prediction process and a smaller memory footprint. Furthermore, the proposed model is easily adaptable for the loss estimation in different types of materials and input signals.

Efficient hysteresis characterization and prediction in 3D-printed magnetic materials using deep learning

Bertolini V.;Faba A.
2025

Abstract

This research proposes a data processing pipeline employing Fourier analysis and deep neural networks to replicate the phenomenon of magnetic hysteresis, in particular frequency components derived from experimental data gathered using a newly developed 3D-printed material. The characterisation of hysteresis is essential for enhancing material performance and constructing precise models to anticipate material behaviour under diverse operating circumstances, especially in 3D-printed materials where properties can be meticulously regulated to ensure successful applications. The experimental signals were used for training and testing a neural network, exploiting Fourier coefficients to condense signals into the frequency components. This compression extracts fewer parameters and thus reduces and optimises the resources required by the neural network. It also improves the generalisation performance of the model, allowing it to make more accurate predictions on unseen data. This therefore optimises traditional modelling that requires a complete representation of hysteresis loops in the time domain, which must be addressed with the use of complex neural networks and large datasets. The experimental results show lower computational costs during the prediction process and a smaller memory footprint. Furthermore, the proposed model is easily adaptable for the loss estimation in different types of materials and input signals.
2025
File in questo prodotto:
Non ci sono file associati a questo prodotto.

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/11391/1594574
Citazioni
  • ???jsp.display-item.citation.pmc??? ND
  • Scopus 0
  • ???jsp.display-item.citation.isi??? 0
social impact