David Alister Lowther

The state-of-the-art design processes for electrical machines (eMachine) are template-based methods and design from scratch approaches. Over the past four decades, significant advancements have been made to make these methods effective in terms of electromagnetic performance. However, the state-of-the-art cannot provide information on the topology and the components of an eMachine from a given non-parametrized computer aided design (CAD). This information is imperative for the automatic selection of an appropriate simulation profile to meet the required key performance indicators (KPIs), enhancing the numerical computation of electromagnetic fields, and improving the user experience through an improved user interface. Moreover, it can be used to generate a parametric version of a given CAD design, which can subsequently be utilized by an optimization tool for geometric refinement and optimal material selection. The first step to achieve this is recognizing the components and their relative special positioning. This paper uses a Mask Region-based Convolutional Neural Network (Mask RCNN), an object detection technique, to identify the components of an eMachine from a CAD design.

This paper investigates the use of multi-fidelity datasets combined with transfer learning to accelerate the design of electric machines (E-machines). By reusing knowledge from prior tasks, transfer learning reduces the cost and time of creating datasets and training models. A transfer-learning-enhanced conditional GAN (TLE-CDC-GAN) is employed, trained on multi-fidelity simulation data to generate E-machine designs conditioned on torque specifications. The results show that transfer learning enhances efficiency, generalizes well across tasks, and works with smaller and low-fidelity datasets. This provides a scalable and cost-effective framework for solving inverse design problems in E-machine development, giving earlier insights for designers and avoiding the frustration of starting over if the design goals change.

The state-of-the-art design processes for electrical machines (eMachine) are template-based methods and design from scratch approaches. Over the past four decades, significant advancements have been made to make these methods effective in terms of electromagnetic performance. However, the state-of-the-art cannot provide information on the topology and the components of an eMachine from a given non-parametrized computer aided design (CAD). This information is imperative for the automatic selection of an appropriate simulation profile to meet the required KPIs, enhancing the numerical computation of electromagnetic fields, and improving the user experience through an improved user interface. Moreover, it can be used to generate a parametric version of a given CAD design, which can subsequently be utilized by an optimization tool for geometric refinement and optimal material selection. The first step to achieve this is recognizing the components and their relative special positioning. This paper uses Mask RCNN, an object detection technique, to identify the components of an eMachine from a CAD design.

The authors explore the possibility of applying a convolutional Naeural Network (CNN) to the solution of coupled electromagnetic and thermal problem, focusing on the classical problem of induction heating systems, traditionally solved by resorting to Finite Element (FE) models. In fact, FE modelling is widely used in the design of induction heating systems due its accuracy, even if the solution of a coupled nonlinear problem is expensive in terms of computational time and hardware resources, notably in 3D analysis. A model based on CNN could be an interesting alternative; in fact, CNN is a learning model selected for its excellent ability to converge, even when trained with a limited dataset. CNNs are able to treat images as input and they are used here as follows: given a temperature map in the workpiece, identify the corresponding vector of current, frequency and process heating time; this mapping is a model of the inverse induction heating problem. Specifically, we consider as an example the induction heating of a cylindrical steel billet, made of C45 steel, placed in a solenoidal inductor coil exhibiting the same axial length of the billet (TEAM 36 problem). A thorough heating process is usually applied before hot working of the billet, as in an extrusion process, but this methodology can be applied also in the design of induction hardening processes. First, a CNN has been trained from scratch by means of a dataset of FE solutions of coupled electromagnetic and thermal problems. For the sake of a comparison, a transfer learning technique is applied using GoogLeNet, i.e. a Deep Convolutional Neural Network able to classify images: starting from the pre-trained GoogLeNet, its training has been subsequently refined with the dataset of solutions from FE analyses. When the training dataset contains a limited number of samples, GoogleNet shows good accuracy in predicting the process parameters; in the case of a high number of samples in the training set, namely beyond a threshold like e.g. 1500, both CNNs show good accuracy of the result.

3-D analysis is a crucial tool for designing and optimizing electrical machines, especially those with significant 3-D effects such as an axial flux permanent magnet (AFPM) machine. Building high-fidelity surrogate models of such topologies requires 3-D FEA. However, where the computational budget is limited for large or medium-scale parametric machine optimization, replacing 3-D FEA with models that confidently and quickly compute the 3-D performance can be extremely helpful. To resolve the 3-D issues, this paper attempts to build effective high-speed reduced order models with an accuracy approaching a full 3-D model. A hybrid modeling approach is presented that couples the 2½-D modeling approaches with simple and effective magnetic equivalent circuits (MEC) models incorporating the 3-D effects. The proposed approach provides a satisfactory compromise regarding computation time and accuracy.

Accurately modeling magnetic hysteresis plays a crucial role in developing precise digital twins for low-frequency electromagnetic systems. However, in large three-dimensional analysis systems, the evaluation of hysteresis performance at hundreds of thousands of points in components containing magnetic steels is challenging. It is of utmost importance that any modeling system can assess the hysteresis performance within the shortest possible timeframe. The utilization of Neural Networks (NNs) offers the potential to achieve this objective. This paper provides a comprehensive review of various NN architectures that can be employed to address this requirement. Two types of learning approaches are explored: completely data-driven approaches and hybrid approaches that combine data and hysteresis laws. A comparative study is conducted to analyze the predictive power and computational cost of the different architectures under investigation.

Physics-informed neural networks (PINNs) have been successfully applied in electromagnetism for the solution of direct problems. However, since PINNs typically do not take system parameters (like geometry or material properties) as input, when embedded in inverse problems or adopted for parametrical studies, in order to output the solution of the governing equations, they require additional training for each new system parameter set. To overcome this issue, we propose a hypernetwork (HNN) that receives system parameters and outputs the network weights of a PINN, which in turn provides the solution of the direct problem. Therefore, once trained, the HNN acts as a parametrized real-time field solver that allows the fast solution of inverse problems, in which the objective(s) are defined a-posteriori (i.e., after HNN’s training). This method is adopted for a coil optimal design task in magnetostatics.