Accurate magnetic hysteresis modeling remains a challenging task that is crucial for optimizing the performance of magnetic devices. Traditional methods, such as recurrent neural networks (RNNs), long short-term memory (LSTM) networks, and gated recurrent units (GRUs), have limitations when it comes to generalizing novel magnetic fields. This generalization is vital for real-world applications.
A team of researchers proposes a new approach using neural operators, specifically the Deep Operator Network (DeepONet) and Fourier Neural Operator (FNO). Neural operators differ from traditional neural networks in that they approximate the operator that maps the magnetic fields, thereby allowing for better generalization of novel fields. Moreover, the researchers introduce a rate-independent Fourier Neural Operator (RIFNO) that can predict material responses at different sampling rates, tackling the rate-independent characteristic of magnetic hysteresis.
The proposed method involves training neural operators using datasets generated from a Preisach-based model of a particular type of material. These datasets include first-order reversal curves and minor loops. The DeepONet architecture consists of two fully connected feedforward neural networks that approximate the magnetic fields. The FNO uses a convolutional neural network architecture with Fourier layers to transform the input tensor and approximate these fields.
The approach was evaluated for performance using three error metrics: relative error in L2 norm, mean absolute error (MAE), and root mean squared error (RMSE). FNO and RIFNO demonstrated higher accuracy and generalization capability compared to traditional recurrent architectures. Specifically, FNO showed the lowest errors, proving its effectiveness in modeling magnetic hysteresis. RIFNO maintained low prediction errors across various testing rates, showing robustness and good generalization under different conditions.
In conclusion, the researchers introduced a novel approach to modeling magnetic hysteresis that outperforms traditional methods in accuracy and generalization. The proposed methods of DeepONet and FNO, along with RIFNO, allow for efficient and accurate modeling of magnetic materials, enabling real-time inference and broadening the applicability of neural hysteresis modeling. This significant research advancement has the potential to revolutionize AI and its applications in magnetic devices. The researchers who contributed to this project deserve acknowledgement for their innovative approach.
