betaFeedback

WorksComputer Science

Neural oscillators for magnetic hysteresis modeling

Abstract

Hysteresis is a ubiquitous phenomenon in science and engineering; its modeling and identification are crucial for understanding and optimizing the behavior of various systems. We develop an ordinary differential equation-based recurrent neural network (RNN) approach to model and quantify the hysteresis, which manifests itself in sequentiality and history-dependence. Our neural oscillator, HystRNN, draws inspiration from coupled-oscillatory RNN and phenomenological hysteresis models to update the hidden states. The performance of HystRNN is evaluated to predict generalized scenarios, involving first-order reversal curves and minor loops. The findings show the ability of HystRNN to generalize its behavior to previously untrained regions, an essential feature that hysteresis models must have. This research highlights the advantage of neural oscillators over the traditional RNN-based methods in capturing complex hysteresis patterns in magnetic materials, where traditional rate-dependent methods are inadequate to capture intrinsic nonlinearity.

Introduction

Magnetic hysteresis pertains to a prevalent observed phenomenon in ferromagnetic and ferrimagnetic materials where the change in magnetization response lags behind variations in the applied magnetic field. Specifically, the hysteresis effect is characterized by a delay in the magnetic flux density (BB) to changes in the applied magnetic field strength (HH), exhibiting history dependency, nonlinearity and non-monotonicity [1]. The relationship between BB and HH fields is represented as a hysteresis curve CC (B−HB-H curve), which plays a pivotal role in comprehending hysteresis and governing the magnetization process during alterations in HH (Figure 1). The hysteresis loop offers insights into material behavior; for instance, the area of the hysteresis loop signifies the energy dissipated as heat during each cycle of magnetization and demagnetization.

Figure 1. B−HB-H magnetic hysteresis curves (FORC = first-order reversal curve)

Accurate hysteresis modeling is pivotal in enhancing the operational efficiency of electrical machines. For instance, in engineering systems that involve the movement of cables, hysteresis requires more sophisticated control strategies to compensate for its effects [2]. Similarly, the efficiency of electrical machines is intrinsically linked to the precise modeling of the hysteresis characteristics exhibited by the steel materials employed [3]. Incorporating a robust model would avoid the costly manufacturing of multiple prototypes. Mathematically, the primary objective of hysteresis modeling is to predict the sequence of BB values that correspond to a given sequence of HH values. However, the relationship between BB and HH defies the mathematical definition of a single-valued function. Consequently, conventional function approximation techniques are not suitable for modeling hysteresis as a function with domain HH and codomain BB [4].

Traditionally, modeling the hysteresis behavior is based on fundamental principles of physics [5]. However, in practical engineering scenarios, the manifestation of hysteresis behavior often stems from complex, large-scale effects and the multiphysical nature of the system, rendering deterministic models inaccurate [5, 1]. Consequently, phenomenological models are employed, establishing connections between desired behaviors and specific underlying phenomena rooted in principles of thermodynamics or elasticity, for instance. Notable phenomenological models include the Preisach [6, 1], Jiles-Atherton [7], and Bouc-Wen models [8, 9]. Generalizing these models across disciplines and incorporating them into systematic modeling approaches, fitting them to experimental data, and integrating them into other mathematical models pose significant challenges [10], such as sophisticated optimization techniques and increased computational burden [11].

To mitigate these limitations of phenomenological models, feed-forward neural networks (FFNNs) are commonly used for modeling magnetic hysteresis [12, 13, 14, 15]. However, owing to the absence of a functional relationship between BB and HH fields, the traditional FFNN approach with input HH and output BB is inadequate and suboptimal. Instead, studies [13, 14] propose HH and B−1B_{-1} as input and BB as output during training, where B−1B_{-1} denotes the previous BB value [4]. The notation B−1B_{-1} is discussed in detail in the ’Method’ section. However, this approach is characterized by two notable limitations. First, it lacks the incorporation of sequential information and fails to capture interdependencies among output values, hence not respecting the underlying physics of the problem. Second, this strategy exhibits limitations in generalizing to new situations, as it relies on single-step prior information during training [16]. Consequently, these models struggle to extrapolate to scenarios beyond training data, limiting broader applications that require robust generalization.

To address the limitations faced by FFNNs, models centered on recurrent neural networks (RNNs) [4, 17] have been employed, which provide a natural framework for modeling the sequential hysteretic nature. However, the models that employ traditional RNNs, gated recurrent unit (GRU) [18], long-short-term memory (LSTM) [19] and their variants exhibit limitations regarding their ability to generalize effectively to unseen HH variations, as we present in the current work. Although these recurrent networks model the underlying relationship and predict hysteresis loops exceedingly well as an interpolation task, the primary objective of achieving robust generalization remains inadequately addressed [20]. An optimal recurrent-based technique should excel in both interpolation tasks and demonstrate reasonable accuracy in generalization, effectively predicting BB sequences for unseen HH sequences.

A possible approach to accomplishing efficient generalization could be to enforce the recurrent architecture to incorporate the underlying dynamics. An efficient way to represent time-varying dynamics involves representation through ordinary differential equations (ODEs) and dynamical systems, recognized for their capacity to model diverse, intricate and nonlinear phenomena across natural, scientific, and engineering domains [21]. This inclusion of inherent dynamics, which should effectively encapsulate crucial physical attributes of the underlying magnetic material, motivates us to employ a system of ODEs to update the hidden states of the recurrent architecture, referred to as neural oscillators. Recently, neural oscillators have shown significant success in machine learning and artificial intelligence and have been shown to handle the exploding and vanishing gradient problem effectively with high expressivity [22, 23, 24, 25, 26]. The recent universal approximation property [27] also supports our belief in modeling magnetic hysteresis with neural oscillators.

Our neural oscillator, referred to as HystrRNN (hysteresis recurrent neural network), is influenced by the principles of the coupled-oscillator recurrent neural network (CoRNN) [22], which integrates a second-order ordinary differential equation (ODE) based on mechanical system principles. CoRNN considers factors such as oscillation frequency and damping in the system. However, for magnetic hysteresis, these physical attributes are less significant. Instead, we focus on embedding the hysteretic nature within the ODE formulation. We leverage phenomenological differential hysteresis models to accomplish this, recognizing that models like Bouc-Wen [8, 9], and Duhem [28] utilize the absolute value function to represent the underlying dynamics. By incorporating this function into our model, we can effectively capture and control the shape of the hysteresis loop. This integration into the recurrent model is expected to facilitate robust generalization, as it inherently captures the shape of the hysteresis loop, preserving symmetry and structure.

In this manuscript, we model nonoriented electrical steel (NO27) to test the validity of the proposed method for magnetic materials. The hysteresis loops employed and modeled in this work are acquired using the Preisach model for an Epstein frame. This model has been adjusted to adhere to the IEC standard, and the core was assembled using 16 strips of NO27-1450H material. More details about the Preisach model are provided in supplementary material SM §B. We train all our models on a major loop (represented by a blue loop in Figure 1(a)) and use the trained model to test two different generalization tasks: predicting first-order reversal curves (FORCs, represented by red curves in Figure 1(b)) and minor loops (represented by a red loop in Figure 1(c)). We perform experiments for four different BB fields with applications relevant to modeling electrical machines.

