Predicting Magnetic Losses with AI: A Physics-Informed Approach

by

Author: Denis Avetisyan

A novel deep learning architecture leverages spectral entropy and feature fusion to dramatically improve the accuracy and robustness of magnetic core loss prediction.

The SEPI-TFPNet model architecture processes 1,024-point time-series data through hierarchical feature extraction and integrates it with scalar physical features via an adaptive multimodal fusion module, guided by a customized loss function that constrains both an empirical model and a data-driven model informed by spectral-entropy discrimination and an explicit physics-prior feature channel, ultimately predicting core loss through a three-layer MLP.
The SEPI-TFPNet model architecture processes 1,024-point time-series data through hierarchical feature extraction and integrates it with scalar physical features via an adaptive multimodal fusion module, guided by a customized loss function that constrains both an empirical model and a data-driven model informed by spectral-entropy discrimination and an explicit physics-prior feature channel, ultimately predicting core loss through a three-layer MLP.

This review details SEPI-TFPNet, a hybrid model integrating physics-informed priors with Bi-LSTM networks and attention mechanisms for precise magnetic core loss analysis.

Accurate prediction of magnetic core loss remains a critical challenge in the design of efficient power electronic systems, despite limitations in traditional modeling approaches. This paper introduces a novel hybrid model, termed ‘Spectral entropy prior-guided deep feature fusion architecture for magnetic core loss,’ which integrates physics-informed empirical models with deep learning techniques-specifically, Bi-LSTM networks, attention mechanisms, and adaptive feature fusion. By leveraging spectral entropy to guide the selection of appropriate empirical priors, the proposed method achieves improved accuracy and robustness in predicting core loss across diverse magnetic materials and excitation waveforms. Could this approach unlock new possibilities for data-driven design and optimization of power electronic components?

The Limits of Conventional Core Loss Modeling

Efficient power converter design hinges on the accurate prediction of core loss, as these losses directly impact overall system performance and energy efficiency. However, traditional core loss prediction methods often fall short when confronted with the complex, non-sinusoidal waveforms prevalent in modern power electronics. These methods, frequently relying on simplifying assumptions about the magnetic flux density, struggle to account for the higher-order harmonics and distortion present in these waveforms. Consequently, designers face challenges in accurately estimating losses, potentially leading to suboptimal component selection, reduced efficiency, and even thermal management issues. The discrepancy between predicted and actual losses underscores the need for more sophisticated modeling techniques capable of handling the intricacies of real-world operating conditions and waveform complexities.

The Steinmetz equation, a cornerstone of core loss prediction for over a century, relies on the assumption of sinusoidal flux density. While offering a computationally efficient method for estimating losses in magnetically driven systems, its applicability drastically decreases when faced with the non-sinusoidal waveforms prevalent in modern power electronics. These real-world signals, rich in harmonics and characterized by rapid switching, introduce distortions that violate the equation’s fundamental premises. Consequently, the Steinmetz equation tends to underestimate losses arising from higher-order harmonics and the rate of flux change – factors that significantly contribute to overall energy dissipation. This limitation underscores the need for more sophisticated models capable of accurately capturing the complex interplay between waveform shape and core loss, particularly as power converters increasingly adopt advanced modulation techniques.

Empirical models for predicting core loss frequently demonstrate limited adaptability when applied beyond the specific datasets used for their development. These models, often derived from curve fitting to experimental results, struggle to account for the nuanced interplay between waveform shape, frequency content, and the resulting energy dissipation within the magnetic core. Consequently, they may provide reasonably accurate predictions under familiar operating conditions, but exhibit substantial errors when faced with novel or complex waveforms-such as those encountered in modern power electronic converters employing pulse-width modulation or multi-level techniques. The intricate relationship between harmonic distortion, peak values, and the rate of change of current or flux density, all contribute to core loss in ways that simple empirical formulas often fail to adequately capture, hindering their generalization capability and necessitating more sophisticated predictive approaches.

The increasing demands placed on modern power converters – higher switching frequencies, complex waveforms, and wider operating ranges – reveal the limitations of established core loss prediction techniques. A truly effective solution must move beyond simplified assumptions and embrace adaptability. This requires a methodology capable of accurately characterizing loss mechanisms under varied excitation conditions, including those present in wide-bandgap semiconductor applications and multi-phase systems. The development of such a robust predictor hinges on advanced modeling approaches, potentially incorporating machine learning or physics-informed algorithms, to capture the nuanced relationship between waveform attributes – like peak current, duty cycle, and harmonic content – and the resulting energy dissipation within the magnetic core. Ultimately, a versatile and precise core loss prediction tool will enable designers to optimize converter performance, minimize energy waste, and enhance system reliability across a spectrum of operational scenarios.

