RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation

arXiv:2603.29745 · eess.SY, cond-mat.mtrl-sci, cs.SY, eess.SP · Submitted 2026-03-31 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Today's paper: "RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation".

Dev: Based on a Pareto investigation, a rather black-box gated recurrent unit (GRU) model structure with a graceful initialization setup was found to offer the most attractive model size vs.

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: So we're looking at the RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation paper today. It seems like they're focusing on making predictions for transient magnetic fields in ferrite materials, which is pretty important for optimizing components.

Dev: Yeah, I saw the abstract mentions they compared different model architectures to see what works best for this kind of prediction task. It sounds like a really practical problem because we need accurate time-resolved and temperature-aware H-field predictions for things that aren't running under steady excitation.

Taro: I'm curious if these predictions translate well when the system encounters unexpected situations, like when the environment misbehaves or the input signal deviates significantly from what they were trained on.

Rosa: Exactly, Taro; that's a big deal for autonomy applications where things aren't always perfectly controlled.

Dev: The paper is quite focused on finding a good trade-off between model size and accuracy, which is something we need to consider when we think about deploying these kinds of models in real-world hardware.

Rosa: What they are proposing seems to be a specific architecture, a GRU model with a graceful initialization setup, which they found was quite attractive for small models in this field.

Dev: That three hundred twenty-five-parameter GRU structure is what caught our attention; parameter efficiency is key when you're dealing with real-time systems where computational resources might be limited.

Taro: It makes sense that they’d prioritize a small model if it still gives us reasonable performance, especially since the other physics-inspired models they tested performed worse.

Rosa: That comparison is telling; the GRU structure seems to have a specific advantage in this regime, even when compared to those models inspired by physical principles.

Dev: The results they reported are quite encouraging too; for that three hundred twenty-five-parameter GRU model, they achieved an average sequence relative error of eight point zero two percent and an average normalized energy relative error of one point zero seven percent across five different materials on unseen test data.

Taro: Eighty percent for the sequence error sounds pretty decent, but I wonder how stable that performance holds if we introduce a completely new material type that wasn't in the training set.

Rosa: That's where the generalization question comes up; they tested it across five materials, which suggests some level of robustness, but we need to know how broad that range really is for different applications.

Dev: The paper also details their feature engineering process, which involves normalizing raw magnetic field and temperature values by finding the maximum absolute value for each material's training set—Hmax, Bmax, and ϑmax—and then dividing by that value.

Title and authors: Taro: That normalization step sounds like a necessary first step to make sure the input scales are consistent before feeding them into the recurrent structure.

Rosa: It’s about standardizing the inputs so the model doesn't get overwhelmed by differences in raw data magnitudes across materials, which is a common headache in experimental work.

Dev: The training cost function they adapted is an RMS error loss, LaRMSE, which incorporates both tracking error on H and pointwise errors on B weighted by the change in B. They then normalize this loss using the RMS value of the full sequence H0:k3 to get a weighted loss L'aRMSE for backpropagation.

Taro: Integrating those two types of error into one cost function shows they were thinking about both tracking accuracy and energy considerations simultaneously, which is smart for a system that needs to be efficient.

Rosa: It’s interesting how they balanced the sequence tracking with the energy aspect in that loss formulation; it suggests they were trying to capture a more complete physical picture of the magnetic behavior during excitation.

Dev: The GRU-P architecture itself includes a warmup process where the first hidden state is created by concatenating the first normalized field value with zeros, and then the input sequence is fed sequentially into the GRU cell using that initial state.

Taro: That warmup mechanism seems designed to get the model's starting point correct before it starts making predictions, which should help stabilize those early outputs when dealing with dynamic excitation.

Rosa: It sounds like a clever way to handle the initial conditions of the recurrence without just starting from scratch, which is something I’ve seen in other time-series modeling.

Dev: Then for the actual H-trajectory estimation, they use a featurized input sequence Xk1:k2 fed sequentially into another part of the GRU structure, where in each iteration, the first element of the hidden state vector is used as the normalized prediction for H.

Taro: Looping back on that first element in every iteration means it's continuously refining its prediction based on its own previous output, which is a classic recursive approach.

Rosa: So they're essentially using the model to predict the next step based on what it already predicted, which is very powerful for trajectory estimation.