The remainder of the manuscript is structured as follows. The section ‘Generalization in hysteresis’ presents the challenge of this manuscript and explains how the task amounts to generalization. The section ‘Method’ formulates the proposed HystRNN method and explains it in detail. The section ‘Numerical experiments’ validates the proposed method through a series of numerical experiments. Finally, the ‘Conclusions’ section collates the key findings and implications of this study.

Generalization in hysteresis

Traditional supervised machine learning methodologies employed to model magnetic hysteresis train the model for (Hi,Bi)∈C1(H_i, B_i) \in C_1, where 1≤i≤N1 \le i \le N, i∈Zi \in\mathbb{Z} and NN is the number of training samples. The trained model is then tested on (Hk,Bk)∈C2(H_k, B_k) \in C_2, where 1≤k≤M1 \le k \le M, k∈Zk \in\mathbb{Z} and MM is the number of testing samples. Traditionally, C2⊂C1C_2 \subset C_1, with Hi≠HkH_i \ne H_k. However, this prediction reduces to an interpolation task [29]. In contrast, we are interested in training the model for (Hi,Bi)∈C1(H_i, B_i) \in C_1 and predicting a hysteresis trajectory for (Hk,Bk)∈C2(H_k, B_k) \in C_2, where C2⊄C1C_2 \not\subset C_1, and C2∩C1=ϕC_2 \cap C_1 = \phi. Here, ϕ\phi denotes the null set. Precisely, we train all our models on the major loop (CmajorC_{\mathrm{major}}) as shown in Figure 1(a). Then the trained model is tested for two different scenarios. First the FORCs (CFORCC_{\mathrm{FORC}}) shown in Figure 1(b) and second the minor loops (CminorC_{\mathrm{minor}}) presented in Figure 1(c).

Modeling FORCs and minor hysteresis loops play a significant role in analyzing magnetic materials. FORC modeling reveals intricate interactions, enabling the differentiation between magnetization components that can be reversed and those that cannot. This knowledge is pivotal in the optimization of magnetic devices like memory and sensor technologies [30, 31, 32]. Minor hysteresis loop modeling complements this by providing insights into localized variations in magnetic behavior.

More insights on how predicting CFORCC_{\mathrm{FORC}} and CminorC_{\mathrm{minor}} entails to a generalization task is provided in Figure 2. In Figure 2(a) and 2(b), the time series of HH and BB fields are shown, respectively. The blue curve represents the training data (HH vs BB is CmajorC_{\mathrm{major}}), and the red curve represents the region in which the prediction is sought (CFORCC_{\mathrm{FORC}} in this case). The black curve behind the shaded region signifies the history that the material has gone through, which is the series of magnetization and demagnetization that is unknown at the time of testing. Hence, this task amounts to extrapolation in time or predicting in a generalized scenario. Similarly, Figure 2(c) and 2(d) represent the case for CmajorC_{\mathrm{major}} and CminorC_{\mathrm{minor}}.

Figure 2. Variation of magnetization and demagnetization in a magnetic material over time

Method

HystRNN utilizes a recurrent structure akin to RNNs, with the difference being in the hidden state update. HystRNN employs ODEs for updating the hidden states. The approach involves two inputs, HH and B−1B_{-1}, which are mapped to BB. The modeling process begins by collecting NeN_e number of experimental data points (Hi,Bi−1=Bi)(H_i, B_{i-1} = B_i) for CmajorC_{\mathrm{major}}, where 1≤i≤Ne1 \le i \le N_e, and i∈Zi \in\mathbb{Z}. Subsequently, (Hj,Bk)(H_j, B_k) and BjB_j are taken as the input and output of HystRNN, respectively, where 2≤j≤Ne2 \le j \le N_e, 1≤k≤Ne−11 \le k \le N_e - 1, k=j−1k = j - 1, and j,k∈Zj,k \in\mathbb{Z}. The number of training points is denoted by N=Ne−1N = N_e - 1. While sharing certain similarities with some feedforward neural network (FFNN) architectures employed for modeling hysteresis, this training approach diverges by incorporating a recurrent relationship that captures longer-time dynamics and output dependencies, which are absent in FFNNs. Next, the hidden states of HystRNN are updated using the following second-order ODE

Test caseL2-norm (↓)Explained variance score (↑)Max error (↓)Mean absolute error (↓)
RNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNN
CFORC1C_{\mathrm{FORC}1}5.02040.85250.77640.2198-0.07210.10810.20070.82525.35972.40892.00751.20302.98881.19671.55500.6149
CFORC2C_{\mathrm{FORC}2}6.48770.52550.47010.3085-0.25450.18750.23950.88445.34841.84281.81771.23713.60380.93270.87230.7613
Cminor1C_{\mathrm{minor}1}5.35061.43821.80280.0438-0.10130.02980.07760.98392.76411.70981.89250.31081.48770.71420.77970.1258
Cminor2C_{\mathrm{minor}2}12.36711.57852.05630.0786-2.70460.02480.06730.96613.74911.57261.74860.36301.97030.65440.73410.1450

Table 1. The generalization performance assessed using the metrics: L2-norm relative error, explained variance error, maximum error, and mean absolute error for the first experiment, where max⁡(B)=1.7\max(B) = 1.7 T. For these metrics, higher (respectively, lower) values are favored for (↑) (respectively, (↓)). The implication of arrows remains consistent for all the following Tables.

y′′=σ1(W1y+W1y′+V1u+b1)+σ2(W2∣y∣2+W2∣y′∣2+V2∣u∣2+b2).(1)y^{\prime\prime} = \sigma_1\left(\mathbf{W}_1y+\mathcal{W}_1y^{\prime}+\mathbf{V}_1u+b_1\right) +\sigma_2\left(\mathbf{W}_2\lvert y\rvert^2+\mathcal{W}_2\lvert y^{\prime}\rvert^2+\mathbf{V}_2\lvert u\rvert^2+b_2\right). \tag*{(1)}

