Conference Agenda

In recent years, the method of transforming Maxwell equations into circuit models with physical values is proposed to solve the three-dimensional electromagnetic problems. However, these existing circuit models greatly simplify the absorbing boundary conditions, so that they cannot accurately simulate the wave propagation problems. In this regard, a circuit model (or admittance transfer function model) based on improved surface impedance absorbing boundary conditions(SIABC) is proposed to solve general electromagnetic problems. Representative numerical examples are used to validate the approach. The results are compared with those of the existing circuit model and the classical finite difference time domain (FDTD) model, showing the proposed model can not only guarantee the accuracy of the solution but also improves the speed.

Simulation of complex electronic based systems frequently require the coupling of a lumped circuit with an electromagnetic field simulations based on the finite element method. This work presents a flexible and natural way to couple the finite element method with lumped circuits in a weak sense with the aid of side constraints based on the magnetic vector potential formulation. A small and simple numerical example shows the feasibility and accuracy of the proposed method.

Recently, deep neural networks (DNNs) have been attracting attention in the field of scientific computation, which are also known as physics informed neural networks.

The feed-forward DNN can be thought of as an approximator of continuous functions for a given activation function. The advent of ReLU activation can broaden the potential of DNN to scientific computations. However, this standard ReLU is not sufficient to solve the elliptic differential equations such as electromagnetic potential equations. Since we need more regularity of the DNN model equipped with ReLU, at least up to second-order derivatives.

The solution for this is the higher-order ReLU functions, which ensure as much regularity as we want. In this paper, we propose a DNN model with these higher-order ReLUs as an activation and investigate the accuracy compared to a sigmoidal activation. Based on this superior approximation property of these higher-order ReLU DNN models, we will accurately solve elliptic differential equations.

In this paper, an improved Enhanced Nodal-Order Reduction (IENOR) methodology is proposed and applied to the model based on the circuit-netlists model to reduce the order of equation sets of three-dimensional (3D) electromagnetic (EM) fields, providing a fast solution approach for 3D EM problems. The proposed reduction-method-based approach is validated by solving the EM field problem of a reflective metasurface cell. The order of this original circuit-netlists model is 122344, and the order of the IENOR reduced model is 66. Moreover, according to the results comparisons of the Radar Cross Section(RCS) for the IENOR model and circuit-netlists model, the IENOR approach not only greatly improves the speed but also guarantees the accuracy of the solution.

We introduce a novel methodology that achieves complementary energy bounds for the ohmic losses and magnetic energy for

an eddy current problem. We devise a novel post-processing fast equilibration technique that provide both bilateral bounds in one

stroke, hence requiring the solution of just one problem formulation. Numerical results confirm the effectiveness of the construction.

Recently a high (polynomial) order extension of the Cell Method has been introduced by the present authors for the numerical solution of the Maxwell equations in the time domain. The basis functions originally chosen are nevertheless non-optimal: although leading to an explicit time stepping method which also preserves fundamental properties of the equations (electromagnetic energy conservation), the basis is badly conditioned as the polynomial degree is increased (a well known property of monomial bases). We herein propose alternative bases which optimise the conditioning through partial orthonormalization in the 2D setting.

It is presented an atomistic micromagnetic model, and this model was used for the calculation of Terbium domain wall. It was developed a computer code for solving the equations, which minimize the exchange, magnetocrystalline and magnetoelastic energy terms. Antiferromagnetism need to be considered in micromagnetic models. The negative terms due to antiferromagnetism must be considered in the Hamiltonian describing the exchange energy term. The exchange energy cannot be truncated at the first term. If negative terms are neglected, the whole result is based on a false minimum.

This paper presents an approach to estimate the multidomain wall motion obtained on video from one grain of a polycrystalline material. The first step consists of an image reprocessing in order to reduce noisy artifacts. Then, an object detection technique is applied to the preprocessed images to locate the position on the domain wall. The location data collected from each image is then used to calculate the motion of the domain walls on all input videos. Although this research is still in progress, we have already achieved encouraging results.