SEPI-TFPNet: A Synergistic Physics-Informed Deep Learning Approach

SEPI-TFPNet utilizes a hybrid approach by integrating the improved generalized Steinmetz equation (iGSE) – a physics-based model describing core loss in magnetic materials – with a deep learning network. The iGSE, expressed as $P = k \cdot f^n \cdot B^m$, provides a foundational understanding of hysteresis loss based on frequency ($f$) and magnetic flux density ($B$). By incorporating this physical prior knowledge into the network architecture, SEPI-TFPNet reduces the reliance on large datasets and improves generalization capability. The deep learning component learns to refine the iGSE parameters and capture complex nonlinearities not explicitly represented in the equation, effectively bridging the gap between physics-based modeling and data-driven prediction of core loss.

An autoencoder is implemented within SEPI-TFPNet to process the high-dimensional input data representing magnetic core characteristics. This neural network architecture learns a compressed, lower-dimensional representation of the input data, effectively performing dimensionality reduction. By minimizing the reconstruction error between the input and the autoencoder’s output, relevant features are extracted while discarding noise and redundant information. This feature extraction process not only reduces computational complexity and improves model training speed, but also enhances generalization performance by focusing on the most salient aspects of the magnetic core data, ultimately increasing the overall efficiency of the core loss prediction model.

Rotation Embedding is a positional encoding technique applied to the time-series data representing magnetic core waveforms. This method calculates the angle of each data point as if originating from the center of the waveform, effectively representing the position of each sample within the cycle. The sine and cosine of this angle are then appended as features to the original data, providing the model with explicit positional information. This encoding is crucial because the dynamics of magnetic hysteresis – and therefore core loss – are dependent on the direction and sequence of magnetization. By incorporating rotational position, the model can more effectively differentiate between waveforms with similar magnitudes but differing directions, improving its ability to accurately predict core loss characteristics and generalize to unseen data.

The SEPI-TFPNet framework demonstrates improved core loss prediction accuracy and generalization capability when benchmarked against conventional methods such as finite element analysis and empirical models. This enhancement stems from the model’s ability to integrate physics-based constraints, derived from the improved generalized Steinmetz equation (iGSE), with data-driven learning. Traditional methods often struggle with complex magnetic behaviors and require extensive computational resources or material-specific calibration. In contrast, SEPI-TFPNet’s hybrid architecture allows it to extrapolate to unseen operating conditions and core geometries with reduced reliance on labeled training data, yielding more robust and reliable loss estimations across a wider range of magnetic cores and excitation waveforms.

Deep Learning Architecture: Capturing Temporal Dynamics with Precision

Convolutional Neural Networks (CNNs) process waveform data by applying filters to generate feature maps, effectively detecting localized patterns across different time segments. These filters, learned during training, convolve with the input signal, producing activations that represent the presence of specific features. Deeper layers within the CNN then operate on these feature maps, extracting increasingly complex and abstract temporal information. This hierarchical feature extraction allows the network to identify subtle, time-dependent characteristics within the waveform without requiring explicit time-frequency analysis, and captures patterns irrespective of their exact position in time. The resulting deep feature maps serve as input for subsequent layers that model longer-range temporal dependencies.

A Bidirectional Long Short-Term Memory (Bi-LSTM) network addresses the limitations of traditional Recurrent Neural Networks (RNNs) by processing sequence data in both forward and reverse directions. This bi-directional approach allows the model to capture contextual information from both past and future time steps, which is crucial for understanding long-range dependencies within the waveform data. Standard LSTMs process sequences unidirectionally, limiting their ability to utilize future context. The Bi-LSTM achieves this by employing two separate LSTM networks: one processing the sequence from beginning to end, and another processing it from end to beginning. The outputs of these two LSTMs are then combined, providing the model with a more comprehensive understanding of each time step’s context within the entire sequence. This is particularly valuable for analyzing waveforms where patterns may extend over considerable durations and where the future can inform the interpretation of the past.

An attention mechanism is implemented to dynamically weigh the contribution of each time step within the input waveform sequence to the loss prediction. This is achieved by calculating attention weights based on the hidden states generated by the Bi-LSTM layer; these weights, typically normalized using a softmax function, represent the relevance of each time step. The weighted sum of these hidden states then forms a context vector, which is concatenated with the Bi-LSTM output and fed into the subsequent Multi-Layer Perceptron for core loss regression. This allows the model to prioritize specific portions of the waveform – for example, transient events or key features – thereby improving the accuracy of loss prediction by focusing computational resources on the most informative data points.

The extracted features from the CNN, Bi-LSTM, and Attention Mechanism are fed into a Multi-Layer Perceptron (MLP) for core loss regression. This MLP serves as the final processing layer, mapping the high-level feature representation to a predicted core loss value. Model training utilizes a custom Double Mean Absolute Percentage Error (DMAPE) loss function, calculated as $DMAPE = \frac{1}{N}\sum_{i=1}^{N} \frac{|y_i – \hat{y}_i|}{(0.5 * (|y_i| + |\hat{y}_i|))}$, where $y_i$ represents the actual core loss and $\hat{y}_i$ is the predicted core loss for the $i$-th sample. The DMAPE function was selected to minimize the impact of outliers and to provide a percentage-based error metric, facilitating performance evaluation across varying core loss magnitudes.