Here, the hidden state of the HystRNN is denoted by y=y(t)∈Rmy=y(t)\in\mathbb{R}^{m}. y′y^{\prime} indicates a time derivative, while y′′y^{\prime\prime} indicates a second-order time derivative. W1,W2,W1,W2∈Rm×m\mathbf{W}_1,\mathbf{W}_2,\mathcal{W}_1,\mathcal{W}_2\in\mathbb{R}^{m\times m}, and V1,V2∈Rm×n\mathbf{V}_1,\mathbf{V}_2\in\mathbb{R}^{m\times n} are the weight matrices, n=N×2n=N\times2, and tt corresponds to the time at which the training data has been collected. u=u(t)∈Rnu=u(t)\in\mathbb{R}^{n} is the input to HystRNN. b1,b2∈Rmb_1,b_2\in\mathbb{R}^{m} are the bias vectors. The activation functions σ1,2:R↦R\sigma_{1,2}:\mathbb{R}\mapsto\mathbb{R} are taken to be σ1,2(u)=tanh⁡(u)\sigma_{1,2}(u)=\tanh(u). By setting z=y′(t)∈Rmz=y^{\prime}(t)\in\mathbb{R}^{m}, (1) becomes a system of first-order ODEs

y′=z,z′=σ1(W1y+W1z+V1u+b1)+σ2(W2∣y∣2+W2∣z∣2+V2∣u∣2+b2).(2)y^{\prime}=z,\qquad z^{\prime}=\sigma_1\left(\mathbf{W}_1y+\mathcal{W}_1z+\mathbf{V}_1u+b_1\right) +\sigma_2\left(\mathbf{W}_2\lvert y\rvert^2+\mathcal{W}_2\lvert z\rvert^2+\mathbf{V}_2\lvert u\rvert^2+b_2\right). \tag*{(2)}

Discretizing the system of ODEs (2) using an explicit scheme for 0<Δt<10<\Delta t<1 leads to

yn=yn−1+Δtzn,zn=zn−1+Δtσ1(W1yn−1+W1zn−1+V1un+b1)+Δtσ2(W2∣yn−1∣2+W2∣zn−1∣2+V2∣un∣2+b2).(3)\begin{aligned} y_n &= y_{n-1}+\Delta t z_n,\\ z_n &= z_{n-1}+\Delta t\sigma_1\left(\mathbf{W}_1y_{n-1}+\mathcal{W}_1z_{n-1}+\mathbf{V}_1u_n+b_1\right)\\ &\quad+\Delta t\sigma_2\left(\mathbf{W}_2\lvert y_{n-1}\rvert^2+\mathcal{W}_2\lvert z_{n-1}\rvert^2+\mathbf{V}_2\lvert u_n\rvert^2+b_2\right). \tag*{(3)} \end{aligned}

Finally, the output B^\hat{B} is computed for each recurrent unit, where B^∈Rn\hat{B}\in\mathbb{R}^{n} with B^=Qyn\hat{B}=Qy_n and Q∈Rn×mQ\in\mathbb{R}^{n\times m}.

Related work

Oscillator networks play a pervasive role across natural and engineering systems like pendulums in classical mechanics, among other instances. A notable trend is emerging where RNN architectures are constructed based on ODEs and dynamical systems [33, 34, 35, 23]. Our study is closely associated with CoRNN, where the oscillation and damping factors are integrated into model construction. In contrast, our approach incorporates hysteretic terms into the model. Another recent study [36], demonstrates employing neural oscillators to extend the applicability of physics-informed machine learning. Our work shares similarities with this study, as both works aim to generalize scientific machine learning and seek to predict quantities of interest beyond the scope of the training data without relying on retraining or

Test caseL2-norm (↓)Explained variance score (↑)Max error (↓)Mean absolute error (↓)
RNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNN
CFORC1C_{\mathrm{FORC}1}5.32341.23080.68630.01090.16250.35660.42300.98912.61341.49221.09940.23701.62530.68710.54330.0563
CFORC2C_{\mathrm{FORC}2}6.40350.95690.50980.01150.15760.38340.46930.99242.60841.29850.90550.25201.74840.57960.45870.0583
Cminor1C_{\mathrm{minor}1}7.59711.62950.80690.03200.14970.29770.32680.98442.37881.33440.94560.18821.49960.59740.43960.0916
Cminor2C_{\mathrm{minor}2}11.82932.17870.88590.12830.06160.27690.30910.92672.26691.17840.79250.29231.52430.56390.36530.1443

Table 2. The generalization performance assessed using the metrics: L2-norm relative error, explained variance error, maximum error, and mean absolute error for the second experiment, where max⁡(B)=1.25\max(B)=1.25 T.

Test caseL2-norm (↓)Explained variance score (↑)Max error (↓)Mean absolute error (↓)
RNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNN
CFORC1C_{\mathrm{FORC}1}6.09070.96920.66250.03420.05340.32900.36490.97053.21841.56611.21940.38781.99920.71050.62470.1296
CFORC2C_{\mathrm{FORC}2}7.26610.67200.47750.03770.02950.36910.41830.97853.21261.29720.94560.40552.16600.58060.52490.1371
Cminor1C_{\mathrm{minor}1}10.23051.60420.90090.0301-0.02160.26990.29090.97742.77031.32220.99470.18571.75200.59230.46190.0855
Cminor2C_{\mathrm{minor}2}18.25282.30691.04980.1580-0.16290.25280.27850.87802.56961.12490.80450.30661.79010.54460.36730.1491

Table 3. The generalization performance assessed using the metrics: L2-norm relative error, explained variance error, maximum error, and mean absolute error for the third experiment, where max⁡(B)=1.3\max(B)=1.3 T.

transfer learning methodologies. However, our work aims to predict trajectories of the hysteresis dynamics, whereas [36] predicts the solutions of partial differential equations in a generalized domain.

Motivation

The hidden state update in HystRNN is motivated by the differential models of hysteresis [8, 9, 28] that describe the phenomenon, incorporating an absolute value function to model the hysteretic nonlinearity. Examples of such phenomenological models include, but are not limited to, the Bouc-Wen and Duhem models, presented in SM §F. These absolute valued components play a crucial role in capturing hysteretic characteristics. These terms allow the models to account for the different responses during magnetization and demagnetization, as well as the effects of history on the behavior of the system. The inclusion of absolute valued terms enhances the ability of the model to capture the intricate dynamics of hysteresis and provides a more realistic representation of the observed phenomena.

Numerical Experiments

A series of numerical experiments encompassing four distinct scenarios is conducted, in which we systematically vary the upper limit of the magnetic field BB. The selection of diverse maximum BB field values corresponds to the specific usage context of the material. As a result, these experiments are geared towards demonstrating the viability of the proposed methodology across a spectrum of electrical machines, all constrained by their respective permissible maximum BB values.

Precisely, we opt for the maximum BB values of 1.7 T, 1.5 T, 1.3 T, and 1.25 T. In each of these instances, we execute a total of four experiments. After training for CmajorC_{\mathrm{major}} until the BB field reaches its saturation point, the experiments involve predicting CFORCC_{\mathrm{FORC}} and CminorC_{\mathrm{minor}}. Notably, the data set is generated from the Preisach model, which characterizes the behavior of non-oriented NO27 steel. It is imperative to preprocess the data before feeding it into the HystRNN model. We employ a normalization step using the min-max scaling technique, as elucidated in SM §C.

For all the numerical experiments, the software and hardware environments used for performing the experiments are as follows: Ubuntu 20.04.6 LTS, Python 3.9.7, NumPy 1.20.3, SciPy 1.7.1, Matplotlib 3.4.3, PyTorch 1.12.1, CUDA 11.7, and NVIDIA Driver 515.105.01, i7 CPU, and NVIDIA GeForce RTX 3080.