This study demonstrated that the mechanical deformation of the hydrogel block in an electric field should be calculated using an electric force density according to the Lorentz–Kelvin formulation, via a comparison between the numerical simulation and experimental results. The conceptual error in conventional electric force density based on the Korteweg–Helmholtz formulation, which is commonly used in electric-fluid coupled simulation model, was revealed and corrected. This correction established homogeneity in the two electric force density formulations that were otherwise known to have different distributions.

Model of indirect induction baking of thin electrically non-conductive layers (paints, lacquers, resins etc.) is presented together with a novel algorithm for controling the process based on the digital twin. This technique is much more flexible than the standard feedback form of control and can take into account temperatures at more points both on the surface and in the volume of the covering layer. Recurrent Neural Network (RNNs) proved to be a powerful tool for solving the mathematical model online and provide the input data for sufficiently fast switching of the current source. The methodology is illustrated with an example.

The multiscale finite element method (MSFEM) allows for the solution of the eddy current problem on stratified cores without having to resolve each steel sheet in the finite element mesh. This drastically reduces the computational complexity while still giving satisfying results. This paper presents an a-posteriori error estimator, which reliably estimates the error with respect to the physical solution. The estimator is based on flux equilibration and introduces no generic constants. While a similar estimator has been presented for the scalar T-formulation, this work extends the applicability to the vector valued A-formulation, using both linear and nonlinear materials. Numerical examples show the estimator to be both reliable and efficient.

In this paper, we propose a new semi-homogenized approach for the modeling of the inductive and capacitive effects in multi-turn windings developed upon the Darwin formulation in the frequency domain. To allow an easy circuit coupling, a relaxed interaction between the electric and magnetic fields is assumed. As a consequence, the formulation splits into two subsequent subproblems: a circuit-coupled magnetodynamic problem and an electrodynamic problem. In such case, the proposed approach considers a semi-homogenized treatment of the windings wherein a fully homogenized winding window is used in the magnetodynamic subproblem; while in its electrodynamic counterpart, the representation of the conductors in the FE geometry is required. A numerical example is considered to validate the results of the proposed formulation.

Left ventricular assist devices (LVADs) have been used to treat patients with heart failure. Despite the clinical success, designing a long–term implantable LVAD remains a challenge. We proposed a hemocompatibility assessment platform (HAP) to identify blood trauma contribution from individual LVAD components. A HAP would help in refining pump components through iterative testing before in vivo testing, thereby preventing unnecessary animal sacrifice and reducing development time and cost. So the HAP does not confound data, it uses an enlarged gap drive motor coupled to a magnetic levitation (maglev) bearing. Characterization of the radial force generated by rotor-stator eccentricity in the drive motor is essential for maglev development. Since most experimental test systems have physical limitations, a method that can overcome those limitations to provide accurate results is valuable. In this study, we established a design method wherein empirical static forces are used to derive a statistical model to predict static force generated by motor eccentricity. The statistical model was then validated numerically.

An incomplete Balancing Domain Decomposition (BDD) method is considered as the preconditioner of an iterative Domain Decomposition Method (DDM) for perturbed magnetostatic problems, where the magnetic vector potential is regarded as an unknown function approximated by the Nedelec curl conforming finite element. To reduce the number of the Degrees Of Freedom (DOF) of coarse spaces in the incomplete BDD method, Virtual Element Methods (VEMs) and Polynomial Element Methods (PEMs) are introduced. Owing to the introduction of VEMs and PEMs and the result of BDD method originally proposed by Mandel, the condition number of coefficient matrices derived from the iterative DDM is evaluated. Therefore, the number of iterations of the iterative DDM can be kept even when the number of the subdomains becomes larger. Moreover, the approximate coarse space by VEMs and PEMs can admit that each subdomain becomes more general polyhedron.

We propose a machine-learning-based workflow to solve the well-known spectral estimation problem of time signals from EM simulations. In time domain, we define a simple yet powerful activation function for a three-layer radial basis function network that is used to fit the signal. Due to the right choice of the activation, we are able to extrapolate the data and, therefore, compute the prediction of the original signal for a long simulation time. This property of our approach enables the calculation of important key quantities of the signal’s spectrum such as resonance frequencies and Q-values with a low error. In order to test our method, we perform the spectral estimation of the response of a highly resonant filter.