A bi-directional long short-term memory (Bi-LSTM) network utilizes sequential data processing with both forward and backward passes to capture contextual information.
A bi-directional long short-term memory (Bi-LSTM) network utilizes sequential data processing with both forward and backward passes to capture contextual information.

Validation and Demonstrably Superior Performance

The development of SEPI-TFPNet relied heavily on the MagNet Dataset, a meticulously curated and expansive database of magnetic material properties. This dataset served as the foundational training ground, exposing the model to a diverse range of material compositions and characteristics crucial for accurate prediction. Rigorous validation against this same dataset ensured the model’s reliability and generalization capability, confirming its ability to accurately predict magnetic behavior beyond the training examples. The comprehensiveness of MagNet, encompassing a wide spectrum of materials and operating conditions, was instrumental in establishing a robust and dependable foundation for SEPI-TFPNet’s performance, ultimately enabling its superior predictive power in power converter design applications.

Rigorous testing reveals that SEPI-TFPNet consistently achieves higher accuracy and reliability than existing state-of-the-art models in magnetic materials prediction. When benchmarked against prominent architectures like New Paderborn, PI-MFF-CN, and New EMPINN, SEPI-TFPNet demonstrates a marked improvement in performance across a range of magnetic materials datasets. This consistent outperformance isn’t merely incremental; it signifies a substantial leap forward in predictive capability, potentially revolutionizing the design and optimization of power conversion systems by enabling more accurate modeling of core magnetic components. The model’s ability to consistently surpass established benchmarks underscores its robustness and suggests a new standard for performance in this critical area of materials science and engineering.

SEPI-TFPNet distinguishes itself through a synergistic approach, seamlessly integrating physics-informed modeling with the capabilities of deep learning. This fusion allows the network to not merely learn from data, but to understand and adhere to the underlying physical principles governing magnetic fields-crucially improving prediction accuracy. Quantitative analysis reveals a substantial performance gain; specifically, SEPI-TFPNet achieves a 36.3% reduction in 95% relative error when benchmarked against the New Paderborn model. This improvement isn’t simply incremental; it signifies a considerable leap in the precision of magnetic field predictions, with potential ramifications for optimizing the design of power converters and minimizing energy dissipation within these systems. The model’s ability to extrapolate beyond the training data, guided by established physical laws, contributes to this enhanced reliability and predictive power.

SEPI-TFPNet exhibits a substantial advancement in predictive accuracy, achieving an 86.4% reduction in 95% relative error compared to the New EMPINN model. This heightened precision translates directly into benefits for power converter design, a field where even minor improvements can yield significant gains. By more accurately modeling the complex physical processes within these converters, SEPI-TFPNet facilitates designs that minimize energy dissipation and maximize operational efficiency. Consequently, implementation of this model promises a pathway toward reduced energy losses, contributing to both cost savings and a smaller environmental footprint in a wide range of electronic applications, from portable devices to large-scale power grids.

The pursuit of accurate magnetic core loss prediction, as detailed in this work, echoes a fundamental tenet of systems design: structure dictates behavior. SEPI-TFPNet, with its fusion of physics-informed priors and deep learning architectures, isn’t simply layering technologies, but crafting a cohesive structure to elicit predictable, robust outcomes. As Donald Davies observed, “If the system survives on duct tape, it’s probably overengineered.” This model eschews unnecessary complexity, instead focusing on the elegant integration of established physical principles with the power of deep learning – a testament to simplicity driving effective performance. The attention mechanism, specifically, acts as a refined structural element, directing focus to the most pertinent features, preventing the model from becoming burdened by irrelevant data.

The Road Ahead

The pursuit of accurate magnetic core loss prediction, as demonstrated by this work, invariably reveals the limitations inherent in any attempt to fully capture a complex physical phenomenon with a finite model. SEPI-TFPNet offers a compelling synthesis of physics-informed priors and deep learning, yet the very architecture hints at persistent challenges. The chosen feature fusion approach, while effective, relies on specific representations learned by the Bi-LSTM network. A more elegant solution would likely emerge from a system capable of discovering relevant features directly from raw data, minimizing the need for pre-defined, potentially brittle, representations.

Future efforts should focus on exploring architectures that prioritize inherent simplicity. If a design feels clever, it’s probably fragile. The current reliance on attention mechanisms, while boosting performance, introduces complexity that may hinder long-term robustness. The true test lies not merely in achieving high accuracy on existing datasets, but in generalizing to unseen operating conditions and material compositions-a task demanding a deeper understanding of the underlying physics, not just skillful pattern recognition.

Ultimately, this field will progress not through increasingly intricate models, but through a return to fundamental principles. A truly robust predictive framework will be one that gracefully handles uncertainty, acknowledges the limits of its knowledge, and prioritizes clarity over complexity. The structure dictates behavior, and a simple, well-understood structure is always preferable to a convoluted, opaque one.

Original article: https://arxiv.org/pdf/2512.11334.pdf

Contact the author: https://www.linkedin.com/in/avetisyan/

See also:

2025-12-15 16:07