Test caseL2-norm (↓)Explained variance score (↑)Max error (↓)Mean absolute error (↓)
RNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNNRNNLSTMGRUHystRNN
CFORC1C_{\mathrm{FORC}1}0.76730.85560.93300.10070.29570.16750.34070.90171.79791.95721.89710.74910.91470.96420.99570.3258
CFORC2C_{\mathrm{FORC}2}0.49680.59420.52680.11640.33310.22090.37180.93301.22391.34921.30400.77250.70940.76860.72530.3304
Cminor1C_{\mathrm{minor}1}1.37191.37234.05500.07840.23180.09030.06400.92241.11121.17671.76880.23940.49810.50220.91770.1306
Cminor2C_{\mathrm{minor}2}1.79231.67216.07490.22570.22910.09070.01570.77430.95500.99881.58200.29910.43960.42610.9110.1717

Table 4. The generalization performance assessed using the metrics: L2-norm relative error, explained variance error, maximum error, and mean absolute error for the fourth experiment, where max⁡(B)=1.5\max(B)=1.5 T.

Baselines

We introduce the concept of employing neural oscillators constructed within the framework of recurrent networks for modeling hysteresis. This choice is driven by the intrinsic sequentiality and memory dependence characteristics of hysteresis. Additionally, recurrent architectures have been successfully employed for interpolation-type hysteresis modeling tasks across various fields [4, 17]. Consequently, motivated by these rationales, we conduct a comparative analysis of HystRNN with traditional recurrent networks such as RNN, LSTM, and GRU. We subject these models to a comprehensive comparison, specifically focusing on their performance in the context of generalization. Our experiments involve delving into the potential of these models by predicting outcomes for untrained HH sequences, thereby exploring their capabilities for generalization beyond trained data.

Hyperparameters

The selected hyperparameters consist of an input size of 2, a single hidden layer with a dimension of 32, and an output size of 1. The optimization process involves the utilization of the Adam optimizer, with a learning rate of 0.01. Training is conducted for 10000 epochs, with a batch size of 1. The hyperparameter Δt\Delta t is chosen to be 0.05. A sequence length of 595 is chosen for all four experiments to train CmajorC_{\mathrm{major}}. The determination of sequence length depends upon the data generated by the Preisach model for CmajorC_{\mathrm{major}}. Uniformity in hyperparameter settings is maintained across all experiments. Furthermore, to ensure fair comparisons with RNN, LSTM, and GRU models, the hyperparameters are also held constant for these methods.

Evaluation metrics

We evaluate the proposed method using four metrics. The first is the L2L2-norm, measuring the Euclidean distance between predicted and actual values. The explained variance score indicates prediction accuracy, capturing variance proportion. Maximum error detects significant prediction discrepancies as potential outliers. The mean absolute error assesses average differences between predictions and actual values for overall precision. Lower L2L2-norm, maximum error, and mean absolute error coupled with higher explained variance signify improved performance. Metric expressions are detailed in SM §D.

Train and test criteria

The trained model is tested in two distinct scenarios involving the prediction of two FORCs and two minor loops. For FORC prediction, testing sequences of lengths 199 and 399 are utilized, respectively. The prediction of minor loops involves a testing sequence with a length of 399 each. As in the case of training the model, these testing sequence lengths depend on the data generated from the Preisach model for evaluating the model. The HystRNN model trained on CmajorC_{\mathrm{major}} is evaluated on CFORCC_{\mathrm{FORC}} and CminorC_{\mathrm{minor}}. This testing sequence is initiated with an input (Hi,Bi)∈CFORC/minor(H_i, B_i) \in C_{\mathrm{FORC/minor}}, where both HiH_i, and BiB_i are provided and Bi+1B_{i+1} is predicted. The output generated from this step, Bi+1B_{i+1}, becomes the subsequent input along with Hi+1H_{i+1}, the known magnetization for the following sequence. Such testing holds paramount importance as practical scenarios lack prior knowledge about the BB values on CFORCC_{\mathrm{FORC}} or CminorC_{\mathrm{minor}}. Thus, the sole available information for generalization stems from the predicted solution in CFORCC_{\mathrm{FORC}} or CminorC_{\mathrm{minor}}.

Experimental Results

Four experiments are carried out to evaluate the performance of HystRNN. The experiments differ by the maximum permitted magnetic flux density of the electrical machine Bmax⁡B_{\max}. Exact Bmax⁡B_{\max} values are indicated for the experiments in SM §G. Performing experiments and exploring the generalization capabilities of the model for varying Bmax⁡B_{\max} values is crucial for understanding and optimizing the performance and efficiency of different and diverse machines. For instance, machines requiring lower magnetic flux densities of Bmax⁡=1.25B_{\max}=1.25 T are typically used as high-efficiency induction motors [37] in industrial settings for tasks like driving conveyor belts, pumps, and compressors. Meanwhile, high-performance applications, for instance, particle accelerators [38] and nuclear magnetic resonance [39], demand higher magnetic flux densities with Bmax⁡=1.7B_{\max}=1.7 T for their operation. We perform experiments for Bmax⁡B_{\max} in this spectrum and examine its performance for generalized scenarios, tailoring designs for diverse needs, ensuring energy efficiency for everyday devices, and pushing technological boundaries for cutting-edge systems.

For all the experiments, the data for the major loop, BmajorB_{\mathrm{major}}, is collected until max⁡(B)\max(B) reaches the saturation value Bmax⁡B_{\max}. HystRNN and other compared methods LSTM, GRU, and RNN are trained on BmajorB_{\mathrm{major}}. Once the model is trained, they are tested for four distinct cases. The first and second test cases correspond to estimating FORC. We denote them as CFORC1C_{\mathrm{FORC1}} and CFORC2C_{\mathrm{FORC2}} respectively. Two distinct FORCs are chosen to study the effect of the distance between the origin of the FORC and Bmax⁡B_{\max}. The third and fourth test cases are performed for predicting minor loops, which we denote by Cminor1C_{\mathrm{minor}_1} and Cminor2C_{\mathrm{minor}_2} respectively. These minor loops vary based on the maximum value of BB to which they are subjected. The origin of CFORC1C_{\mathrm{FORC}_1}, CFORC2C_{\mathrm{FORC}_2} and the maximum BB value of the minor loop is provided in SM §G.

Detailed performance metrics for HystRNN are outlined in Tables 1 to 4, corresponding to experiments 1 through 4, respectively. The Tables also facilitate a comprehensive comparative analysis with RNN, LSTM, and GRU. The metrics notably emphasize the superior performance of our proposed method HystRNN across all numerical experimentation scenarios.

