Researchers are grappling with the challenge of accurately modeling the complex magnetic behavior of synchronous machines. This is an important step towards improving the control and efficiency of synchronous machines. Junyi Li, Tim Foißner and Floran Martin from Aalto University, together with Antti Piippo from ABB Oy and Marko Hinkkanen from Aalto University, present a new physics-based neural network approach that directly incorporates gradient networks into the fundamental equations governing magnetic fields. This collaboration represents a significant advance over traditional methods such as lookup tables, requiring less training data while ensuring physically realistic and reliable model extrapolation. The resulting model has been validated using both measured and finite element datasets from a 5.6 kW permanent magnet synchronous reluctance machine and promises to enable robust model inversion and optimal trajectory generation for advanced control applications.
New modeling techniques for electric motors promise more efficient and reliable performance. This advancement addresses the long-standing challenge of accurately simulating the complex magnetic behavior within these machines, potentially enabling improvements in everything from industrial automation to electric vehicle performance. By integrating physics-based neural networks, researchers have achieved unprecedented accuracy and efficiency in modeling complex magnetic behavior within electrical machines.
This effort addresses long-standing challenges in electric drive technology, creating dynamic models that accurately capture magnetic saturation and spatial harmonics essential for optimal performance and control. The team’s innovation lies in an architecture that embeds the fundamental equations governing electromagnetic fields directly into the neural network itself, ensuring inherent physical consistency.
By modeling the magnetic field energy gradient, the system automatically satisfies the principle of energy balance, which is a key requirement for reliable operation and robust control strategies. This new modeling framework goes beyond the limitations of traditional methods such as lookup tables and standard machine learning techniques. These traditional approaches often require large training datasets, struggle to extrapolate beyond known data, and can produce output that lacks smoothness.
In contrast, the proposed architecture requires significantly less training data, guarantees monotonicity, a property that ensures predictable behavior, and produces very smooth output. These properties unlock the potential for robust model inversion and optimal trajectory generation, essential for advanced control applications in electric drives.
At the heart of this advance is the use of gradient networks, a special type of neural network that can universally approximate physically valid magnetic behavior. Unlike previous Hamiltonian neural networks that rely on numerical differentiation, this approach directly models conservative vector fields, eliminating key sources of error and improving computational efficiency.
By directly learning the relationships between magnetic flux, current, and rotor angle, the model accurately represents the complex interactions of forces within the machine. Validation using both measured data from a 5.6 kW permanent magnet synchronous reluctance machine and finite element method simulations demonstrated the accuracy and physical consistency of the model, even when trained with limited data.
This research paves the way to more efficient and reliable electric drives, enabling improved performance in a wide range of applications, from electric vehicles to industrial automation. The ability to accurately model complex magnetic behavior with minimal data opens new possibilities for real-time control, predictive maintenance, and the design of next-generation electrical machines.
Energy conservation and reversible relationships through monotonic gradient networks
The proposed physics-based neural network accurately models a 5.6 kW permanent magnet synchronous reluctance machine and demonstrates physically consistent behavior even with limited training data. Importantly, the model inherently satisfies energy balance through the direct implementation of the gradient network, ensuring that the reciprocity condition is met without any additional constraints.
This is achieved by modeling the stator current and electromagnetic torque as gradients of a scalar field energy function, which is a fundamental aspect of the study. This architecture universally approximates any physically realizable magnetic behavior, offering significant advantages over traditional look-up tables and standard machine learning approaches. A monotonic gradient network was adopted to ensure the convexity of the field energy and established a unique and reversible relationship between the flux linkage and the current.
This allows the formulation of both current and flux linkage maps, increasing the versatility and robustness of the model. Incorporating Fourier functionality successfully captures spatial harmonics while maintaining a lossless and conservative structure of the field, improving accuracy in complex scenarios. In this study, we also investigated activation functions and proposed computationally efficient p-norm gradient activation as an alternative to the more commonly used softmax functions.
The model’s ability to accurately represent magnetic saturation and angular dependence is a significant achievement and addresses important challenges in high-fidelity mechanical modeling. By directly modeling conservative vector fields, this work avoids the need for numerical differentiation in scalar neural networks, improving both data efficiency and generalization ability. This approach enables robust model inversion and optimal trajectory generation, which are essential components for advanced control applications in electric drives.
Physically consistent modeling with monotone gradient networks and Fourier features
Gradient networks underpin this research, directly modeling conservative vector fields to accurately represent the electromagnetic behavior of synchronous machines. Rather than relying on traditional neural networks or lookup tables, this study implements an architecture in which the stator current and electromagnetic torque are defined as the gradient of a scalar field energy function.
This innovative approach inherently guarantees physical consistency, especially energy balance and reciprocity, by design. A monotonic gradient network is employed to ensure the convexity of the field energy, establishing a unique and reversible relationship between flux linkage and current, which is important for both forward and reverse mapping. Fourier functions are integrated into the network to capture the complexity of real-world machines.
These features allow the model to represent spatial harmonics in the magnetic field while preserving the lossless and conservative structure of the energy field. This is a significant methodological advance as it avoids the limitations of analytic functions and the memory demands of high-dimensional lookup tables. The chosen architecture universally approximates all physically realizable magnetic behavior and provides a robust alternative to traditional modeling techniques.
Validation included utilizing both measured datasets and datasets generated by finite element method (FEM) simulations from a 5.6 kW permanent magnet synchronous reluctance machine. This dual approach ensured model accuracy across a variety of operating conditions and provided a benchmark against established simulation techniques. This methodology prioritizes data efficiency and requires less training data than traditional machine learning approaches while maintaining a high degree of accuracy and physical validity.
Physics-based neural networks enhance modeling of electrical machine performance and reliability
Scientists have long struggled to create accurate digital twins of complex physical systems, and this research represents an important step forward in that effort. The challenge is not to reproduce behavior, but to reproduce it in a way that respects the underlying physical laws. Traditional machine learning models are great at pattern recognition, but they can easily produce unrealistic output, especially when extrapolating beyond training data.
This new approach employs physics-based neural networks to embed these physical constraints directly into the model itself. The impact goes beyond simply improving simulation fidelity. Accurate and efficient modeling is essential to optimize the performance of electrical machines such as synchronous reluctance motors, leading to increased energy efficiency and reduced operating costs.
Importantly, the ability to reliably extrapolate enables predictive maintenance and the design of more robust control systems, potentially preventing failures before they occur. Reducing the need for large-scale training data is also a major benefit, lowering the barrier to entry for applying these techniques to new and diverse systems. However, this is not a universal solution.
Current validation focuses on specific motor sizes and operating conditions. Extending this approach to significantly larger or more complex machines will undoubtedly introduce new challenges. Furthermore, although the model enforces energy balance, it does not address all possible physical constraints, and the choice of which constraints to include remains an important design decision.
In the future, we can expect to see more widespread integration of physics-based machine learning across various engineering disciplines. The next frontier lies in the development of automated methods that identify and incorporate relevant physical laws, potentially creating self-aware models that can adapt to unexpected situations. This will help us seamlessly blend the power of data-driven learning with the elegance of fundamental physics, unlocking a new era of intelligent systems.
