Cross-Material Magnetic Core Loss Prediction via Operating-Condition Matching and Partial Optimal Transport

arXiv:2610.14964 · cs.LG · Submitted 2026-10-10 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Cross-Material Magnetic Core Loss Prediction via Operating-Condition Matching and Partial Optimal Transport".

Jane: The gist The proposed method, T-PCOT, achieves state-of-the-art predictive performance in magnetic core loss transfer learning by organizing crossmaterial alignment through operating-condition support and partial optimal transport.

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

Paper summary: Tom: We’ve gone through the whole paper on "Cross-Material Magnetic Core Loss Prediction via Operating-Condition Matching and Partial Optimal Transport." The authors, Cong Yao, Yu Zheng, Zehao Hou, and Rui Xu contributed equally to this work.

Jane: They tackled the challenge of predicting core loss when you have limited measurements by proposing T-PCOT. It claims they improve performance by separating condition support for prediction from the partial optimal transport for alignment.

Lu: The key mechanism is using frequency, flux excursion, temperature, and waveform shape descriptors to define a sparse graph of comparable operating conditions. This graph then drives a sparse linear program to create a fixed transport plan that summarizes matched mass M and target coverage κj.

Meng: So, the implication for deployment is that you build this alignment plan once based on measurements and then freeze it during training so inference is fast because there’s no need for a transport solver at runtime.

Tom: And they show concrete results: macro MRE drops from twelve point six zero percent to nine point six five percent, and macro P95 drops from thirty point nine four percent to twenty-five point zero six percent. That’s a solid improvement over the T-MKMMD baseline.

Jane: In simple terms, it means that by using operating condition support, you get better prediction errors than global alignment methods when comparing materials. It shows that focusing on supported pairs helps reduce those errors significantly across different material pairs and label budgets.

Lu: The design involves condition-driven sparse matching, capacity-constrained partial transport, and objective-preserving weighted training. This sets a new direction for how we can approach crossmaterial alignment in core loss prediction.

Meng: It’s about controlling extreme errors when your source condition coverage contracts, which is something we need to watch out for as we scale things up.

Tom: Exactly. We’re suggesting adapting that accepted transport mass to predictive needs and controlling those extreme errors when source condition coverage shrinks are directions for further development with T-PCOT.

Conclusion: Tom: So, we're wrapping up on this T-PCOT paper about core loss prediction using operating conditions and transport plans for aligning different materials across different measurements.

Jane: Right, so the title is "Cross-Material Magnetic Core Loss Prediction via Operating-Condition Matching and Partial Optimal Transport." It sounds really technical. What is it trying to solve, basically?

Lu: It’s solving the problem of predicting core loss when you have limited data by separating how you predict from how you align things. The authors use a shared Transformer backbone for both tasks.

Meng: So they're using the same AI model structure for two different jobs—predicting and aligning—which is a neat way to share knowledge, but what’s the actual trick?

Tom: The trick is this: they define these operating conditions using things like frequency and temperature, and then they use a sparse graph based on those conditions to guide the matching.

Jane: I see. So it's not just matching any two materials together, but only matching ones that have similar operating conditions first? That makes sense for physical reality.

Lu: Exactly. They define these condition descriptors—like "log10 f," "log10 A," and temperature—and then calculate a cost between pairs based on how far apart those condition values are.

Tom: And then they use a sparse linear program, which is this optimization thing, to figure out the best way to assign mass from your sources to your targets.

Meng: So they build this transport plan once before training? That sounds like it could save a ton of time during actual inference if that planning step doesn't need recalculating constantly.

Jane: It seems designed for efficiency, which is huge when you’re dealing with real-world hardware where you can’t run a complex optimization every time you want a prediction.

Lu: The training objective combines the standard regression loss with this transport objective, which concentrates the alignment only on those supported pairs defined by the condition graph.

Tom: And when they did the experiments, they showed that this method cuts down error metrics—like mean relative error and that P95 error—compared to other methods like T-MKMMD.

Jane: So what does that mean for a person who just uses this for their own work? It means you get much more reliable predictions across different materials.

Lu: It shows that organizing the alignment through these operating conditions gives a real predictive advantage over just using a global alignment approach.

Meng: The ablation studies showed that having source capacity limits actually mattered a lot, meaning accepting some less-optimal matches can actually help improve those error metrics.

Tom: So to sum up, T-PCOT organizes alignment through condition support and uses partial transport to get lower prediction errors than the state-of-the-art baseline.

Jane: It suggests that for crossmaterial work, focusing on matching things that are physically similar in terms of their operating environment is a very smart way forward.

Lu: And the authors point out that this design, with its condition-driven sparse matching and capacity constraints, gives us a solid foundation for designing alignment strategies in core loss prediction.

Tom: It’s about adapting the accepted transport mass to actually serve our predictive needs while keeping an eye on those extreme errors when your source condition coverage shrinks.

Jane: That’s a big direction to watch—controlling those extreme errors is something we really need to focus on next with this kind of approach.