Experiment 1

In Figure 3, the top two rows display the predictions of CFORC1,2C_{\mathrm{FORC}_{1,2}} respectively, wherein training exclusively occurs on CmajorC_{\mathrm{major}}, indicated by the blue color. Predictions are represented with black color, and ground truth is represented with red color. The colors are kept consistent for all the following experiments. The top two rows show that LSTM (Figure 3(a), 3(d)) and GRU (Figure 3(b), 3(e)) fail drastically to capture the shape of the FORC accurately. In contrast, HystRNN effectively captures the structure and symmetry of reversal curves as shown in Figure 3(c) and 3(f). The last two rows show the prediction for minor loop Cminor1,2C_{\mathrm{minor}_{1,2}}. For this case also predictions from LSTM (Figure 3(g), 3(j)) and GRU (Figure 3(h), 3(k)) are inaccurate. Neither LSTM nor GRU could form a closed loop for the predicted trajectory, posing a major challenge to compute the energy loss without a closed region, as energy loss depends on the surface area of the hysteresis loop. In contrast, our proposed method HystRNN predicts the structure of the loop very well and efficiently models the minor loop (Figure 3(i), 3(l)).

Figure 3. Experimental vs predicted hysteresis trajectories for experiment 1, where max⁡(B)=1.7\max(B) = 1.7 T. The blue curve represents the training loop CmajorC_{\mathrm{major}}. The red curve represents the ground truth for CFORC/minorC_{\mathrm{FORC/minor}} and the black curve represents the prediction of the model. Top two rows: predictions for CFORC1C_{\mathrm{FORC1}} and CFORC2C_{\mathrm{FORC2}} respectively. Bottom two rows: predictions for Cminor1C_{\mathrm{minor1}} and Cminor2C_{\mathrm{minor2}} respectively. The colors are used consistently for the following figures.

Experiment 2

The top two rows of Figure 4 present that LSTM (Figure 4a, 4(d)) and GRU (Figure 4b, 4(e)) fail to predict the trajectory of FORC by a huge margin. On the other hand, HysRNN shows close agreement with the ground truth for predicting CFORC1,2C_{\mathrm{FORC}_{1,2}} (Figure 4c, 4(f)). Also, the prediction of HystRNN for CFORC1C_{\mathrm{FORC}_1} is slightly better than for CFORC2C_{\mathrm{FORC}_2}, exemplifies that the model performs better when the origin of FORC is closer to max⁡(B)\max(B). A possible reason for this behavior could be the resemblance in the trajectories of CmajorC_{\mathrm{major}} and a FORC from a higher origin value. The last two rows of Figure 4 present the predictions of Cminor1,2C_{\mathrm{minor}_{1,2}} respectively. For this case, LSTM (Figure 4g, 4(j)) and GRU (Figure 4h, 4(k)) almost form a loop-like shape; however, they are very off from compared to the ground truth. HysRNN, on the other hand, captures the loop shape efficiently, as presented in Figure 4i and 4(l).

LSTM

(a) LSTM

GRU

(b) GRU

HystRNN

(c) HystRNN

LSTM

(d) LSTM

GRU

(e) GRU

HystRNN

(f) HystRNN

LSTM

(g) LSTM

GRU

(h) GRU

HystRNN

(i) HystRNN

LSTM

(j) LSTM

GRU

(k) GRU

HystRNN

(l) HystRNN

Figure 4. Experimental vs predicted hysteresis trajectories for experiment 2, where max⁡(B)=1.25\max(B) = 1.25 T. Top two rows: predictions for CFORC1C_{\mathrm{FORC}1} and CFORC2C_{\mathrm{FORC}2} respectively. Bottom two rows: predictions for Cminor1C_{\mathrm{minor}1} and Cminor2C_{\mathrm{minor}2} respectively.

Experiment 3

The predictions for the model, the ground truth, and the training data are presented in Figure 5. As presented in Figure 5(a), 5(b), 5(d), and 5(e) predictions by LSTM and GRU models exhibit a lack of accuracy. In contrast, predictions of our model HystRNN for the reversal curve are notably precise, as evidenced in Figure 5(c) and Figure 5(f). The final two rows within Figure 5 present that HystRNN accurately captured the characteristics of the minor loop, as showcased in Figure 5(i) and 5(l). Prediction by GRU manages to capture a resemblance of the loop, although not entirely, as revealed in Figure 5(h) and 5(k). On the other hand, LSTM performs poorly, failing to capture the intricate structure of the minor loop, as depicted in Figure 5(g) and 5(j).

Figure 5. Experimental vs predicted hysteresis trajectories for experiment 3, where max⁡(B)=1.3\max(B)=1.3 T. Top two rows: predictions for CFORC1C_{\mathrm{FORC1}} and CFORC2C_{\mathrm{FORC2}} respectively. Bottom two rows: predictions for Cminor1C_{\mathrm{minor1}} and Cminor2C_{\mathrm{minor2}} respectively.

Experiment 4

The predictions for the model, the ground truth, and the training data are presented in Figure 6. Predictions of the reversal curve show agreement with the nature observed in previous experiments. In this case, max⁡(B)\max(B), origin of CFORC2C_{\mathrm{FORC}_2}, and maximum BB value of Cminor2C_{\mathrm{minor}_2} vary significantly, posing a challenge for both LSTM and GRU. However, HystRNN outperforms them for each case, as shown in Figure 6. The results underscore the performance of HystRNN as, for neither of the cases, the accuracy of LSTM or GRU is comparable to our proposed method. Additional visual results for all the RNN experiments are supplemented in SM §E.

LSTM hysteresis trajectory for experiment 4, top row

(a) LSTM

GRU hysteresis trajectory for experiment 4, top row

(b) GRU

HystRNN hysteresis trajectory for experiment 4, top row

(c) HystRNN

LSTM hysteresis trajectory for experiment 4, second row

(d) LSTM

GRU hysteresis trajectory for experiment 4, second row

(e) GRU

HystRNN hysteresis trajectory for experiment 4, second row

(f) HystRNN

LSTM hysteresis trajectory for experiment 4, third row

(g) LSTM

GRU hysteresis trajectory for experiment 4, third row

(h) GRU

HystRNN hysteresis trajectory for experiment 4, third row

(i) HystRNN

LSTM hysteresis trajectory for experiment 4, bottom row

(j) LSTM

GRU hysteresis trajectory for experiment 4, bottom row

(k) GRU

HystRNN hysteresis trajectory for experiment 4, bottom row

(l) HystRNN

Figure 6. Experimental vs predicted hysteresis trajectories for experiment 4, where max⁡(B)=1.5\max(B) = 1.5 T. Top two rows: predictions for CFORC1C_{\mathrm{FORC1}} and CFORC2C_{\mathrm{FORC2}} respectively. Bottom two rows: predictions for Cminor1C_{\mathrm{minor1}} and Cminor2C_{\mathrm{minor2}} respectively.

Conclusions

We introduced a novel neural oscillator, HystRNN, aimed to advance magnetic hysteresis modeling within extrapolated regions. The proposed oscillator is based upon the foundation of coupled oscillator recurrent neural networks and inspired from phenomenological hysteresis models. HystRNN was validated by predicting first-order reversal curves and minor loops after training the model solely with major loop data. The outcomes underscore the superiority of HystRNN in adeptly capturing intricate nonlinear dynamics, outperforming conventional recurrent neural architectures such as RNN, LSTM, and GRU on various metrics. This performance is attributed to its capacity to assimilate sequential information, history dependencies, and hysteretic features, ultimately achieving generalization capabilities. Access to the codes and data will be provided upon publication.