Dev: The implementation uses the JAXthree Python library to handle things like GPU and TPU utilization and just-in-time compilation, which speaks directly to their need for efficient execution on hardware.

Title and authors: Taro: That’s crucial for achieving the kind of real-time inference we talked about earlier; you can’t run complex models if the loop rate is too slow or latency is too high.

Rosa: It really shows they thought about the deployment side, not just the theoretical modeling part, which I appreciate seeing in a paper like this on RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation.

Dev: Looking ahead at improvements suggested in this work, one key suggestion is to focus on improving core loss prediction accuracy for transient magnetic fields by deploying the GRU-P architecture with that specific warmup mechanism.

Taro: That makes sense; if we can nail the transient field prediction, we solve a big part of the problem for things like electric motor drives or power factor correction applications.

Rosa: And it opens up possibilities for much more accurate design of magnetic components because you can estimate core losses with higher fidelity than before.

Dev: Another area they push is enabling time-resolved and temperature-aware H-field prediction for magnetic components operating under non-stationary excitation waveforms, like those seen in electric motor drives.

Taro: That’s exactly where the real challenge lies; dealing with things that aren't just simple sine waves requires a model that can adapt dynamically to the input changes.

Rosa: If we can handle those non-stationary conditions reliably, it means these models could be useful in environments where the excitation itself is constantly shifting.

Dev: They also look at achieving high parameter efficiency while maintaining excellent prediction accuracy on unseen test data, specifically targeting low Sequence Relative Error and Normalized Energy Relative Error scores.

Taro: The three hundred twenty-five-parameter result already showed good efficiency, but pushing those error metrics lower would really prove its utility for demanding applications where precision is paramount.

Rosa: It’s about showing that you don't need massive models to get high accuracy in this specific type of magnetic field prediction task.

Dev: They also mention developing a robust training cost function that combines sequence tracking error and energy-related metrics using an adapted RMS error loss, weighting the quadratic tracking error on H with pointwise errors weighted by the change in B to account for energy considerations.

Taro: That weighted loss idea seems like it addresses those practical issues of balancing prediction fidelity with physical energy constraints during operation.

Rosa: It’s a sophisticated way to train the model, trying to make sure it learns not just what the field looks like, but how that field relates to the energy state of the core.

Title and authors: Dev: They also explore improving generalization across different material types by systematically investigating a Pareto front of various model architectures, allowing researchers to select the optimal model size versus accuracy trade-off based on specific application requirements.

Taro: That’s a very practical approach for anyone trying to use this in a real design flow; you can pick the right tool for the job instead of forcing one architecture onto everything.

Rosa: It gives researchers flexibility, which is always valuable when you're trying to apply complex models outside of a perfectly controlled lab setting.

Dev: They also propose incorporating physically motivated regularization terms, like Physics-Informed Neural Networks or a differentiable version of phenomenological models like Preisach, into the training loss function as a regularization signal.

Taro: Adding physical constraints directly into the learning process should help prevent the model from predicting non-physical behaviors that we saw with some of those earlier phenomenological models.

Rosa: That brings in that desire for interpretability; if you can bake in known physical laws, the resulting predictions are usually more trustworthy.

Dev: Finally, they suggest exploring hybrid architectures like GRU-L, which directly parameterizes linear models to predict material permeability in real-time alongside the field prediction.

Taro: That would be really interesting because it gives us an estimate of a fundamental physical property—permeability—which is often hard to measure directly.

Rosa: So, we're not just getting a field prediction anymore; we’re getting insight into the material properties themselves, which has huge implications for material science.

Dev: Overall, the work on RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation shows a solid approach to using parameter-efficient recurrent networks for magnetic field prediction.

Taro: The ability to model transient magnetic fields without needing slow first-principles simulations for every prediction step is a significant practical benefit for real-time control systems.

Rosa: It gives us a powerful modeling backbone that can be tuned based on whether we need the speed of a small model or the precision of a larger one, and I think this paper sets up good ground for future work in this area.

Dev: We should keep an eye on how they tackle those generalization issues across more diverse material types as they move forward with their research on RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation.

Taro: I agree, and I hope we see these efficient models integrated into more complex autonomous systems soon.

Rosa: Well, it's been really interesting discussing this paper; we'll be ready for the next one when it comes.

The paper's summary: Rosa: So, to wrap up what we've seen so far, this paper introduces an AI model called GRU-P which is designed to predict how the magnetic field changes over time based on what we observe in terms of flux density and temperature.