Cong Yao, Yu Zheng, Zehao Hou, Rui Xu

Changsha University of Science and Technology · University of Arkansas

cs.LG

Submitted: 2026-10-10

Updated: 2026-10-10

The gist: The gist The proposed method, T-PCOT, achieves state-of-the-art predictive performance in magnetic core loss transfer learning by organizing crossmaterial alignment through operating-condition

Key concepts

Operating Condition Description
This defines a set of key descriptors—like frequency, flux excursion, and temperature—that characterize the magnetic core's state. These descriptors form a sparse graph where the cost between two conditions is calculated based on their distance, guiding which operating conditions are considered comparable for matching.
Partial Transport
This uses a sparse linear program to determine how to optimally match source and target data points. It minimizes the sum of costs while respecting limits on how much data can be matched from each source and target, ensuring the resulting alignment plan is feasible.
Pretraining Objective
The training loss combines two parts: one for standard regression (predicting core loss) and another for transport. The transport term specifically forces the network to concentrate its attention on pairs of operating conditions that are deemed well-matched by the partial optimal transport plan.

Terminology

Summary

The gist The proposed method, T-PCOT, achieves state-of-the-art predictive performance in magnetic core loss transfer learning by organizing crossmaterial alignment through operating-condition support and partial optimal transport.

Problem Formulation

T-PCOT addresses the challenge of cross-material magnetic core loss prediction by separating the roles of operating conditions in prediction and alignment. The method utilizes a shared Transformer backbone for both tasks, where a pointwise projection maps the waveform to 24 channels, followed by sinusoidal positional encoding and two Transformer encoder layers. A projected operating-condition vector is added to the encoded sequence before two shared fully connected blocks of width 40. These features share the regression weights, while conditions influence the choice and strength of pairwise interactions without directly satisfying feature-consistency penalty through their own projection.

Operating-Condition Description

The method defines a sparse graph of comparable operating conditions using frequency, flux excursion, temperature, and waveform shape descriptors. The descriptor v(x) is defined as log10 f, log10 A, T − 50/40, p L−1 P k b 2 k A, log10 L−1 X k d 1.5 k!, std(d) mean(d).. Pairwise cost Cij is calculated based on the Euclidean distance between source and target descriptors, scaled by the number of active dimensions da. The neighborhood scale and candidate graph size are determined by rules involving source and target pool sizes, such as K = min(ns, max(64, 2kτ)).

Partial Transport

The alignment path is formulated as a sparse linear program to assign masses to candidate pairs. The objective is to minimize the sum of pairwise costs CijΠij plus a rejection price τXjrj. Constraints include source capacity limits, XjΠij ≤ ai, and target mass allocation, XiΠij + rj = bj. This plan is constructed once from measured inputs and remains fixed during network training. The resulting plan summarizes total matched mass M = Pij Πij, target coverage κj = nuP i Πij, and source utilization ηi = nsP j Πij, ensuring M, κj, ηi ∈ [0, 1].

Training and Prediction

The pretraining objective L(θ) combines uniform source regression with the transport objective. The loss is defined as L(θ) = X i ai(lbθ(x s i)−l s i) squared + λ Xij Πij Dθ(i, j), where LOT = Pij ΠijDθ(i, j) concentrates alignment on supported pairs. For minibatch optimization, a combined sampling scheme qij is used to preserve both terms of the objective. After pretraining, supervised fine-tuning updates all prediction-network weights using target labels in the same log-loss space. Inference executes only the regression path without a neighbor search or transport solver.

Improved Cross-Material Prediction

Experiments demonstrate that T-PCOT reduces mean relative error (MRE) and 95th-percentile error (P95) relative to Transformer-based multiple-kernel maximum mean discrepancy (T-MKMMD) in all nine comparisons. Macro MRE decreases from 12.60% to 9.65%, and macro P95 from 30.94% to 25.06%, corresponding to relative reductions of 23.41% and 19.01%. The controlled comparisons show that organizing crossmaterial alignment through operating-condition support offers a predictive advantage over the compared global alignment method. Ablation studies confirm that source capacity has a clearer effect on both metrics, showing that accepting additional, more costly matches can be beneficial for improving MRE and P95.

Conclusion

T-PCOT successfully organizes alignment through operating-condition support to achieve lower prediction errors than T-MKMMD across various material pairs and label budgets. The method's design, which includes condition-driven sparse matching, capacity-constrained partial transport, and objective-preserving weighted training, positions operating-condition support as a useful basis for designing crossmaterial alignment in core loss prediction. The findings suggest adapting the accepted transport mass to predictive needs and controlling extreme errors when source condition coverage contracts are directions for further development.

Improvements for AI systems

  1. Bold header: Condition-driven sparse matching

Identify comparable source-target pairs by defining a sparse graph of comparable operating conditions using descriptors like frequency, flux excursion, temperature, and waveform shape. This allows for local representation alignment by restricting interactions to neighboring conditions rather than relying on a global feature distribution alignment.

  1. Bold header: Capacity-constrained partial transport

Implement a capacity-constrained partial transport plan where Source capacities control the reuse of individual measurements, ensuring that the plan explicitly manages which source measurements contribute to target predictions and which target mass is left unmatched, as described by Equation (3).

  1. Bold header: Objective-preserving weighted training

Utilize an importance-weighted sampling scheme during minibatch optimization to preserve both uniform source regression and the transport objective, ensuring that the training loss term retains the alignment pressure even when using a subset of positive transport entries.

Sources

Related papers