Supplementary Material

Nomenclature

The table provided below presents the abbreviations utilized within this paper.

SYMBOLDESCRIPTION
CORNNCOUPLED-OSCILLATORY RECURRENT NEURAL NETWORK
GRUGATED RECURRENT UNIT
LSTMLONG SHORT-TERM MEMORY
ODEORDINARY DIFFERENTIAL EQUATION
RNNRECURRENT NEURAL NETWORK

Table 5. Abbreviations used in this paper

Preisach model of hysteresis and data generation for magnetic experiments

The modeling of magnetic materials is a critical step when designing electromechanical devices, mostly to predict the behavior of the ferromagnetic core. Often, this is achieved by a so-called anhysteretic curve that describes the ideal magnetic behavior of a soft-magnetic material which is not subjected to any hysteresis effects. However, for increased insight it is desired to include this hysteresis behavior into the simulation.

One of the most widely known models to describe the hysteresis relation is the Preisach model of hysteresis. It is a phenomenological model that describes the hysteresis effect by a set of hysteresis operators, scattered on a triangular domain called the Preisach plane, which are scaled by a weight function, following

f(t)=∬α≥βP(α,β)γ^α,βu(t)dαdβ(4)f(t) = \iint_{\alpha\ge\beta} P(\alpha,\beta)\hat{\gamma}_{\alpha,\beta}u(t)d\alpha d\beta \tag*{(4)}

where:

  • uu is the input,

  • ff is the output,

  • tt is a moment in time,

  • α\alpha & β\beta are two switching variables,

  • PP is the Preisach weight function,

  • γ^\hat{\gamma} represents the hysteresis operators.

Typically, the weight function is constructed from well-known probability density functions, for which the parameters are tuned to closely match the magnetic material behavior. Alternatively, the weight function can also be obtained from magnetic measurement data, which generally results in the best correspondence, albeit at the cost of a more complicated identification procedure. Furthermore, in its general form, the model output is not affected by the rate of change of the input, and the model describes solely the static scalar hysteresis behavior. However, various generalization exist to include, e.g. dynamical eddy currents effects, or the vector hysteresis phenomenon. Ultimately, in this work the static scalar Preisach model was applied, with a weight function fitted on a set of measurement data.

The measurement data in this work was obtained by a Brockhaus MPG 200 soft-magnetic material tester, using an Epstein frame calibrated to correspond with the IEC standard. A set of concentric hysteresis loops up to a maximum of 1.7 T was measured for NO27-1450H, obtained under quasi-DC excitation. Here, quasi-DC indicates that the rate of change of the magnetic flux density was controlled such that any eddy current fields are negligible, and the static hysteresis behavior is obtained. The Epstein sample strips used were obtained using spark eroding and cut in the rolling direction.

Data Normalization

The procedure used to normalize the data for training all the models is described in the following subsections.

Min-Max Scaling

Min-Max Scaling is a technique used to transform numerical data so that it falls within a specific range, typically between 0 and 1. The purpose of this normalization is to bring all the features to a similar scale, which can be particularly important for algorithms that are sensitive to the scale of input data.

Xmin⁡X_{\min} represents the minimum value in the original dataset XX.

Xmax⁡X_{\max} represents the maximum value in the original dataset XX.

XscaledX_{\mathrm{scaled}} is the result of scaling the original dataset.

The formula for scaling the data is:

Xscaled=X−Xmin⁡Xmax⁡−Xmin⁡×2−1X_{\mathrm{scaled}} = \frac{X-X_{\min}}{X_{\max}-X_{\min}} \times2 - 1

In this formula:

  • (X−Xmin⁡)(X-X_{\min}) calculates the difference between each data point and the minimum value.

  • (Xmax⁡−Xmin⁡)(X_{\max}-X_{\min}) represents the range of the data.

  • Dividing by (Xmax⁡−Xmin⁡)(X_{\max}-X_{\min}) scales the data to the range [0,1][0,1].

  • Multiplying by 2 and subtracting 1 scales the data to the range [−1,1][-1,1].

So, XscaledX_{\mathrm{scaled}} contains the normalized data in the range of [−1,1][-1,1].

Inverse Min-Max Scaling

After Min-Max Scaling to revert the scaled data back to its original values. The formula for this Inverse Min-max scaling is given as:

Xoriginal=Xscaled+12×(Xmax⁡−Xmin⁡)+Xmin⁡X_{\mathrm{original}} = \frac{X_{\mathrm{scaled}}+1}{2} \times(X_{\max}-X_{\min}) + X_{\min}

In this formula:

  • Xscaled+12\frac{X_{\mathrm{scaled}}+1}{2} scales the scaled data back to the range [0,1][0,1].

  • (Xmax⁡−Xmin⁡)(X_{\max}-X_{\min}) represents the original range of the data.

  • Multiplying by (Xmax⁡−Xmin⁡)(X_{\max}-X_{\min}) scales the data back to the original range.

  • Adding Xmin⁡X_{\min} adjusts the scaled data to its original position.

Error Metrics

The error metrics used in this study are described in the following subsections.

L2L^2 norm

The formula for the relative L2L^2 norm of B^\hat{B} with respect to ground truth BB is given by:

Relative L2 norm=∥B^−B∥2∥B∥2\text{Relative } L^2 \text{ norm} = \frac{\lVert\hat{B}-B\rVert_2}{\lVert B\rVert_2}

where:

  • ∥B^−B∥2\lVert\hat{B}-B\rVert_2 is the Euclidean distance between vectors B^\hat{B} and BB,

  • ∥B∥2\lVert B\rVert_2 is the Euclidean norm (magnitude) of vector BB.

Explained variance score

The formula for the explained variance score is given by:

Explained Variance Score=1−∑i=1n(Bi−B^i)2∑i=1n(Bi−Bˉ)2\text{Explained Variance Score} = 1 - \frac{\sum_{i=1}^{n}(B_i-\hat{B}_i)^2}{\sum_{i=1}^{n}(B_i-\bar{B})^2}

where:

  • nn is the number of data points,

  • BiB_i represents the ground truth at the ii-th data point,

  • B^i\hat{B}_i represents the predicted value at the ii-th data point,

  • Bˉ\bar{B} represents the mean of the ground truth.

Max error

The formula for the maximum absolute error is given by:

Max Absolute Error=max⁡i=1n∣Bi−B^i∣\text{Max Absolute Error} = \max_{i=1}^{n} \lvert B_i - \hat{B}_i \rvert

where:

  • nn is the number of data points,

  • BiB_i represents the ground truth at the ii-th data point,

  • B^i\hat{B}_i represents the predicted value at the ii-th data point,

  • ∣x∣\lvert x \rvert represents the absolute value of xx.

Mean absolute error

The formula for the mean absolute error is given by:

Mean Absolute Error=1n∑i=1n∣Bi−B^i∣\text{Mean Absolute Error} = \frac{1}{n} \sum_{i=1}^{n} \lvert B_i - \hat{B}_i \rvert

where:

  • nn is the number of data points,

  • BiB_i represents the ground truth at the ii-th data point,

  • B^i\hat{B}_i represents the predicted value at the ii-th data point,

  • ∣x∣\lvert x \rvert represents the absolute value of xx.