Dev: Exactly, Rosa; it’s essentially taking those messy real-world measurements and feeding them into a recurrent structure to estimate the next state of that magnetic field trajectory.

Taro: I think the main point they are driving home is how this system attempts to handle those non-stationary excitation waveforms, which is where things usually get complicated in practical applications.

Rosa: That's right, Taro; it focuses on making predictions for transient magnetic fields even when the input signal isn't steady, which is a big deal for anything that moves or changes state.

Dev: From an engineering standpoint, the architecture they propose has a specific warmup process to stabilize those initial predictions before it gets going with the main trajectory estimation part.

Taro: And I'm interested in how well this system performs when you throw it into a scenario where things go unexpectedly, like sudden environmental shifts or input deviations.

Rosa: That's the million-dollar question for me; we need to know if this model is robust enough to operate outside of a perfectly controlled laboratory setting for any meaningful duration.

Dev: The paper shows some promising results on unseen test data, achieving relatively low error scores, which suggests a good level of generalization across different material types they tested.

Taro: Low error scores are important, but we need to know if that accuracy holds up when you introduce a completely novel material where the physical properties are very different from what was in the training set.

Rosa: That's a fair challenge; while five materials were used for testing, it's hard to guarantee broad applicability without more extensive validation across an even wider range of conditions.

Dev: The model’s parameter efficiency is also a major point; they managed to achieve decent accuracy with only about three hundred twenty-five parameters, which is really promising for deployment constraints.

Taro: That efficiency is what makes it attractive for real-time systems, but we have to be careful that you don't sacrifice too much physical fidelity just to save on computational resources.

Rosa: So it’s a trade-off between being fast enough to run and being accurate enough to be useful in the field, which is exactly the kind of challenge we face in robotics.

Dev: Right, and that training cost function they developed, combining tracking error with energy metrics, shows they’re trying to make sure the predictions are physically plausible during operation.

Taro: That focus on physical plausibility through loss functions is something I really appreciate because it helps prevent the AI from learning weird behaviors that wouldn't happen in reality.

Rosa: It sounds like they’ve put a lot of thought into making this model not just mathematically sound, but also physically grounded in the behavior of magnetic materials.

Dev: And the use of JAXthree for implementation means they’ve already considered how to get this running on hardware efficiently, which is a huge hurdle for us when we think about deployment.

Taro: I'm still curious about the long-term vision; if this model can reliably predict these field trajectories, what kind of autonomy features could we unlock with that capability?

Rosa: That’s where I want to focus; imagine robotic systems that can anticipate magnetic field changes in front of them without needing slow simulations running constantly.

Dev: And if the inference rate is high enough, it could mean very low latency control loops, which is critical for anything involving physical movement or interaction with magnetic components.

Taro: If we can get reliable predictions under dynamic conditions, it opens up possibilities for much more sophisticated autonomous navigation where the environment itself isn't static.

Rosa: So we’ve seen the core idea and some promising results; now the real test is seeing if this model can live in a messy, unpredictable physical world over an extended period.

The paper's improvements: Rosa: So, we've looked at how they build this GRU-P model to handle those magnetic field predictions, and now I want to talk about what they suggest as ways to make it even better for real-world use.

Dev: Right, Rosa; the authors don't just stop at the initial version; they outline several specific improvements aimed at boosting accuracy and robustness across different scenarios.

Taro: I'm particularly interested in their suggestion to integrate physically motivated regularization terms, like using a differentiable version of a Preisach model to guide the AI during training.

Rosa: That makes sense, Taro; baking physical laws directly into the loss function should help prevent the GRU from generating predictions that are fundamentally impossible in physics.

Dev: From an engineering standpoint, that regularization could help stabilize the model’s learning process when dealing with noisy or incomplete sensor data during operation.

Taro: It's a way to enforce known material constraints, which is essential for autonomy because we need systems that respect the laws of nature even when things go sideways.

Rosa: And they also suggest exploring hybrid architectures like GRU-L, which would allow the model to directly predict material properties like permeability while estimating the field simultaneously.

Dev: If the AI can provide an estimate of permeability in real-time alongside field data, that’s a significant step up in terms of actionable information for a control system.

Taro: That moves the model from just predicting a value to providing insight into the underlying material characteristics, which is way more useful when we're trying to understand complex systems.

