RNN Cuts Cost of FEM Simulations with Hysteresis and Eddy Currents

Learn how a recurrent neural network reduces FEM simulation costs for laminated cores with hysteresis and eddy currents. Open-source surrogate model inside.

lunes, 20 de julio de 2026 • 6 min read • Q2BSTUDIO Team

Machine learning para electromagnetismo computacional

The global energy transition and the massive electrification of sectors such as mobility, aerospace and renewables have placed electromagnetic components at the center of industrial innovation. Traction motors, wind generators and distribution transformers demand laminated ferromagnetic cores that operate with minimum losses and maximum power density. In this context, numerical simulation using the finite element method (FEM) has become the essential tool for predicting behavior before manufacturing prototypes. However, when the goal is to rigorously capture phenomena such as magnetic hysteresis and eddy currents in each sheet of the stack, computational requirements grow exponentially, turning each design iteration into a lengthy process.

The challenge lies in the duality between physical accuracy and temporal viability. Classical models that solve the electromagnetic field inside each lamination, point by point and at every nonlinear integration step, generate huge system matrices that can increase calculation time by extreme factors compared to an approximation that only considers saturation. For an engineering department that must deliver optimizations under tight deadlines, this asymmetry is unsustainable. The answer does not always lie in acquiring more computing power, but in rethinking the architecture of simulation software, adopting hybrid approaches where the development of custom software plays a strategic role in connecting legacy calculation engines with new intelligent optimization layers.

From a business perspective, having custom software adapted to internal R&D workflows allows manufacturers to overcome the limitations of generic CAE suites. At Q2BSTUDIO we understand that every electrical machine manufacturer has its own methodologies, differentiated material libraries and specific validation criteria. Therefore, we accompany our clients in modernizing their simulation environments through solutions that integrate traditional FEM engines with artificial intelligence components, ensuring that the technological transition respects previous investments and enhances analytical capabilities without operational friction.

Within the spectrum of machine learning techniques, recurrent neural networks (RNNs) stand out as natural candidates for modeling the response of magnetic materials. Unlike standard dense networks, RNNs maintain an internal state that evolves with each new input, giving them contextual memory. This property is fundamental when addressing hysteresis, a phenomenon where the current magnetization state depends on the entire previous excitation history. A well-designed recurrent architecture can learn to predict the resulting flux density or field intensity in the laminated core from temporal sequences, eliminating the need to solve partial differential equations in the domain of each individual sheet.

The practical integration of this approach transforms FEM calculation dynamics. Instead of diving into the microphysics of each lamination at every Gauss point of the domain, the simulator queries the neural model as if it were an external material routine. The RNN, previously trained with thousands of synthetically generated magnetic field histories that emulate real operating conditions —frequency variations, amplitude, waveform and temperature—, returns the constitutive response in fractions of a millisecond. The result is a marginal increase in overall computational cost, comparable to a simulation that neglects hysteresis, while retaining fidelity close to the full reference model. This efficiency opens the door to massive parametric studies and topological optimizations that were previously unfeasible.

The robustness of the method rests on the diversity of the training set. The broader the spectrum of magnetic conditions simulated during the learning phase, the greater the surrogate's generalization capability. A model trained with sequences covering everything from steady-state sinusoidal regimes to complex transients with harmonics can be applied indistinctly in 2D magnetodynamic FEM simulations of permanent magnet motors, reluctance generators or power transformers, without requiring specific retraining for each geometry. This versatility reduces project preparation time and standardizes result quality across the entire product portfolio of a company.

Deploying these capabilities in an industrial environment demands an appropriate technological infrastructure. Training deep recurrent architectures and subsequently serving them as real-time inference components benefits enormously from cloud AWS/Azure environments, where GPU-accelerated compute clusters scale automatically according to demand. In parallel, the use of managed artificial intelligence platforms allows monitoring model drift, scheduling periodic retraining with experimental data from real tests, and maintaining a continuous improvement cycle that closes the loop between simulation, manufacturing and validation.

The benefits of accelerating FEM calculations are not limited to the analysis department. When simulation times are drastically reduced, data on losses, heating, efficiency and torque density can be channeled into BI/Power BI systems that offer real-time executive visibility. Product managers and technical directors have access to interactive dashboards where they compare design scenarios, identify thermal bottlenecks and make informed decisions about materials and geometries without relying on static reports that take weeks to prepare. Simulation ceases to be a technical silo and becomes a strategic business asset.

In this ecosystem of digital models and interconnected data flows, cybersecurity acquires relevance comparable to numerical accuracy. Trained surrogate models encapsulate proprietary knowledge about alloys, heat treatments and differential electromagnetic topologies. Their exposure through internal APIs or storage in shared repositories demands encryption policies, multi-factor authentication and role-based access control. Any breach in the confidentiality of these intangible assets could compromise years of R&D, so the software architecture must incorporate security layers from its design, not as a subsequent add-on.

Looking toward the horizon, the next maturity level is represented by AI agents capable of orchestrating autonomous design flows. These systems do not merely query a neural model to obtain a material response; they can automatically iterate on geometric variables, select sheet thicknesses, adjust stacking and evaluate thermal and mechanical constraints as they progress. By combining the speed of the RNN surrogate with Bayesian or evolutionary optimization algorithms, agents explore the design space continuously, proposing configurations that the human engineer might not have considered. This paradigm transforms the engineering center into a predictive innovation unit, where human creativity is enhanced by systematic machine exploration.

Adopting these technologies does not necessarily imply a disruptive revolution of existing systems. The neural model can be encapsulated as an independent library, with standard interfaces that couple to established FEM frameworks without rewriting the simulator core. This custom software and microservices architecture philosophy is precisely what we promote at Q2BSTUDIO: evolving our clients' technology stacks through clean integrations, well-documented APIs and decoupled components that facilitate long-term maintenance. Thus, innovation in materials and numerical methods advances at the pace of business, not the other way around.

In conclusion, the convergence between classical numerical methods and artificial intelligence models is redefining the boundaries of advanced electromagnetic design. Recurrent neural networks do not replace physical understanding of the problem; they amplify it, filtering microscopic complexity to deliver high-fidelity results in times compatible with modern industry. For organizations that bet on this path, the differential will reside not only in isolated hardware or algorithms, but in the ability to integrate everything within a coherent digital strategy. With the support of technology partners specialized in software development, cloud infrastructure and data governance, the road to next-generation predictive simulation is already open.

OUR SERVICES

How we can help you

Do you have a project in mind?

Tell us your vision and we'll turn it into a software solution. Whatever the scope, we make your idea real.