Additional results for RNN

Additional results for RNN, for experiment 1 through 4 are provided in Figure 8 to Figure 11, respectively.

Figure 8. Exp 1: RNN prediction for (a:) CFORC1C_{\mathrm{FORC}1}; (b:) CFORC2C_{\mathrm{FORC}2}; (c:) Cminor1C_{\mathrm{minor}1}; (d:) Cminor2C_{\mathrm{minor}2}

Figure 11. Exp 4: RNN prediction for (a:) CFORC1\mathcal{C}_{\mathrm{FORC}1}; (b:) CFORC2\mathcal{C}_{\mathrm{FORC}2}; (c:) Cminor1\mathcal{C}_{\mathrm{minor}1}; (d:) Cminor2\mathcal{C}_{\mathrm{minor}2}

Phenomenological hysteresis differential models

The Duhem hysteresis model is given by,

B˙(t)=a∣H˙(t)∣⋅[bH(t)−B(t)]+cH˙(t)\dot{B}(t) = a \lvert\dot{H}(t) \rvert\cdot[bH(t) - B(t)] + c\dot{H}(t)

where, a,ba,b, and cc are the parameters that control the shape of the hysteresis loop. The dot over a quantity denotes the time derivative of the quantity.

The Bouc-Wen model for hysteresis is given by,

B˙(t)=αH˙−β∣H˙(t)∣⋅B(t)⋅∣B(t)∣n−1−γH˙∣B(t)∣n\dot{B}(t) = \alpha\dot{H} - \beta\lvert\dot{H}(t)\rvert\cdot B(t) \cdot\lvert B(t)\rvert^{n-1} - \gamma\dot{H}\lvert B(t)\rvert^n

where, α\alpha, β\beta, and γ\gamma are Bouc-Wen parameters that control the shape of the hysteresis loop.

Details regarding the conducted experiments

Experiment 1

For the first experiment, max⁡(B)=1.7\max(B) = 1.7 T. The origin of CFORC1,2C_{\mathrm{FORC}_{1,2}} is taken to be 1.25 T and 0.5 T respectively. The chosen maximum BB value of Cminor1,2C_{\mathrm{minor}_{1,2}} is 1.25 T and 1.1 T. The predictions for the model, the ground truth, and the training data are presented in Figure 3.

Experiment 2

For the second experiment, max⁡(B)=1.25\max(B) = 1.25 T. The origin of CFORC1,2C_{\mathrm{FORC}_{1,2}} is taken to be 1.2 T and 1.0 T respectively. The chosen maximum BB value of Cminor1,2C_{\mathrm{minor}_{1,2}} is 1.0 T and 0.8 T. The predictions for the model, the ground truth, and the training data are presented in Figure 4.

Experiment 3

For the third experiment, max⁡(B)=1.3\max(B) = 1.3 T. The origin of CFORC1,2C_{\mathrm{FORC}_{1,2}} is taken to be 1.0 T and 0.75 T respectively. The chosen maximum BB value of Cminor1,2C_{\mathrm{minor}_{1,2}} is 1.0 T and 0.75 T.

Experiment 4

For the fourth experiment, max⁡(B)=1.5\max(B) = 1.5 T. The origin of CFORC1,2C_{\mathrm{FORC}_{1,2}} is taken to be 1.25 T and 0.75 T respectively. The chosen maximum BB value of Cminor1,2C_{\mathrm{minor}_{1,2}} is 0.9 T and 0.7 T.

The MPG 200 desk by Brockhaus, with Epstein frame and sample strips shown on the right.

Figure 7. The MPG 200 desk by Brockhaus, with Epstein frame and sample strips shown on the right.

Figure 9. Exp 2: RNN prediction for (a:) CFORC1C_{\mathrm{FORC}1}; (b:) CFORC2C_{\mathrm{FORC}2}; (c:) Cminor1C_{\mathrm{minor}1}; (d:) Cminor2C_{\mathrm{minor}2}

Figure 10. Exp 3: RNN prediction for (a:) CFORC1C_{\mathrm{FORC}1}; (b:) CFORC2C_{\mathrm{FORC}2}; (c:) Cminor1C_{\mathrm{minor}1}; (d:) Cminor2C_{\mathrm{minor}2}