Rosa: I think that capability would be fantastic for developing smarter, more adaptable robotic systems that can react intelligently to their surroundings.

Dev: And these suggested improvements in loss functions and architectures are clearly focused on pushing those error metrics—the SRE and NERE scores—even lower for more demanding applications.

Taro: Lower error scores mean the system is performing better when the world misbehaves, which is exactly what we need for reliable autonomous decision-making.

Rosa: So, these improvements are really about taking a solid modeling result and refining it into something that can actually perform reliably in complex, uncontrolled environments.

Dev: Indeed, and this focus on improving generalization across more varied material types through architecture selection gives us a much better tool for designing systems that can handle diverse hardware.

Taro: If we can select the right model size based on the specific constraints of an application, it makes deploying these kinds of AI solutions much more practical for different research groups.

Rosa: It sounds like the future work is really about making this modeling backbone versatile and resilient enough to be a useful tool in many different engineering domains, not just one specific lab setup.

Conclusion: Rosa: So we've covered the GRU-P model, its impressive parameter efficiency of three hundred twenty-five parameters, and how they trained it using that adapted RMS error loss function for those magnetic field predictions.

Dev: That's right; we saw how they handled the sequence tracking error by weighting it against the change in flux density to incorporate energy considerations into the training process.

Taro: I still think the most important thing is how this AI handles those unpredictable scenarios, because that’s what matters when you try to build autonomous systems that have to deal with real-world chaos.

Rosa: Exactly, Taro; the goal here is to create a modeling backbone that can give us accurate field estimations without needing computationally expensive first-principles simulations for every single step.

Dev: And for me, the performance on unseen test data suggests it has a decent chance of being usable in real-time control loops if we can manage the latency effectively.

Taro: I'm still focused on the long game; if this works reliably outside a controlled lab environment, we could unlock applications in robotics that need to anticipate magnetic field changes dynamically and adapt quickly.

Rosa: That’s a huge potential application, Taro; imagine robotic systems that can react instantly to changing magnetic environments without waiting for slow simulations.

Dev: The paper shows the framework is designed for efficiency on hardware, which means we're looking at a pathway toward deploying this kind of modeling backbone in embedded control systems sooner than we might have thought.

Taro: It's encouraging to see such an efficient architecture being put into practice for such a complex physics problem; it shows that data-driven methods can be quite effective when paired with smart engineering techniques.

Rosa: So, to summarize, the paper on "RHINO-MAG: Recursive H-Field Inference based on Observed Magnetic Flux Density under Dynamic Excitation" gives us a highly efficient GRU model that predicts magnetic field trajectories with reasonable accuracy even under dynamic conditions.

Dev: It’s a solid foundation for reducing the computational load on complex simulations, provided we can keep the inference loop rate high enough for control purposes.

Taro: I think we need to keep pushing for those improvements they suggested, especially integrating physical regularization, because that’s what will really give us confidence in deploying this system autonomously.

Rosa: We're really excited about the direction this research is heading; it feels like a lot of the hard work needed to move these models from theoretical concepts into practical engineering tools.

Dev: I agree; we need to keep checking those loop rates and failure modes as we start thinking about how this AI will actually integrate into our control hardware.

Taro: I'm looking forward to seeing how they tackle those generalization issues across more diverse physical systems in future work, because that’s the next big hurdle for autonomy.

University of Siegen

eess.SY, cond-mat.mtrl-sci, cs.SY, eess.SP

Submitted: 2026-03-31

Updated: 2026-09-09

DOI: 10.1109/TPEL.2026.3733311

Code: https://github.com/PaulShuk/MagNetX

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 58/100

The gist: Based on a Pareto investigation, a rather black-box gated recurrent unit (GRU) model structure with a graceful initialization setup was found to offer the most attractive model size vs.

Key concepts

GRU Model: A black-box gated recurrent unit (GRU) model structure was found to be the most attractive for this prediction task due to its good trade-off between model size and accuracy. It is used for predicting transient magnetic fields.
Feature Engineering Normalization: This process involves normalizing raw magnetic field and temperature values by dividing them by the maximum absolute value observed in each material's training set (Hmax, Bmax, ϑmax). This standardizes input scales to prevent the model from being overwhelmed by differences in raw data magnitudes across materials.
RMS Error Loss (LaRMSE): This adapted cost function incorporates both tracking error on H and pointwise errors on B weighted by the change in B. It is normalized using the RMS value of H0:k3 to create a weighted loss L'aRMSE for backpropagation, balancing tracking accuracy with energy considerations.
Recursive Prediction: The model uses a recursive approach where the first element of the hidden state vector is used as the normalized prediction for H in each iteration. This means it continuously refines its prediction based on its own previous output, which is effective for estimating H trajectories.

Terminology

Core — 34× in this paper · explained in 2 episodes
Accuracy — 30× in this paper · explained in 1 episode
Hysteresis — 28× in this paper · explained in 3 episodes
Regularization — 11× in this paper · explained in 3 episodes
Parameter Efficiency — 10× in this paper · explained in 3 episodes
Ablation — 9× in this paper · explained in 1 episode
Loss Function — 9× in this paper · explained in 2 episodes
Generalization — 6× in this paper · explained in 13 episodes
Ablation study — 4× in this paper · explained in 1 episode
Feature Engineering — 4× in this paper · explained in 3 episodes
Ablation Studies — 3× in this paper · explained in 1 episode
Adaptation — 3× in this paper · explained in 1 episode
Fidelity — 3× in this paper · explained in 2 episodes
Robustness — 3× in this paper · explained in 5 episodes
Archetype — 2× in this paper · explained in 1 episode
Chair of Interconnected Automation Systems — 2× in this paper
Hybrid Architectures — 2× in this paper · explained in 3 episodes
Interpretability — 2× in this paper · explained in 15 episodes
System Identification — 2× in this paper · explained in 3 episodes
Activation Function — 1× in this paper · explained in 1 episode
Adam — 1× in this paper · explained in 2 episodes
Adjoint Method — 1× in this paper · explained in 1 episode
Benchmark — 1× in this paper · explained in 2 episodes
Domain Walls — 1× in this paper · explained in 2 episodes
Failure Modes — 1× in this paper · explained in 3 episodes
Foundation Model — 1× in this paper · explained in 3 episodes
Foundation Models — 1× in this paper · explained in 7 episodes
Generalizability — 1× in this paper · explained in 6 episodes
Multi-objective Optimization — 1× in this paper · explained in 4 episodes
Physical Constraints — 1× in this paper · explained in 7 episodes
Pruning — 1× in this paper · explained in 4 episodes
Quantization — 1× in this paper · explained in 12 episodes
Transfer Learning — 1× in this paper · explained in 10 episodes
Transformer — 1× in this paper · explained in 2 episodes
Verification — 1× in this paper · explained in 2 episodes

Summary

Based on a Pareto investigation, a rather black-box gated recurrent unit (GRU) model structure with a graceful initialization setup was found to offer the most attractive model size vs. model accuracy trade-off in the small-model regime, while the examined physics-inspired models performed worse. For a GRU-based model architecture with only 325 parameters (trained separately per material), an average sequence relative error of 8.02 % and an average normalized energy relative error of 1.07 % across five different materials were achieved on unseen test data. With this excellent parameter efficiency, the proposed model won the first place in the performance category of the MC2.

The modeling task is to estimate the scalar component of the magnetic field along the excitation and measurement direction, denoted as Hˆk with k ∈ [k1, k2], based on previously observed magnetic field Hk with k ∈ [k0, k1−1], magnetic flux density Bk with k ∈ [k0, k2], and core temperature ϑ. This is sought in the form of a model of the form Hˆ k1:k2 = M(Bk0:k2, Hk0:k1−1, ϑ).

The feature engineering involves normalization by computing the maximum absolute value for raw magnetic field, magnetic flux density, and temperature values (Hmax, Bmax, and ϑmax) independently for each material training data set. Normalization is performed by simple division: z˜ = z / zmax with z ∈ [H, B, ϑ]. The featurized input matrix Xk0:k2 is built as:

Xk0:k2 = [B˜ k0 B˜ k0+1... B˜ k2 ∆B˜ k0 ∆B˜ k0+1... ∆2B˜ k0 ∆2B˜ k0+1... ϑ.

The training cost function utilized an adapted root mean squared (RMS) error as the loss function: LaRMSE = vuuut Pk2 k=k1 (H˜ k − Hˆ˜ k) squared · (B˜ k − B˜ k−1) squared / (k2 − k1 + 1). Additionally, the loss is normalized with the RMS value of the full sequence H0:k3 from which H˜ k0:k2 and B˜ k0:k2 are sampled. The weighted loss L′aRMSE = LaRMSE · Hmax / (k3 + 1) Σ k=0 H2 k!−1/2 (8) is backpropagated to obtain the gradient with respect to model parameters θ.

The proposed GRU-P model architecture involves a warmup process:

The first hidden state is created by concatenating the first normalized field value with zeros g′k0 = [H˜ k0 0]. The featurized input sequence Xk0+1:k1−1 = [xk0+1, xk0+2,..., xk1−1] is then fed sequentially into the GRU cell (Equation 13). The corresponding hidden state is created analogously to the first hidden state, where the first element is always set to the true normalized magnetic field value, while the remaining elements of the vector are filled with the hidden state produced by (Equation 11) in the last iteration g′k = H˜ k g(1:) k. The final hidden state of this warmup process g′k1−1 is then given to the H-trajectory estimation process (lower part of Fig. 3). Here, the featurized input sequence Xk1:k2 = [xk1, xk1+1,..., xk2] is fed sequentially into (Equation 15), while the hidden state from the last iteration is always looped back. In each iteration, the first element (index 0) of the hidden state vector is used as the normalized prediction for the magnetic field, i.e., Hˆ˜ k = Hˆ˜θ (xk, gk−1) = g(0) k (Equation 16), which finally produces the estimated H-field trajectory Hˆ k1:k2 after concatenation and denormalization.

The model is implemented using the JAX3 Python library for seamless utilization of accelerator hardware (GPUs, TPUs) and just-in-time (JIT) compilation. The GRU-P model architecture won the first place in the performance track in the MC2, showing an excellent accuracy per parameter efficiency especially in the small-model regime. With only 325 parameters and an average SRE score of 8.02 % and an average NERE score of 1.

Improvements for AI systems

Here are specific improvements for existing AI systems, based on the RHINO-MAG paper:

Improve core loss prediction accuracy for transient magnetic fields by deploying a parameter-efficient Gated Recurrent Unit (GRU) model, specifically the GRU-P architecture with a warmup mechanism.
Enable time-resolved and temperature-aware H-field prediction for magnetic components operating under quasistationary or non-stationary excitation waveforms (e.g., power factor correction, electric motor drives).
Achieve high parameter efficiency (as demonstrated by the GRU with only 325 parameters) while maintaining excellent prediction accuracy on unseen test data, specifically targeting low Sequence Relative Error (SRE) scores and Normalized Energy Relative Error (NERE) scores.
Develop a robust training cost function that combines elements of both sequence tracking error (SRE) and energy-related metrics (NERE), utilizing an adapted Root Mean Squared error loss function that weights quadratic tracking error on H with pointwise errors weighted by the change in B to account for energy considerations.
Improve model generalization across different material types by systematically investigating a Pareto front of various model architectures (GRU variants, LSTM variants, and phenomenological models like Preisach), allowing researchers to select the optimal model size vs. accuracy trade-off based on specific application requirements (e.g., small-model regime for FEM simulations).
Enhance the ability of data-driven models to handle material property variations by incorporating physically motivated regularization terms (Physics-Informed Neural Networks, PINN) into the training loss function, using a differentiable version of phenomenological models like Preisach as a regularization signal.
Improve model interpretability and physical insight by exploring hybrid architectures such as GRU directly parameterizing linear models (GRU-L), which predicts material permeability in real-time, providing an estimate of the physical property alongside the field prediction.
Develop advanced state representation methods for complex magnetic phenomena by exploring Vector Field GRUs (GRU-V), where hidden states are arranged in 2D grids to approximate discretized magnetization vector fields, potentially offering richer spatial context than standard RNN states.

The improved AI system will be capable of:

Accurately predicting the time-resolved magnetic field trajectory without requiring computationally prohibitive first-principles micromagnetic simulations (like LLG PDE solvers) for every prediction step.
Optimizing magnetic component design by providing highly accurate estimates of core losses under complex, non-sinusoidal excitation conditions and varying operating temperatures.
Performing rapid, on-device or real-time inference due to the high parameter efficiency of the proposed GRU architecture, making it suitable for integration into embedded control systems or simulation loops (like FEM).
Serving as a versatile modeling backbone that can be tuned—either by selecting a small, efficient model for fast prototyping or by utilizing larger models with more complex structures when higher accuracy is demanded.