References

  1. [1]Giorgio Bertotti and Isaak D Mayergoyz. The Science of Hysteresis: 3-volume set. Elsevier, 2005.
  2. [2]Konstantinos Vlachas, Konstantinos Tatsis, Konstantinos Agathos, Adam R Brink, and Eleni Chatzi. A local basis approximation approach for nonlinear parametric model order reduction. Journal of Sound and Vibration, 502:116055, 2021.DOI
  3. [3]Doğa Ceylan, Reza Zeinali, Bram Daniels, Konstantin O Boynov, and Elena A Lomonova. A novel modeling technique via coupled magnetic equivalent circuit with vector hysteresis characteristics of laminated steels. IEEE Transactions on Industry Applications, 59(2):1481–1491, 2022.DOI
  4. [4]Guangzeng Chen, Guangkae Chen, and Yunjiang Lou. Diagonal recurrent neural network-based hysteresis modeling. IEEE Transactions on Neural Networks and Learning Systems, 33(12):7502–7512, 2021.DOI
  5. [5]Giorgio Bertotti. Hysteresis in magnetism: for physicists, materials scientists, and engineers. Gulf Professional Publishing, 1998.DOI
  6. [6]Ferenc Preisach. Über die magnetische nachwirkung. Zeitschrift für physik, 94(5-6):277–302, 1935.
  7. [7]David C Jiles and David L Atherton. Theory of ferromagnetic hysteresis. Journal of magnetism and magnetic materials, 61(1-2):48–60, 1986.
  8. [8]R Bouc. Forced vibrations of mechanical systems with hysteresis. In Proc. of the Fourth Conference on Nonlinear Oscillations, Prague, 1967, 1967.
  9. [9]Yi-Kwei Wen. Method for random vibration of hysteretic systems. Journal of the engineering mechanics division, 102(2):249–263, 1976.
  10. [10]Abhishek Chandra, Bram Daniels, Mitrofan Curti, Koen Tiels, Elena A Lomonova, and Daniel M Tartakovsky. Discovery of sparse hysteresis models for piezoelectric materials. Applied Physics Letters, 122(21), 2023.DOI
  11. [11]Miaomiao Lin, Changming Cheng, GuanZhen Zhang, Baoxuan Zhao, Zhike Peng, and Gu ang Meng. Identification of bouc-wen hysteretic systems based on a joint optimization approach. Mechanical Systems and Signal Processing, 180:109404, 2022.
  12. [12]Claudio Serpico and Ciro Visone. Magnetic hysteresis modeling via feed-forward neural networks. IEEE Transactions on Magnetics, 34(3):623–628, 1998.DOI
  13. [13]Dimitre Makaveev, Luc Dupré, Marc De Wulf, and Jan Melkebeek. Modeling of quasistatic magnetic hysteresis with feed-forward neural networks. Journal of Applied Physics, 89(1):6737–6739, 2001.DOI
  14. [14]Simone Quondam-Antonio, Francesco Riganti-Fulgeini, Antonino Laudani, Gabriele-Maria Lozito, and Riccardo Scorretti. Deep neural networks for the efficient simulation of macro-scale hysteresis processes with generic excitation waveforms. Engineering Applications of Artificial Intelligence, 121:105940, 2023.
  15. [15]M Farrokh, FS Dizaji, and MS Dizaji. Hysteresis identification using extended preisach neural network. Neural Processing Letters, pages 1–25, 2022.
  16. [16]Chih-Lyang Hwang and Chau Jan. Recurrent-neural-network-based multivariable adaptive control for a class of nonlinear dynamic systems with time-varying delay. IEEE transactions on neural networks and learning systems, 27(2):388–401, 2015.
  17. [17]Mahmoud Saghafifar, Anjan Kundu, and Andrew Nafalski. Dynamic magnetic hysteresis modelling using elman recurrent neural network. International Journal of Applied Electromagnetics and Mechanics, 13(1-4):209–214, 2002.
  18. [18]Kyunghyun Cho, Bart Van Merriënboer, Caglar Gulcehre, Dzmitry Bahdanau, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. Learning phrase representations using rnn encoder-decoder for statistical machine translation. arXiv preprint arXiv:1406.1078, 2014.
  19. [19]Sepp Hochreiter and Jürgen Schmidhuber. Long short-term memory. Neural computation, 9(8):1735–1780, 1997.pubmed.ncbi.nlm.nih.gov/9377276
  20. [20]Diego Serrano, Haoran Li, Shukai Wang, Thomas Guildo, Min Luo, Vineet Bansal, Niraj K Jha, Yuxin Chen, Charles R Sullivan, and Minjie Chen. Why magnet: Quantifying the complexity of modeling power magnetic material characteristics. IEEE Transactions on Power Electronics, 2023.
  21. [21]Steven L Brunton and J Nathan Kutz. Data-driven science and engineering: Machine learning, dynamical systems, and control. Cambridge University Press, 2022.
  22. [22]T Konstantin Rusch and Siddhartha Mishra. Coupled oscillatory recurrent neural network (cornn): An accurate and (gradient) stable architecture for learning long time dependencies. arXiv preprint arXiv:2010.00951, 2020.arxiv.org/abs/2010.00951
  23. [23]T Konstantin Rusch and Siddhartha Mishra. Unicornn: A recurrent model for learning very long time dependencies. In International Conference on Machine Learning, pages 9168–9178. PMLR, 2021.arxiv.org/abs/2103.05487
  24. [24]T Konstantin Rusch, Siddhartha Mishra, N Benjamin Erichson, and Michael W Mahoney. Long expressive memory for sequence modeling. arXiv preprint arXiv:2110.04744, 2021.arxiv.org/abs/2110.04744
  25. [25]Alejandro Queiruga, N Benjamin Erichson, Liam Hodgkinson, and Michael W Mahoney. Stateful ode-nets using basis function expansions. Advances in Neural Information Processing Systems, 34:21770–21781, 2021.arxiv.org/abs/2106.10820
  26. [26]T Konstantin Rusch, Ben Chamberlain, James Rowbottom, Siddhartha Mishra, and Michael Bronstein. Graph-coupled oscillator networks. In International Conference on Machine Learning, pages 18888–18909. PMLR, 2022.arxiv.org/abs/2202.02296
  27. [27]Samuel Lanthaler, T Konstantin Rusch, and Siddhartha Mishra. Neural oscillators are universal. arXiv preprint arXiv:2305.08753, 2023.arxiv.org/abs/2305.08753
  28. [28]JinHyoung Oh and Dennis S Bernstein. Semilinear duhem model for rate-independent and rate-dependent hysteresis. IEEE Transactions on Automatic Control, 50(5):631–645, 2005.DOI
  29. [29]Arbaaz Khan, Saikou Ceesay, Yuan-Po Teng, Ruoli Wang, Haipeng Yue, and David Lowther. Generalizable deep neural network based multi-material hysteresis modeling. In 2022 IEEE 20th Biennial Conference on Electromagnetic Field Computation-Long papers (CEFC-LONG), pages 1–4. IEEE, 2022.DOI
  30. [30]Mohammad Reza Zamani Kouhpanji and Bethanie JH Stadler. First-order reversal curve (forc) measurements for decoding mixtures of magnetic nanowires. Magnetic Measurement Techniques for Materials Characterization, pages 651–663, 2021.DOI
  31. [31]Dustin Gilbert. Forc diagrams in magnetic thin films. Magnetic Measurement Techniques for Materials Characterization, pages 629–650, 2021.DOI
  32. [32]Alexandru Stancu. Characterization of magnetic nanostructures with the first-order reversal curves (forc) diagram technique. Magnetic Measurement Techniques for Materials Characterization, pages 605–628, 2021.DOI
  33. [33]Ricky TQ Chen, Yulia Rubanova, Jesse Bettencourt, and David K Duvenaud. Neural ordinary differential equations. Advances in neural information processing systems, 31, 2018.arxiv.org/abs/1806.07366
  34. [34]Yulia Rubanova, Ricky TQ Chen, and David K Duvenaud. Latent ordinary differential equations for irregularly-sampled time series. Advances in neural information processing systems, 32, 2019.
  35. [35]Bo Chang, Minmin Chen, Eldad Haber, and Ed H Chi. Antisymmetricrnn: A dynamical system view on recurrent neural networks. arXiv preprint arXiv:1902.09689, 2019.arxiv.org/abs/1902.09689
  36. [36]Taniya Kapoor, Abhishek Chandra, Daniel M Tartakovsky, Hongrui Wang, Alfredo Nunez, and Rolf Dollevoet. Neural oscillators for generalization of physics-informed machine learning. arXiv preprint arXiv:2308.08989, 2023.arxiv.org/abs/2308.08989
  37. [37]Aina He, Anding Wang, Shiqiang Yue, Chengliang Zhao, Chuntao Chang, He Men, Xinmin Wang, and Run-Wei Li. Dynamic magnetic characteristics of fe78si3b9 amorphous alloy subjected to operating temperature. Journal of Magnetism and Magnetic Materials, 408:159–163, 2016.
  38. [38]Ingars Reinholds, Valdis Kalkis, and Roberts D Maksimovs. The effect of ionizing radiation and magnetic field on deformation properties of high density polyethylene/acrylonitrile-butadiene composites. Journal of Chemistry and Chemical Engineering, 6(3), 2012.
  39. [39]Yukie Mori, Yoshio Sakaguchi, and Hisaharu Hayashi. Magnetic field effects on chemical reactions of biradical radical ion pairs in homogeneous fluid solvents. The Journal of Physical Chemistry A, 104(21):4896–4905, 2000.

Paper details

vlnc:W_wtpxv9ut
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents